Venues会议/期刊 › CoRL 2026 › Program日程
All 687 accepted papers by session and poster number. 574 are linked to arXiv, 183 have code; the rest link to OpenReview. 555 have a one-line takeaway, including every oral and spotlight on arXiv.按日程与海报编号列出全部 687 篇录用论文：574 篇已关联 arXiv，183 篇附代码，其余链接至 OpenReview。555 篇附一句话要点，覆盖 arXiv 上的全部 Oral 与 Spotlight。
November 9–12, 2026, Austin, Texas. Orals and spotlights are at the Hilton Austin, posters at the JW Marriott.2026 年 11 月 9–12 日，美国得克萨斯州奥斯汀。Oral 与 Spotlight 在 Hilton Austin，海报在 JW Marriott。 Sessions, titles and poster numbers come from the日程、标题与海报编号来自 official list官方列表 (2026-11-07); arXiv links are matched by title, so check there for last-minute changes. Code links come from arXiv comments, Hugging Face and the authors' project pages.arXiv 链接按标题匹配；临时变动请以官方页面为准。代码链接来自 arXiv 备注、Hugging Face 与作者项目页。
Code found: 41% of orals and spotlights, 24% of the other papers. This counts only links we could find, so the real share is higher.已找到代码：Oral 与 Spotlight 中占 41%，其余论文中占 24%。这里只统计我们能找到的链接，实际比例更高。
Keyword counts on the titles of all accepted papers: a paper can count in several themes, and a paper that does not name a theme in its title is not counted. Click a theme to list its papers.基于全部录用论文标题的关键词统计：一篇论文可计入多个主题，标题未提及的不计入。点击主题可列出对应论文。
A public signal, not a quality ranking: a star count can belong to a larger project, and papers without code cannot appear here.这只是公开信号，并非质量排名：星标可能属于更大的项目，没有代码的论文也不会出现在这里。
Press ☆ on the papers you want to see: you get your own schedule, sorted by time, with a link to share and a calendar file. The list stays in this browser; nothing is sent anywhere.点论文上的 ☆ 即可生成个人日程：按时间排序，可分享链接、导出日历文件。清单只保存在你的浏览器里，不会上传。
Mon 11/9, 9:00–9:30 AM · Hilton Austin · 9 papers
CoRL 2026 Spotlight. Splits the generalization gap of sequential instruction-following policies into marginal, compositional and context-action shift, viewed through instruction-space coverage.CoRL 2026 Spotlight。从指令空间覆盖出发，把序列任务策略的泛化差距分解为边际偏移、组合偏移与上下文-动作偏移。
CoRL 2026 Spotlight. Training-free steering that makes Cartesian, embodiment-agnostic policies aware of the deployed robot's body, using joint-space guidance for constraints such as whole-body collision avoidance.CoRL 2026 Spotlight。无需训练的引导框架：用关节空间引导让与本体无关的笛卡尔策略感知实际机器人本体，处理全身避碰等约束。
CoRL 2026 Spotlight. Similar prompts can make VLAs behave very differently, so this framework interactively searches for language sequences that raise closed-loop task performance.CoRL 2026 Spotlight。语义相近的指令可能让 VLA 行为迥异，该框架通过交互搜索语言序列来提升闭环任务表现。
CoRL 2026 Spotlight. Gives VLAs memory at several timescales: abstract long-term state (which recipe stages are done) plus short-term recall of recent events and occluded objects.CoRL 2026 Spotlight。为 VLA 提供多时间尺度记忆：长期记住抽象状态（如菜谱进行到哪一步），短期记住近期事件与被遮挡物体。
Fully open successor to MolmoAct: a flow-matching action expert on an embodied-reasoning VLM, the OpenFAST tokenizer and a 720 h bimanual YAM dataset; reports beating π0.5 on 7 sim and real benchmarks.MolmoAct 的全开源后继：在具身推理 VLM（MolmoER）上挂接流匹配连续动作专家，并发布开源动作分词器 OpenFAST 与 720 小时双臂 YAM 遥操作数据集；论文报告在 7 个仿真与真机基准上超过 π0.5 等基线，权重、训练代码与数据全部开放。
Reframes VLA pretraining as goal-conditioned RL: contrastive learning gives the model a sense of whether a language goal is reachable. Pretrained on 167B tokens; strongest on long-horizon tasks.将 VLA 预训练重构为目标条件强化学习：用对比学习让模型判断语言目标是否可达。基于 1670 亿 token 预训练，在长时序任务上优势最明显。
Finds that fine-tuning erodes the semantic structure VLAs inherit, and that this tracks task failure. A training-only anchor restores it, reporting up to +21.5% on out-of-distribution tasks.发现微调会破坏 VLA 继承的语义结构，且这与任务失败同步。提出仅在训练时使用的语义锚定方法，报告分布外任务提升最高 21.5%。
CoRL 2026 Spotlight. VLAs can ignore language and follow vision shortcuts from dataset bias. Introduces LIBERO-CF, a counterfactual benchmark for language following.CoRL 2026 Spotlight。VLA 可能忽略语言、沿用数据偏差形成的视觉捷径。论文提出反事实基准 LIBERO-CF，用于评测语言遵循能力。
CoRL 2026 Spotlight. A controlled study of zero-shot VLA transfer that isolates embodiment change from tasks, scenes and protocols, and separates strict zero-shot from looser definitions.CoRL 2026 Spotlight。对 VLA 零样本跨形态迁移的对照研究：将本体变化与任务、场景、流程隔离，并把严格零样本与宽松定义区分开。
Mon 11/9, 10:30–11:00 AM · Hilton Austin · 3 papers
CoRL 2026 Oral. Cross-embodiment VLA that treats human interaction traces as a shared language, pre-trained on UniHand-2.0: over 35,000 hours of multimodal data across 30 robot embodiments.CoRL 2026 Oral。跨形态 VLA，把人类交互轨迹当作通用“母语”，在 UniHand-2.0 上预训练：超过 3.5 万小时多模态数据，覆盖 30 种机器人形态。
CoRL 2026 Oral. Test-Time-Training policies scale visuomotor context to 8K timesteps, three orders of magnitude longer, without added inference latency, enabling one-shot imitation from human video.CoRL 2026 Oral。Test-Time-Training 策略把视觉运动上下文扩展到 8K 时间步，长出三个数量级且不增加推理延迟，由此支持从人类视频单次模仿等新能力。
CoRL 2026 Oral. A steerable generalist foundation model: follows diverse instructions in unseen scenes and reaches zero-shot cross-embodiment tasks such as folding laundry, with no task-specific finetuning.CoRL 2026 Oral。可引导的通用基础模型：在未见环境中执行多样指令，跨形态零样本完成叠衣等任务，无需任务专用微调。
Mon 11/9, 11:00–11:30 AM · Hilton Austin · 3 papers
CoRL 2026 Oral. Argues VLAs fail at dynamic tasks because one observation brings motion ambiguity and state aliasing, and adds learned temporal context to resolve both.CoRL 2026 Oral。指出 VLA 难以应对动态任务，是因为单帧观测带来运动歧义与状态混淆，并通过学习时序上下文同时解决这两个问题。
CoRL 2026 Oral. DreamZero: a World Action Model on a pretrained video diffusion backbone, jointly predicting video and actions, so it learns skills from heterogeneous robot data without repetitive demonstrations.CoRL 2026 Oral。DreamZero：基于预训练视频扩散骨干的世界动作模型，联合预测视频与动作，可从异构机器人数据学习技能，无需重复示范。
Mon 11/9, 11:30 AM–12:00 PM · Hilton Austin · 3 papers
CoRL 2026 Oral. Measures diversity of imitation datasets directly from demonstrations, with entropy defined through a signature kernel, so no model has to be trained first.CoRL 2026 Oral。直接从示范数据度量模仿学习数据集的多样性：基于签名核定义熵，无需先训练模型。
CoRL 2026 Oral. Argues simulation can be enough: large, diverse synthetic data gives zero-shot real transfer for static and mobile manipulation. MolmoBot-Engine is an open procedural data pipeline.CoRL 2026 Oral。挑战“仅靠仿真不够”的观点：足够大规模、多样的合成数据即可零样本迁移到真实静态与移动操作。MolmoBot-Engine 是开源的程序化数据管线。
CoRL 2026 Oral. Audits LIBERO, CALVIN, SimplerEnv, RoboCasa and RoboTwin 2.0 against four failure modes, from shortcut solvability to creeping overfitting. LIBERO and CALVIN fail several diagnostics.CoRL 2026 Oral。用捷径可解、统计不显著、渐进过拟合、数据来源依赖四种失效模式审计 LIBERO、CALVIN、SimplerEnv、RoboCasa 和 RoboTwin 2.0；其中 LIBERO 与 CALVIN 未通过多项诊断。
Mon 11/9, 2:30–3:00 PM · Hilton Austin · 9 papers
CoRL 2026 Spotlight. A fully native 1-bit VLA with ternary weights, built on BitNet b1.58 2B4T, designed for edge devices instead of relying on post-hoc compression.CoRL 2026 Spotlight。原生 1-bit VLA，权重三值化，基于 BitNet b1.58 2B4T，面向边缘设备设计，而非依赖事后压缩。
CoRL 2026 Spotlight. Model-based RL that uses a chunk world model as a differentiable training-time evaluator, targeting compounding error, inference latency and BPTT instability.CoRL 2026 Spotlight。将分块世界模型用作训练期可微评估器的模型式强化学习，针对累积误差、推理延迟与 BPTT 不稳定问题。
CoRL 2026 Spotlight. Diffusion sequence prediction with selective re-noising, so segments of the horizon can be refined again using context from the rest.CoRL 2026 Spotlight。在扩散序列预测中选择性重新加噪，让时间跨度内的片段可以借助其余片段的上下文再次修正。
EagleVLA: asynchronous VLA inference for low-power onboard chips such as Jetson Orin, tackling perception-execution misalignment and slow reaction; a CoRL 2026 Spotlight, formerly titled Jetson-PI.EagleVLA：面向 Jetson Orin 等低功耗机载设备的 VLA 异步推理方法，针对异步执行带来的感知-执行错位与反应慢两个问题；CoRL 2026 Spotlight，曾用名 Jetson-PI。
Federated RL for heterogeneous clients that aggregates diffusion behavior priors with optimal transport rather than averaging policies, plus a low-variance DICE value baseline.面向异构客户端的联邦强化学习：不再平均策略，而是用最优传输聚合扩散行为先验，并加入低方差的 DICE 价值基线。
CoRL 2026 Spotlight. Instead of regenerating action chunks from scratch, keeps a persistent action buffer with staggered flow timesteps, so flow-based VLAs get more reactive while preserving long-horizon context.CoRL 2026 Spotlight。不再每次从头生成动作块，而是维护一个流时间步错开的持久动作缓冲区，让基于流的 VLA 更具反应性，同时保留长时域上下文。
A policy predicts each action and how long to hold it. Reports matching high-frequency control at under half the rate on an RC car and Unitree Go1, plus simulated VLA inference.策略同时预测动作及其保持时长。报告在遥控车、Unitree Go1 和仿真 VLA 上以不到一半的控制频率达到高频控制的性能。
CoRL 2026 Spotlight. Shows world models are a stealthy data-poisoning entry point: compromised policies can be deployed even when the ground-truth training data looks safe.CoRL 2026 Spotlight。展示世界模型可成为隐蔽的数据投毒入口：即使真实训练数据看似安全，部署的策略仍可能被破坏。
Mon 11/9, 5:00–6:30 PM · JW Marriott · 228 papers · Lone Star (3rd floor): posters 1–116 · Griffin (2nd floor): posters 117–232
Training-only 3D flow heads (point-level and gripper-level) teach a 2D VLA how the scene moves; no extra sensors or inference cost. Improves success, zero-shot robustness and data efficiency.仅在训练时加入 3D 流预测头（点级与末端级双重监督），让 2D VLA 学到场景动力学；推理时无需额外传感器或计算。提升成功率、零样本鲁棒性与数据效率。
VLA with sparse experts: sensor availability gates modality experts, so it degrades gracefully when depth is missing, and task intent routes action experts. Strong on LIBERO, RoboCasa GR1 and real dual-arm tasks.稀疏专家 VLA：传感器是否可用决定模态专家的开关，缺少深度时也能平稳退化；任务意图再路由到动作专家。在 LIBERO、RoboCasa GR1 和真实双臂任务上表现强。
Finds VLAs over-rely on the wrist camera; predicts future interaction intent from past actions to route between global and wrist views per stage. Reports +10.1 pp over π0.5 on RoboTwin 2.0 and 71.7% → 85.0% real.发现现有 VLA 过度依赖腕部相机；根据历史动作预测未来交互意图，按任务阶段在全局与腕部视角间动态路由；报告在 RoboTwin 2.0 上比 π0.5 高 10.1 个百分点，真机成功率由 71.7% 提升到 85.0%。
VLA that lifts VLM features into 3D scene tokens, predicts how scene points will move, and conditions a 3D flow action denoiser on that motion. Reports SOTA on CALVIN without robot-action pretraining.把 VLM 特征提升为 3D 场景 token，预测场景点将如何运动，并以此为条件驱动 3D 流匹配动作去噪器。报告在无机器人动作预训练下 CALVIN 达到最优。
Action tokenizer whose first quantization level is trained as a semantic 'action language' while deeper levels keep detail; pretrained on 2.4M trajectories. Reports gains over FAST on grounding and long-horizon tasks.动作 tokenizer：残差量化的第一层被训练成表达粗粒度运动意图的语义“动作语言”，更深层保留细节；在 240 万条轨迹上预训练。报告在多模态对齐和长程任务上优于 FAST。
Action tokenizer adapted from neural audio codecs (multi-scale RVQGAN with a Vocos-style decoder). Reports higher success than binning, FAST and earlier VQ tokenizers on LIBERO-10, RoboMimic and real tasks.借鉴神经音频编解码器的动作 tokenizer（多尺度 RVQGAN + Vocos 式解码器）。报告在 LIBERO-10、RoboMimic 与真机任务上成功率高于分箱、FAST 和已有 VQ 方法。
Derives design rules for action tokenizers from VLA optimisation (token overlap, low redundancy, independence) and builds one: a 2.2B SmolVLM2 reaches 95.5% on LIBERO with no robot pretraining.从 VLA 优化角度总结动作 tokenizer 的设计原则（时间重叠、低冗余、模态互信息、独立性）并据此构建 ActionCodec：2.2B 的 SmolVLM2 在无机器人预训练下 LIBERO 达 95.5%。
Discrete VLA that generates actions autoregressively across temporal blocks and with masked diffusion inside each block, gaining speed and supporting real-time chunking via temporal in-painting.离散 token VLA：块与块之间自回归、块内做掩码离散扩散，兼顾时间连贯性与推理速度，并可通过时间补全支持实时分块执行。
Plan-then-execute VLA that first predicts a short sequence of coarse action tokens summarising the intended trajectory, then the executable tokens conditioned on it. Largest gains on long multi-stage tasks.先规划后执行的 VLA：先预测概括未来轨迹的少量粗粒度动作 token，再以此为条件生成可执行动作 token；规划与执行共享同一离散动作词表。长程多阶段任务提升最大。
Interleaves text and image reasoning in one model: a forward chain predicts the next subgoal image, an inverse chain reasons about the actions to reach it, then actions are generated. Biggest gains on long tasks.在同一模型中交替进行文本与图像推理：正向推理预测下一个子目标图像，逆向推理据此分析如何到达，再生成动作。长程任务提升最大。
Reframes VLA pretraining as goal-conditioned RL: contrastive learning gives the model a sense of whether a language goal is reachable. Pretrained on 167B tokens; strongest on long-horizon tasks.将 VLA 预训练重构为目标条件强化学习：用对比学习让模型判断语言目标是否可达。基于 1670 亿 token 预训练，在长时序任务上优势最明显。
Uses 3D object motion as a shared interface: object motion and robot actions are modelled jointly with masking, so one pretrained model serves as policy, dynamics model and few-shot adapter. Robot, human and sim data.以 3D 物体运动轨迹作为统一接口，用掩码生成目标联合建模物体运动与机器人动作，同一套预训练参数可作为策略、动力学模型或少样本适配器。混合机器人、人类视频与仿真数据预训练。
Jointly trains a latent action model, latent policy and action decoder, so plentiful action-free video plus a few labelled robot demos can train a policy. Reports beating prior latent-action baselines.联合训练潜在动作模型、潜在策略和动作解码器，用大量无动作标注的视频加少量带标注示教即可训练策略。报告在仿真和真实任务上优于已有潜在动作方法。
Aligns robot visual features with human ones using paired human-robot demos, then pretrains a latent-action VLA on human and robot videos. Reports 4.481 average length on CALVIN ABC→D.利用成对的人-机演示把机器人视觉表征与人类表征对齐，再用人类和机器人视频预训练潜在动作 VLA。报告 CALVIN ABC→D 平均长度 4.481。
FiberTune: action-supervised VLA fine-tuning can collapse visual structure the action loss barely constrains; a training-time objective keeps those residuals, with no added inference cost.FiberTune：动作监督微调会压缩动作损失约束较弱的视觉结构（沿局部动作纤维的残差坍塌）；提出训练期目标函数，对齐冻结的视觉教师以保留这些残差，不增加推理开销。
Finds that fine-tuning erodes the semantic structure VLAs inherit, and that this tracks task failure. A training-only anchor restores it, reporting up to +21.5% on out-of-distribution tasks.发现微调会破坏 VLA 继承的语义结构，且这与任务失败同步。提出仅在训练时使用的语义锚定方法，报告分布外任务提升最高 21.5%。
Latent action model constrained to compose additively over short horizons (identity, inverse, cycle), so latents capture motion rather than scene detail and give better supervision for policies.对潜在动作空间施加短时可加组合约束（恒等、逆、循环一致），使潜在动作只编码运动而非场景细节，并为下游策略学习提供更好的监督。
Latent actions with polar structure in hyperbolic space: radius encodes how much changes, direction encodes what kind. Reports gains over latent-action baselines and pretrained VLAs.在双曲空间中构造极坐标结构的潜动作：半径表示变化多大，方向表示变化类型。报告优于其他潜动作方法和预训练 VLA。
Supervises VLAs with domain-invariant future latents to cut reliance on visual shortcuts. Reports 69.1% on LIBERO-Plus, 11.4 points above the strongest baseline.用与域无关的未来潜变量监督 VLA，减少对视觉捷径的依赖。报告 LIBERO-Plus 平均成功率 69.1%，比最强基线高 11.4 个百分点。
Fights shortcut learning when fine-tuning VLAs: predicted masks of task-relevant objects and the robot serve as an auxiliary grounding loss; nothing changes at inference. Tested on four foundation models and π0.5.针对 VLA 微调中的捷径学习：预测任务相关物体与机器人的掩码，作为辅助对齐损失引导策略关注正确区域，推理时不增加任何开销。在四个基础模型和真机 π0.5 上验证。
Trains the VLA executor with refined subtask language plus an explicit budget on how much visual evidence it may use. Reports mean subtask success up from 54.2% to 79.0% over π0.5 on eight real tasks.用更细的子任务语言重标注轨迹，并对执行器可使用的视觉证据施加显式预算来训练 VLA。报告在八个真机任务上平均子任务成功率比 π0.5 从 54.2% 提升到 79.0%。
Gives a VLA memory as motion rather than frames: a short past video becomes continuous trajectory-field tokens that current visual tokens query. Smoother, more direct long-horizon execution.让 VLA 记住“运动”而非“帧”：把一小段历史视频转成时间连续的轨迹场 token，供当前视觉 token 查询。长程操作更平滑、更直接。
Turns robot state into VLM-vocabulary tokens that, with the instruction, choose which image patches the VLM processes. Keeping ~12% of visual tokens beats the full-token baseline on CALVIN with lower latency.把机器人状态离散成 VLM 词表 token，与指令一起决定哪些图像块进入 VLM 计算。仅保留约 12% 的视觉 token 就在 CALVIN 上超过全 token 基线，同时降低延迟。
Fully open successor to MolmoAct: a flow-matching action expert on an embodied-reasoning VLM, the OpenFAST tokenizer and a 720 h bimanual YAM dataset; reports beating π0.5 on 7 sim and real benchmarks.MolmoAct 的全开源后继：在具身推理 VLM（MolmoER）上挂接流匹配连续动作专家，并发布开源动作分词器 OpenFAST 与 720 小时双臂 YAM 遥操作数据集；论文报告在 7 个仿真与真机基准上超过 π0.5 等基线，权重、训练代码与数据全部开放。
Plug-in adapter that projects the robot state onto the image to build a grounded state token and FiLM-modulate nearby features. Reports +8.7% for Diffusion Policy in sim and 10.6% average real-world gain.即插即用适配器：把机器人状态投影到图像平面，采样局部视觉特征构成“落地”的状态 token，并用 FiLM 调制对应特征。报告仿真中 Diffusion Policy 提升 8.7%，真机平均提升 10.6%。
Attention pooling over intermediate CNN features, weighted by the robot's current proprioception, to keep control-relevant local detail. Reports +36.2% and +41.2% over pooling and explicit-crop methods.以机器人当前本体状态为查询，对 CNN 中间特征做注意力池化，保留与控制相关的局部信息。报告比常规池化与显式裁剪方法分别高 36.2% 和 41.2%，仅需训练 5.8% 的编码器参数。
Seeker learns where to look from action supervision alone, giving progress-aware crops on frozen DINOv3 features. Reports real-robot success 48.3% to 76.7%, and 20% to 60% under lighting and background shifts.Seeker 仅凭动作监督学习“看哪里”，在冻结的 DINOv3 特征上给出随任务进展变化的裁剪区域。报告真机成功率从 48.3% 升到 76.7%，光照与背景变化下从 20% 升到 60%。
Turns guidance from 3D foundation models, VLMs or human pose into differentiable 3D attractor/repeller energies that steer VLA action sampling. Reports gains for π0.5 and GR00T N1.6 in sim and real.把 3D 基础模型、VLM、人体姿态等指导信号统一表示为 3D 空间中可微的吸引/排斥能量函数，引导 VLA 动作采样。报告提升 π0.5 和 GR00T N1.6 在仿真与真机上的表现。
Adds separate tactile, force and torque streams to a pretrained VLA through joint cross-modal attention, with staged training and an auxiliary task that predicts future physical signals. Real-robot tests.为预训练 VLA 增加独立的触觉、力和力矩信号流，通过跨模态联合注意力融合，配合分阶段训练和预测未来物理信号的辅助任务。真机实验验证。
Lifelong VLA learning with short-term and long-term LoRA adapters mixed by a task-aware gate, plus cheap stochastic replay. Learns new skills while keeping old ones; tested on a real xArm.终身学习 VLA：短期与长期两条 LoRA 适配通路由任务感知门控融合，再配合低存储的随机回放，兼顾学新与保旧。在真实 xArm 上验证。
RL post-training of VLAs with an online-updated world model whose measured reliability decides how much imagined data and how long rollouts to use. Reports 79.7% on perturbed LIBERO-PRO.VLA 强化学习后训练：在线更新的世界模型与奖励头作为模拟器，按其可信度动态决定想象数据比例和推演长度。报告扰动版 LIBERO-PRO 平均成功率 79.7%。
A task-manager VLM re-predicts the remaining subtasks and a 2D keypoint trace at every step; the VLA executor just follows the trace. Failed steps stay in the plan, so replanning comes for free. Sim + real Franka.由任务管理 VLM 每一步重新预测剩余子任务和一条 2D 关键点轨迹，VLA 执行器只需沿轨迹执行。失败的步骤会留在计划里，因此无需额外逻辑即可自动重规划。仿真与真实 Franka 实验。
Bounded latent memory for imitation policies on tasks where a cue disappears before the decision; it is keyed by path signatures of the robot's own trajectory and attaches via adapters. Real-robot branching tasks.为“线索在决策前已消失”的任务设计的模仿策略记忆：固定大小的潜在记忆以机器人自身轨迹的路径签名为键读写，通过适配器接入现有策略。在真机分支任务上验证。
~60M-parameter policy with event memory for tasks where the deciding cue comes long before the action, plus the real-robot Camo-Dataset. Reports decision success up from 22.5% to 80.8% and 87.1% on LIBERO-10.约 6000 万参数的视觉运动策略，带事件记忆，面向“决定性线索早于动作很久出现”的任务，并发布真机基准 Camo-Dataset。报告决策成功率从 22.5% 提升到 80.8%，LIBERO-10 达 87.1%。
Memory for VLAs that keeps only sparse keyframes the policy itself predicts will matter later. Adds the RoboTwin-MeM benchmark; reports +40% average success over memory-augmented VLAs across 17 sim and 4 real tasks.为 VLA 加入稀疏记忆：由策略自己预测哪些关键帧之后会有用并只保存这些帧。同时提出 RoboTwin-MeM 基准；报告在 17 个仿真和 4 个真机任务上平均成功率比带记忆的 VLA 高 40%。
CoRL 2026 Spotlight. Gives VLAs memory at several timescales: abstract long-term state (which recipe stages are done) plus short-term recall of recent events and occluded objects.CoRL 2026 Spotlight。为 VLA 提供多时间尺度记忆：长期记住抽象状态（如菜谱进行到哪一步），短期记住近期事件与被遮挡物体。
Safe task planning for embodied agents: keeps a long-term semantic graph of the scene so an LLM risk predictor can catch hazards outside the current view and trigger replanning. IS-Bench + real robot.具身智能体的安全任务规划：持续维护场景的长期语义图记忆，让 LLM 风险预测器发现视野外的隐患并触发重规划。在 IS-Bench 与真机上验证。
Brain-inspired memory for embodied agents (temporal, spatial, semantic, episodic) with a spatial knowledge graph. Reports +26.5 points for Qwen2.5-VL-72B on EmbodiedBench.受大脑启发的具身智能体记忆框架（时间、空间、语义、情景记忆），配合动态空间知识图谱。报告让 Qwen2.5-VL-72B 在 EmbodiedBench 上平均成功率提高 26.5 个百分点。
Persistent memory for mobile manipulators: a 3D scene graph updated with what the robot learns by acting (locked cabinets, object locations). Reports 80% progress over a 25-task, 3-hour run.移动操作机器人的持久记忆：在三维场景图中记录通过交互学到的知识（如柜子上锁、物体位置）。报告在连续 25 个任务、3 小时的仿真中任务进度达 80%。
Uses a VLM only at training time to pick task-relevant current and past information, distilled into a 'workspace token' the policy reads at deployment. Beats conditioning on curated past frames.仅在训练时用 VLM 挑出完成任务所需的当前与历史信息，蒸馏成策略部署时读取的“工作区 token”。效果优于直接输入精选历史帧。
CoRL 2026 Oral. Test-Time-Training policies scale visuomotor context to 8K timesteps, three orders of magnitude longer, without added inference latency, enabling one-shot imitation from human video.CoRL 2026 Oral。Test-Time-Training 策略把视觉运动上下文扩展到 8K 时间步，长出三个数量级且不增加推理延迟，由此支持从人类视频单次模仿等新能力。
CoRL 2026 Oral. Argues VLAs fail at dynamic tasks because one observation brings motion ambiguity and state aliasing, and adds learned temporal context to resolve both.CoRL 2026 Oral。指出 VLA 难以应对动态任务，是因为单帧观测带来运动歧义与状态混淆，并通过学习时序上下文同时解决这两个问题。
Always-on dual-arm robot: a π0.5 backbone with three light async modules that watch the live stream, track embodied state and recall its own past actions. Reports 45% vs 28% on the combined task.“永远在线”的双臂机器人：在 π0.5 主干上加三个轻量异步模块，分别监看实时流、跟踪本体状态、回忆自己过去的动作，两臂可并行执行。报告综合任务成功率 45%，最强基线 28%。
Finds that a few replay samples, where new and old observations look alike but need different actions, prevent most forgetting. Dropping just 10% of them raises forgetting more than 4.5x on LIBERO.发现回放缓冲中少数“观测相似但所需动作冲突”的样本（记忆锚点）承担了防遗忘的主要作用：仅剔除其中 10% 就会让 LIBERO 上的遗忘增加 4.5 倍以上。
CoRL 2026 Oral. DreamZero: a World Action Model on a pretrained video diffusion backbone, jointly predicting video and actions, so it learns skills from heterogeneous robot data without repetitive demonstrations.CoRL 2026 Oral。DreamZero：基于预训练视频扩散骨干的世界动作模型，联合预测视频与动作，可从异构机器人数据学习技能，无需重复示范。
CoRL 2026 Spotlight. Shows world models are a stealthy data-poisoning entry point: compromised policies can be deployed even when the ground-truth training data looks safe.CoRL 2026 Spotlight。展示世界模型可成为隐蔽的数据投毒入口：即使真实训练数据看似安全，部署的策略仍可能被破坏。
Lightweight world action model: compact video backbone with future supervision in a downsampled latent, and an action expert that pools several layers to predict chunks in one pass. Reports 0.44B trainable params, 72 ms.轻量世界动作模型：紧凑视频主干在降采样潜空间中做未来监督，动作专家融合多层特征一次前向预测动作块；报告可训练参数仅 0.44B，推理延迟 72 毫秒。
World-language-action model: one autoregressive backbone predicts text subtasks, subgoal images and actions; world prediction can be off at inference or on for test-time scaling. Reports 92.94% on RoboTwin 2.0 Clean.世界-语言-动作模型：单一自回归主干同时预测文本子任务、子目标图像与动作；推理时可关闭世界预测，也可开启做测试时扩展。报告在 RoboTwin 2.0 Clean 上成功率 92.94%。
Builds a robot foundation model on video instead of language: an open video planner trained on internet-scale human video makes zero-shot plans that are turned into robot actions. Model and dataset are released.以视频而非语言为主干构建机器人基础模型：在互联网规模人类视频上训练开源视频规划器，零样本生成视频计划再转为机器人动作；模型与数据集均已开源。
Adds episodic memory tokens to video-world-model policies so actions can depend on events no longer in view. Reports LIBERO-Mem success up from 5% to 42.5% and 75–80% on real memory tasks.为视频世界模型策略加入情景记忆 token，使动作能依赖已不在视野中的事件。报告 LIBERO-Mem 成功率从 5% 提升到 42.5%，真机记忆类任务约 75–80%。
CoRL 2026 Spotlight. A fully native 1-bit VLA with ternary weights, built on BitNet b1.58 2B4T, designed for edge devices instead of relying on post-hoc compression.CoRL 2026 Spotlight。原生 1-bit VLA，权重三值化，基于 BitNet b1.58 2B4T，面向边缘设备设计，而非依赖事后压缩。
EagleVLA: asynchronous VLA inference for low-power onboard chips such as Jetson Orin, tackling perception-execution misalignment and slow reaction; a CoRL 2026 Spotlight, formerly titled Jetson-PI.EagleVLA：面向 Jetson Orin 等低功耗机载设备的 VLA 异步推理方法，针对异步执行带来的感知-执行错位与反应慢两个问题；CoRL 2026 Spotlight，曾用名 Jetson-PI。
CoRL 2026 Spotlight. Instead of regenerating action chunks from scratch, keeps a persistent action buffer with staggered flow timesteps, so flow-based VLAs get more reactive while preserving long-horizon context.CoRL 2026 Spotlight。不再每次从头生成动作块，而是维护一个流时间步错开的持久动作缓冲区，让基于流的 VLA 更具反应性，同时保留长时域上下文。
A policy predicts each action and how long to hold it. Reports matching high-frequency control at under half the rate on an RC car and Unitree Go1, plus simulated VLA inference.策略同时预测动作及其保持时长。报告在遥控车、Unitree Go1 和仿真 VLA 上以不到一半的控制频率达到高频控制的性能。
CoRL 2026 Oral. A steerable generalist foundation model: follows diverse instructions in unseen scenes and reaches zero-shot cross-embodiment tasks such as folding laundry, with no task-specific finetuning.CoRL 2026 Oral。可引导的通用基础模型：在未见环境中执行多样指令，跨形态零样本完成叠衣等任务，无需任务专用微调。
CoRL 2026 Spotlight. Similar prompts can make VLAs behave very differently, so this framework interactively searches for language sequences that raise closed-loop task performance.CoRL 2026 Spotlight。语义相近的指令可能让 VLA 行为迥异，该框架通过交互搜索语言序列来提升闭环任务表现。
CoRL 2026 Spotlight. Training-free steering that makes Cartesian, embodiment-agnostic policies aware of the deployed robot's body, using joint-space guidance for constraints such as whole-body collision avoidance.CoRL 2026 Spotlight。无需训练的引导框架：用关节空间引导让与本体无关的笛卡尔策略感知实际机器人本体，处理全身避碰等约束。
CoRL 2026 Spotlight. A controlled study of zero-shot VLA transfer that isolates embodiment change from tasks, scenes and protocols, and separates strict zero-shot from looser definitions.CoRL 2026 Spotlight。对 VLA 零样本跨形态迁移的对照研究：将本体变化与任务、场景、流程隔离，并把严格零样本与宽松定义区分开。
CoRL 2026 Oral. Cross-embodiment VLA that treats human interaction traces as a shared language, pre-trained on UniHand-2.0: over 35,000 hours of multimodal data across 30 robot embodiments.CoRL 2026 Oral。跨形态 VLA，把人类交互轨迹当作通用“母语”，在 UniHand-2.0 上预训练：超过 3.5 万小时多模态数据，覆盖 30 种机器人形态。
Knife peeling as a test case for tasks with subjective quality: force-aware imitation first, then preference-based fine-tuning with a learned reward. Reports 100% success on cucumbers, apples and potatoes.以用刀削皮为例研究“质量难以量化”的精细操作：先做力感知模仿学习，再用学习到的奖励模型按人类偏好微调。报告在黄瓜、苹果、土豆上成功率 100%。
Smartphone app for robot-free policy iteration: shows the current policy's predicted trajectory in AR so users collect corrective data where it fails. Reports up to 2x sample efficiency over offline demos.基于手机的免机器人策略迭代系统：用 AR 显示当前策略的预测轨迹，用户可在策略薄弱处采集纠正数据。报告样本效率最高为纯离线演示的 2 倍。
Shows fixed rules for deriving labels from human feedback often disagree with what people mean, and learns these implications instead. Better preference estimates, incl. a real pizza-making study.发现从人类反馈推导未选动作标签的固定规则常与人的真实意图不符，改为学习这种“隐含含义”。偏好估计更准，并在真实做披萨协作实验中验证。
Shows VLAs can do implicit human-robot collaboration, but action chunks that cross task transitions make the robot help too early; an inference-time steering fix is tested in a 16-person assembly study.表明 VLA 可用于隐式人机协作，但跨越任务切换点的动作块会让机器人过早提供帮助；提出推理时引导方法缓解，并在 16 人装配用户研究中验证。
One VLA backbone runs as an independent copy on each robot, conditioned on its own view and a robot ID prompt, for decentralized team tasks. Reports a 64-point gain over from-scratch decentralized models.同一个 VLA 主干在每台机器人上独立运行，只依赖本机观测和机器人身份提示，实现去中心化多机协作。报告比从零训练的去中心化模型提升 64 个百分点。
TouchSafeBench: 2,940 Habitat 3.0 episodes testing whether VLMs can tell if a robot is colliding or about to collide with people or the scene. Best average Macro-F1 stays below 50%.TouchSafeBench：基于 Habitat 3.0 的 2940 个场景，测试 VLM 能否判断机器人正在或即将与人或环境碰撞。最佳平均 Macro-F1 仍低于 50%。
SocialNav-SUB: a VQA benchmark testing whether VLMs understand real social navigation scenes (spatial, temporal, social reasoning). The best VLM still trails a rule-based baseline and human consensus.SocialNav-SUB：检验 VLM 能否理解真实社交导航场景（空间、时序与社会推理）的视觉问答基准。最好的 VLM 仍落后于规则基线和人类共识。
Task-aware robot-to-human handovers: a VLM generates images of how a person would hold the object for the task, matched live to the person's hand. 83.3% of 12 users felt it understood the task better.面向任务的机器人递物：用 VLM 生成人在该任务下如何握持物体的图像假设，并与实时观测的人手姿态匹配。12 名参与者中 83.3% 认为它比已有方法更懂任务。
Proactive assistant robot: a VLM reads a state-machine memory of the user's routine to decide when to help, and a VLA executes. The memory is revised daily by reflection, without retraining.主动协助机器人：VLM 读取以有限状态机表示的用户日常流程记忆，决定何时提供何种帮助，再由 VLA 执行；记忆每天通过反思更新，无需重新训练。
Learns who owns which objects from everyday observation and keeps a long-term spatial memory refreshed by curiosity-driven exploration, so commands like 'find my bottle' work.从日常观察中学习“哪个物品属于谁”，并通过好奇心驱动的探索刷新长期空间记忆，从而执行“找我的水杯”这类个性化指令。
RL-trained AI coach that steps back as the learner improves, modelled as a game between learner and coach. A 33-person FPV drone-racing study shows faster human skill gains than prior AI coaching.用强化学习训练的 AI 教练：随学习者进步逐步减少辅助，把教练与学习者建模为非合作动态博弈。33 人第一视角无人机竞速研究显示，人类技能提升快于已有 AI 教练方法。
Humanoid gaze system: a VLM picks where to look and a Transformer-VAE trained on retargeted human recordings generates coordinated eye, neck and eyelid motion. Lowest joint error on a 2-second benchmark.人形机器人注视系统：VLM 选择注视目标，基于重定向人类录像训练的 Transformer-VAE 生成协调的眼、颈与眼睑运动。在 2 秒基准上关节误差最低。
When teaching data under-specifies some reward features, the robot targets preference queries at those features and explains what to look at. Up to 30.8% better than standard active preference learning with few queries.当示教数据对某些奖励特征监督不足时，机器人把偏好询问集中在这些特征上，并附带解释提示人该关注什么。少量询问下比标准主动偏好学习高最多 30.8%。
Studies whether single-robot teleop data can train decentralised coordinated box-pushing; tests with up to 40 robots show coordination is hard to learn just by watching other robots.研究能否只用单机遥操作数据训练去中心化的多机协同推箱；多达 40 台机器人的实验表明，仅靠被动观察其他机器人很难学会协调。
Fleet-scale offline-to-online RL that keeps post-training a generalist VLA from autonomous rollouts and human interventions. On 16 dual-arm robots and eight tasks it reports 95% average success.车队规模的离线到在线强化学习：利用自主执行与人工干预数据持续后训练通用 VLA。在 16 台双臂机器人、八个任务上报告平均成功率 95%。
Studies how value-function reliability drives offline-to-online RL for manipulation, with a history-conditioned value estimator. Reports 86% on millimeter-level chip insertion.系统研究价值函数可靠性如何影响操作任务的离线到在线强化学习，提出历史条件价值估计器。报告毫米级芯片插装成功率 86%。
Federated RL for heterogeneous clients that aggregates diffusion behavior priors with optimal transport rather than averaging policies, plus a low-variance DICE value baseline.面向异构客户端的联邦强化学习：不再平均策略，而是用最优传输聚合扩散行为先验，并加入低方差的 DICE 价值基线。
Uses an imitation policy's predicted action chunks as short task-space references for dense progress rewards in sparse-reward RL, plus guided pretraining. Better early sample efficiency and fewer safety violations.把模仿策略预测的动作块转成短时末端参考轨迹，为稀疏奖励强化学习提供稠密进度奖励，并用于引导预训练。提升早期样本效率、减少安全违规。
Agentic intervention model for human-in-the-loop RL: detects stalled exploration from value trends and steers the policy back to high-value states. Reports 57% fewer human interventions.用于人在回路强化学习的智能干预模型：根据价值变化趋势发现无效探索，并自动把策略拉回高价值状态。报告人工干预减少 57%，成功率提升 8.6%。
Human-in-the-loop RL that treats each intervention as a preference (human action over policy action) and corrects credit on suboptimal segments. Reports +24.5% average success and 1.3x faster convergence.人在回路强化学习：把每次人工干预视为偏好（人类动作优于策略动作），对次优片段做信用重新分配。报告五个真机任务平均成功率提升 24.5%，收敛快 1.3 倍。
Training-free progress reward: reads a video-language model's token probability that the instructed task is done, rather than asking it for a number. Introduces ManiRewardBench (130 tasks, four robots).免训练的进度奖励：不让视频语言模型输出数值，而是读取其判断“任务已完成”的 token 概率作为稠密奖励。并提出 ManiRewardBench（130 个任务、4 种机器人）。
Video comparator trained once on general videos that turns one successful demo into a dense progress reward, paying only for confident forward progress. Best RL success on 9 sim and 4 real tasks.只在通用视频上训练一次的视觉比较器，可把一条成功演示变成稠密的进度奖励，并只奖励置信度高的前进。在 9 个仿真和 4 个真机任务的强化学习中成功率最高。
Training-free reward editing: compares successful and failed rollouts in hindsight to penalize trap states that foundation-model rewards rate highly. Reports at least 3x the base policy's performance.免训练的奖励修正：事后对比成功与失败轨迹，惩罚被基础模型奖励高估、最终却导致失败的“陷阱状态”。报告性能至少是基础策略的 3 倍。
System for stable, sample-efficient RL fine-tuning of pretrained VLAs; reports 30/30 success on every task (e.g. plugging in string lights, pool shots) after ~19 minutes of robot data on average.稳定且样本高效的 VLA 强化学习微调系统；报告在插灯串插头、台球击球入袋等所有任务上均 30/30 成功，平均只需约 19 分钟在线机器人数据。
Uses 3D world generative models to build 100 diverse interactive scenes for RL fine-tuning of a VLA. Reports sim success from 9.7% to 79.8% and real-world success from 21.7% to 75%.用 3D 世界生成模型构建 100 个多样的可交互场景，对 VLA 做强化学习微调。报告仿真成功率从 9.7% 提升到 79.8%，真机从 21.7% 提升到 75%。
Fine-tunes an LLM with solver-based RL rewards to write executable mixed-integer programs for multi-modal motion planning (e.g. walk, then grasp), instead of answering directly.用求解器结果作为奖励对 LLM 做强化学习微调，让它为多模态运动规划（如先走再抓）编写可执行的混合整数规划代码，而不是直接给出答案。
Post-training recipe that turns an off-the-shelf VLM into a free-form language reasoner guiding a low-level policy: mid-train on expert reasoning traces, then rubric-based RL from offline action data.把现成 VLM 训练成用自由语言推理来指导底层策略的“机器人推理器”：先用专家推理轨迹做中训练，再用离线动作数据做基于评分规则的单步强化学习。
Kinesthetic teaching for real-world RL that keeps human guidance within the policy's kinematic limits, plus an impedance control stack. Reports 23–48% shorter cycle time vs HIL-SERL on industrial insertion.面向真实世界强化学习的示教框架：人手拖动示教时轨迹始终满足策略可执行的运动学限制，并配套阻抗控制栈。报告工业插装任务节拍时间比 HIL-SERL 缩短 23–48%。
Success Guided Sampling focuses massively parallel RL on task configurations at the edge of the policy's ability. With up to a million envs it solves quadruped terrain and assembly tasks; assembly transfers zero-shot.成功引导采样把大规模并行强化学习集中在策略能力边界附近的任务配置上。用最多约一百万个并行环境解决四足多地形与装配任务，装配策略零样本迁移到真机。
Agent harness that automates sim RL from environment setup to trained policy: staged specialist agents, executable gates, parallel trials and reusable knowledge. 16 tasks across 6 benchmarks, some moved to real robots.自动化仿真强化学习全流程的智能体框架：从环境搭建、奖励设计到调参训练，由分阶段的专职智能体、可执行检查点、并行试验与经验复用完成。覆盖 6 个基准 16 个任务，部分迁移到真机。
CoRL 2026 Spotlight. Model-based RL that uses a chunk world model as a differentiable training-time evaluator, targeting compounding error, inference latency and BPTT instability.CoRL 2026 Spotlight。将分块世界模型用作训练期可微评估器的模型式强化学习，针对累积误差、推理延迟与 BPTT 不稳定问题。
Action-to-action flow-matching policy conditioned on object-centric 'what' features and image-plane 'where' cues, improving robustness to distractors and object shifts while keeping few-step inference.以物体为中心的动作到动作流匹配策略：把观测拆成语义“是什么”特征与图像平面“在哪里”线索作为条件，在保持少步推理的同时提升对干扰物和位置偏移的鲁棒性。
Tests six flow-matching VLAs and reports that one-step Euler action generation largely matches multi-step sampling at scale — a useful check before adding special one-shot objectives to speed up control (CoRL 2026).在六个基于流匹配的 VLA 上实验，报告单步 Euler 生成动作在规模足够时与多步采样效果基本相当——在为提速引入专门的单步生成目标之前值得先看（CoRL 2026）。
CoRL 2026 Spotlight. Diffusion sequence prediction with selective re-noising, so segments of the horizon can be refined again using context from the rest.CoRL 2026 Spotlight。在扩散序列预测中选择性重新加噪，让时间跨度内的片段可以借助其余片段的上下文再次修正。
Fully open behavior-cloning stack: ABC-130K (3,500 hours, 130K episodes, 195 tasks), hardware, training code and a sim co-training recipe whose sim scores track real-world results.完全开源的行为克隆全栈：ABC-130K 遥操作数据集（3500 小时、13 万条、195 个任务），外加硬件、训练代码和仿真协同训练方案，其仿真评测结果与真机表现相关。
CoRL 2026 Spotlight. VLAs can ignore language and follow vision shortcuts from dataset bias. Introduces LIBERO-CF, a counterfactual benchmark for language following.CoRL 2026 Spotlight。VLA 可能忽略语言、沿用数据偏差形成的视觉捷径。论文提出反事实基准 LIBERO-CF，用于评测语言遵循能力。
CoRL 2026 Spotlight. Splits the generalization gap of sequential instruction-following policies into marginal, compositional and context-action shift, viewed through instruction-space coverage.CoRL 2026 Spotlight。从指令空间覆盖出发，把序列任务策略的泛化差距分解为边际偏移、组合偏移与上下文-动作偏移。
CoRL 2026 Oral. Audits LIBERO, CALVIN, SimplerEnv, RoboCasa and RoboTwin 2.0 against four failure modes, from shortcut solvability to creeping overfitting. LIBERO and CALVIN fail several diagnostics.CoRL 2026 Oral。用捷径可解、统计不显著、渐进过拟合、数据来源依赖四种失效模式审计 LIBERO、CALVIN、SimplerEnv、RoboCasa 和 RoboTwin 2.0；其中 LIBERO 与 CALVIN 未通过多项诊断。
For tasks people cannot teleoperate well, an LLM writes state-based guardrails that filter teleop and policy commands. Reports data-collection success rising from 0-10% to 70-90%.针对人难以遥操作完成的任务，由 LLM 编写基于状态的护栏来过滤遥操作与策略指令。报告数据采集成功率从 0–10% 提升到 70–90%。
Lets text-only LLMs do few-shot bimanual manipulation without fine-tuning by treating the arms as leader and follower agents. Reports 70.5% average success on 13 TWIN tasks.把双臂视为“主-从”两个智能体，让纯文本 LLM 无需微调即可通过上下文学习完成少样本双臂操作。报告在 TWIN 的 13 个任务上平均成功率 70.5%。
Diffusion policy that denoises one action in several coordinate frames at once with frame-specialised denoisers and fuses the results. Beats oracle single-frame choice on nine bimanual tasks; two real mobile tasks.扩散策略在多个坐标系中同步去噪同一动作：各坐标系有专门的去噪器，再把噪声预测融合回规范坐标系。在九个双臂任务上优于“最优单坐标系”选择，并在两个真实移动操作任务上验证。
Diffusion policy for whole-body mobile manipulation: a Fourier encoder turns per-pixel 3D points into tokens gated in by depth validity, and a hybrid decoder gives arm and base separate action subspaces.面向全身移动操作的扩散策略：傅里叶编码器把逐像素 3D 坐标变成几何 token，并按深度可靠性门控注入视觉基础模型特征；混合解码器为机械臂与底盘分配独立的动作子空间。
Mobile manipulation policy that infers motion intent to reweight perception and decodes base and arm actions with a coupled flow-matching head. Reports +23.6 to +33.0 points over DSPv2 on three tasks.移动操作策略：推断运动意图以动态分配感知注意力，并用显式建模底盘-机械臂交互的流匹配解码器输出动作。报告在三个任务上比 DSPv2 高 23.6–33.0 个百分点。
HiRe-MoMa fine-tunes a frozen mobile-manipulation policy in the real world by learning residual corrections, with only occasional human guidance.HiRe-MoMa 在真实环境中通过学习残差修正来微调冻结的移动操作策略，只需偶尔的人类引导。
One wearable both teleoperates a wheeled bimanual robot and records robot-free demos; arms are mechanically aligned, so no IK or retargeting. With a 10-episode robot-grounded seed, 100 robot-free episodes suffice.同一套穿戴设备既能遥操作轮式双臂机器人，也能在机器人不在场时录制示教；手臂与机器人机械同构，无需逆运动学或重定向。只需 10 条有机器人参与的种子数据，其余 100 条可脱离机器人录制。
Collects VLM-agent data by having teachers act through the agent's own skill interface, including where to look. Reports 71.76% on 170 held-out open-world mobile-manipulation episodes and 38% zero-shot on a real robot.让示教者直接通过智能体自身的技能接口（包括“往哪看”）执行动作来采集 VLM 智能体训练数据。报告在 170 个开放世界移动操作测试集上成功率 71.76%，零样本真机 38%。
Creative tool use (cut a cake with a plate): a VLM picks the object to use as a tool and the relevant parts, then vision foundation models ground those parts as 3D regions from one RGB-D image. No end-to-end training.创造性用工具（没有刀就用盘子切蛋糕）：VLM 选出可充当工具的物体及其相关部位，再由视觉基础模型从单张 RGB-D 图像把这些部位定位为 3D 区域。无需端到端训练。
Maps observations into a latent space organised by affordances (e.g. 'movable') rather than appearance, and adds new affordances when uncertain. Reports 94% inference accuracy, 15+ points above prior work.把视觉观测映射到按功能（如“可移动”）而非外观组织的潜在空间，并在不确定时自动发现新的可供性。报告推断准确率 94%，比已有方法高 15 个百分点以上。
Adds a learned, phase-conditioned affordance field to SE(3)-equivariant policies, anchoring observations and actions on the active region instead of the gripper frame. Improves MimicGen and multi-object real tasks.为 SE(3) 等变策略加入随阶段变化的可学习可供性场，以当前作用区域而非夹爪坐标系作为观测与动作的锚点。在 MimicGen 与多物体真机任务上都有提升。
Open-vocabulary 3D affordance detection: MLLM output tokens pull multi-scale geometry from a frozen 3D VQ-VAE through gated cross-attention before mask decoding. Reports about +8% mIoU.开放词汇 3D 可供性检测：多模态大模型的输出 token 在解码掩码前，通过门控交叉注意力从冻结的 3D VQ-VAE 中检索多尺度几何特征。报告 mIoU 提升约 8%。
VLM agent that plans, monitors and recovers by treating a VLA or world-action model as an interruptible tool it can stop and steer mid-rollout. Comes with RoboVoLo, a long-horizon open-vocabulary benchmark.VLM 智能体把 VLA/世界动作模型当作可随时中断的工具，在执行中途调度和纠正，从而完成规划、监控与恢复。配套发布开放词汇长程操作基准 RoboVoLo。
Zero-shot long-horizon manipulation: a VLM proposes and checks several generated video plans, actions follow object or hand flow, and execution is monitored with local corrections such as poking.零样本长程操作：VLM 提出并验证多条生成的视频计划，动作由物体流或手部流转换而来，执行中持续监控并做戳推等局部纠正。
Composes simple behaviours (semantic constraints from foundation models, keypoints from imagined video, tactile/force corrections) into one SE(3) pose command per step. 8 real tasks, incl. seating a GPU.把语义约束（基础模型）、想象视频中的关键点预测、触觉/力实时修正等简单行为在 SE(3) 上组合成每一步的位姿指令。8 个真机任务，包括把显卡插入 PCIe 槽。
Co-trains a bimanual policy on sim data (for robot-valid actions) and human videos (for real visuals), then fine-tunes on 80 real episodes per task. Reports 62.5% on held-out scenes, 53.7 points above real-only.用仿真数据（提供机器人可执行动作）和人类视频（提供真实视觉）联合预训练双臂策略，再在每任务 80 条真机数据上微调。报告在未见场景上成功率 62.5%，比纯真机数据高 53.7 个百分点。
Mobile dexterous grasping learned in sim: a whole-body controller does reaching and active vision, local RL experts the object, all distilled into one point-cloud policy; reports zero-shot real transfer.在仿真中学习移动灵巧抓取：全身控制器负责底盘-手臂协同到达与主动视觉，局部强化学习专家负责物体交互，二者蒸馏成一个基于点云的视觉策略；报告可零样本迁移到真实世界的数百个物体与场景。
Splits bimanual policy features into a swap-symmetric part (coordination that holds when arms trade roles) and a swap-antisymmetric part (role-specific behaviour), recomposed as residuals on each arm's features.把双臂策略特征分解为交换对称分量（两臂互换角色时不变的协调结构）与交换反对称分量（随角色变化的行为差异），再以残差形式加回各臂的表示。
Represents skills as short interaction segments of relative SE(3) motion between task objects, independent of scene layout and robot. Needs much less data than prior policies on RLBench and COLOSSEUM.把技能表示为任务物体之间相对 SE(3) 运动的短交互片段，与场景布局和机器人无关。在 RLBench 与 COLOSSEUM 上所需数据明显少于以往方法。
Contrastive RL struggles when contact changes the dynamics mode. Interaction-weighted Resampling samples around before/during/after contact, reporting a 19.8% average gain in sim, including robot air hockey.接触会改变动力学模式，使对比式强化学习难以处理操作任务。交互加权重采样在接触前、中、后阶段重点采样，仿真中平均提升 19.8%，任务包括机器人空气曲棍球。
Lets a policy trained only on static objects handle moving ones: demos are augmented by moving the object to a counterfactual future pose and warping the action chunk; a pose predictor supplies the target at test time.让只用静止物体演示训练的策略也能抓取运动物体：把物体平移到反事实的未来位姿并相应变形动作块来扩增数据，部署时由位姿预测器给出目标位置。
Learns a relational GNN heuristic over a simple symbolic model, with a motion planner revealing hidden geometric preconditions during training. Its top-ranked action is feasible in 94% of states (22% random).在简单符号模型上学习关系图神经网络启发式，训练时由运动规划器揭示隐藏的几何前提。其排名第一的动作在 94% 的状态下几何可行（随机选择为 22%）。
Makes dense VLM features spatially consistent by rebuilding each patch from language-labelled visual prototypes; improves open-vocabulary retrieval and 3D mapping. Also ships a compact 25M dense VLM for robots.通过由带语言描述的视觉原型重建每个图像块，让稠密 VLM 特征在空间上更一致，提升开放词汇检索与三维建图；并提供 25M 参数的小型稠密 VLM。
A large foundation model on a remote workstation gives high-level guidance while a light onboard Edge Adapter refines actions at high rate. Reports 40% higher success than baselines with delays up to 6 s.大模型在远端工作站提供高层指引，机载轻量 Edge Adapter 高频修正动作，两路异步运行并端到端微调。在通信延迟最高 6 秒的真实导航中，成功率比基线高 40%。
Trains a privileged-state expert, freezes its latent interface and action head, then fits a light adapter mapping frozen-VLM features of image + instruction into it: supervised alignment, not end-to-end RL.先用特权状态训练专家策略并冻结其潜空间接口与动作头，再训练轻量适配器把冻结 VLM 提取的图像与指令特征映射到该潜空间：用监督式对齐代替端到端策略优化。
One diffusion transformer jointly generates long-horizon actions and short-horizon latent visual foresight, avoiding multi-step rollouts. Supports image-goal, point-goal and other goal types in one model.用同一个扩散 Transformer 同时生成长时域动作和短时域潜空间视觉预测，避免多步自回归推演；一个模型支持图像目标、点目标等多种目标类型。
Social navigation where the VLM writes a costmap instead of choosing actions, checks upstream tracking, and a fast safety check guards against latency. Real-world tests.社交导航：让 VLM 生成代价地图而非直接决策，并校验上游跟踪，另加高频安全检查应对延迟；真机验证。
FlowPilot: a mapless, monocular-RGB sidewalk navigation policy pre-trained with anchored flow matching on fleet data, then tuned on a small set of human interventions by preference learning.FlowPilot：仅用单目 RGB、无需地图的人行道导航策略，先用锚定流匹配在大规模机器人车队数据上预训练，再用少量人工接管数据做偏好学习微调。
Multi-robot social navigation: robots fuse features to forecast occupancy maps with uncertainty, and a constrained multi-agent RL policy avoids regions predicted occupied or uncertain. Unity sim + real robots.多机器人社交导航：多机融合特征预测未来占据栅格及其不确定性，约束多智能体强化学习策略据此避开被预测占据或高度不确定的区域。Unity 仿真与真机验证。
For each candidate path, a 3D world model imagines several possible scene layouts behind occlusions; a future-aware value map combines them to pick waypoints. Tested on a Go2 quadruped.对每条候选路径，用 3D 世界模型想象遮挡后面多种可能的场景，再用面向未来的价值地图汇总这些想象来选择航点。在 Go2 四足机器人上测试。
Flips the problem: instead of adapting the robot, decides where to place a few fiducial markers so a planned trajectory can be executed reliably, using a shielded conditional diffusion model. Sim + hardware.反过来思考：不改机器人，而是决定在环境中哪里放少量视觉标记，使规划好的轨迹能可靠执行；方法为带安全约束采样的条件扩散模型。仿真与真机验证。
Navigates cluttered spaces by deciding, with VLM affordance reasoning, whether to swipe, pick up or avoid each blocking object. Zero-shot; reports +27 points over the best baseline on a real mobile manipulator.在杂乱环境中导航：用 VLM 推理每个挡路物体该推开、拿起还是绕开。零样本，报告在真实移动操作机器人上比最强基线高 27 个百分点。
Physics-based benchmark for navigating from buildings to streets in one episode: 200 scenes, 10,000 ObjectNav / VLN / PlaceNav episodes. Zero-shot VLAs succeed most; the modular method is safest.物理仿真基准，考察在一个连续回合中从室内走到街道：200 个场景、1 万个 ObjectNav / VLN / 地点导航回合。零样本测试中端到端 VLA 成功率最高，模块化方法最安全。
Adds simulated odour transport to Habitat-Sim using Fire Dynamics Simulator, giving a smell-and-vision navigation benchmark for finding invisible gas leaks and early-stage fires, with a multimodal RL baseline.用 Fire Dynamics Simulator 为 Habitat-Sim 加入气味传输仿真，构成视觉加嗅觉的导航基准，任务是定位不可见的气体泄漏和早期火源，并提供多模态强化学习基线。
An image generation model reads a language command and paints the goal and traversability masks onto a bird's-eye-view image; cross-view localization curbs drift. A quadruped finishes a 285 m outdoor route.用图像生成模型理解语言指令，在鸟瞰图上标出目标与可通行区域，并以跨视角定位抑制里程计漂移。四足机器人借此完成 285 米的户外长距离导航。
Navigation where goals change mid-episode without notice: memory is split into global/local spatial and long/short temporal beliefs, with goal-conditioned rewards. Tested on a Spot quadruped.研究回合中目标会无提示地改变的导航：把记忆分解为全局/局部空间与长/短时序信念，并配合目标条件奖励。在 Spot 四足机器人上实机验证。
Has a VLM pick among sparse 'points of interest' (frontiers, likely targets, stairs) tied to waypoints and raw views; planners do the motion. The checkable choices allow RLVR on local VLMs without CoT labels.让视觉语言模型在稀疏的“兴趣点”（探索前沿、疑似目标、楼梯等，每个绑定航点与原始视角）中做选择，底层规划器负责运动。这种可验证的决策使其能用 RLVR 训练本地 VLM，无需人工思维链标注。
Labels which frontier explains a human demonstrator's next moves, then trains a VLM to choose among BEV-memory candidates. Candidate-level interface transfers zero-shot to wheeled, quadruped and humanoid robots.通过前瞻人类示教，标注最能解释其后续搜索方向的前沿，再训练 VLM 在 BEV 记忆中的候选点之间做选择。这种候选级接口零样本迁移到轮式、四足和人形机器人。
Scores delivery-robot navigation by money: cost, revenue and break-even, using Isaac Sim plus public filings and injury data. Reports that none of seven baselines is economically viable yet.以经济账评估配送机器人导航：在 Isaac Sim 中结合公开财报与伤害数据计算成本、收入与盈亏平衡。报告七种基线方法目前都无法盈利。
Decouples spatial estimation from MLLM planning via an interactive metric world map plus counterfactual reasoning. Reports zero-shot 48.8% SR on R2R-CE and 42.2% on RxR-CE; runs on a TurtleBot 4 and a drone.通过可交互的度量世界表示把空间状态估计与多模态大模型的语义规划解耦，并引入反事实推理。报告零样本 R2R-CE 成功率 48.8%、RxR-CE 42.2%，并在 TurtleBot 4 与自制无人机上实机部署。
Calibrates VLM semantic maps with online conformal prediction so each object class gets a safety margin with coverage guarantees as the vocabulary grows. Reports 61.6% fewer semantic collisions on MP3D.用在线共形预测校准 VLM 语义地图，使不同类别物体获得有覆盖率保证的安全距离，且支持词表扩展。报告在 MP3D 上语义碰撞减少 61.6%，四足真机覆盖率 90.73%。
Uses calibrated barrier risk (CVaR along the MPC rollout) as an early warning to switch to a safer planning mode. Reports 83.3% fewer collisions than a safety-filter-only baseline.以校准后的障碍函数风险（MPC 预测轨迹上的 CVaR）作为预警信号，及时切换到更安全的规划模式。报告碰撞比仅用安全滤波器的基线减少 83.3%。
Joint task and motion planning for up to nine robots among dense obstacles: waypoints parameterize motion, and a curriculum with modified RLVR passes motion-planner feasibility back to the task planner (BoxNet3D-OBS).面向密集障碍中多达九台机器人的任务与运动联合规划：用航点参数化运动轨迹，并以课程式训练和改进的 RLVR 将运动规划器的可行性反馈传回任务规划器（BoxNet3D-OBS 基准）。
First autonomous vine (soft growing) robot: 19 cameras along its body feed an end-to-end visuomotor policy learned from demos, which steers through branches, up slopes and into confined spaces.首个自主藤蔓（软体生长）机器人系统：沿机身分布的 19 个摄像头为从演示学到的端到端视觉运动策略提供反馈，能穿过分叉结构、爬坡并进入狭窄空间。
RL post-training for navigation diffusion policies with perturbation-based exploration and group-normalized Q-scores, across embodiments. Reports sim success from 61.20% to 84.28% and real hard cases from 10% to 65%.面向导航扩散策略的强化学习后训练：保留预训练先验的扰动探索 + 组归一化 Q 分重加权匹配，跨多种本体训练。报告仿真成功率从 61.20% 升到 84.28%，真机困难场景从 10% 升到 65%。
Visual-navigation policies say where to go but not how far. PathLift reads sensor depth (RGB-D, stereo or LiDAR) and fits per-waypoint magnitude and lateral residuals, turning unitless paths metric in one pass.视觉导航策略只给方向、不给距离。PathLift 读取传感器深度（RGB-D、双目或激光雷达），为每个航点拟合长度与横向残差，一次前向就把无单位轨迹变为度量路径。
Treats camera motion as an action: given an image and a language intent ('look behind the box'), predict the next relative camera pose. Supervision is mined from egocentric video by hindsight labeling.把相机运动当作一种动作：给定当前图像和语言意图（如“看看箱子后面”），预测下一视角的相对相机位姿。监督信号通过事后标注从第一人称视频中挖掘。
Aligns two RGB-only egocentric captures of a room taken at different times: builds object-level 3D scene graphs with foundation models and lets a VLM match objects across them. No training or depth sensor.对齐同一房间在不同时间的两段纯 RGB 第一人称录像：用基础模型构建物体级 3D 场景图，再让 VLM 跨两段录像匹配物体实例。无需训练或深度传感器。
Localizes an RGB camera on a floorplan online, querying the map directly and refining poses with diffusion instead of building an offline database. Reports SOTA single-frame and sequential results in real time.直接在平面图上在线定位 RGB 相机：实时查询地图并用扩散模型精化位姿，无需离线数据库。报告单帧与序列定位均达最优且实时运行。
Feature matcher that refines correspondences recurrently in continuous coordinates instead of on fixed grids; reports higher accuracy than Efficient LoFTR at over 10x its speed on embedded hardware.在连续坐标中循环细化对应点、不再依赖固定网格的特征匹配方法；报告精度高于 Efficient LoFTR，在嵌入式硬件上速度超过其 10 倍。
Restructures deep feature matchers so cross-image interaction happens only in the last layers; per-image context blocks are cached offline. Reports over 2x lower compute, plus 10-14% from entropy-guided token reduction.重构深度特征匹配器，把跨图像交互推迟到最后几层，单图上下文模块可离线缓存。报告计算量降低一半以上，熵引导的 token 裁剪再省 10–14%。
Relative 6-DoF pose between two image groups (relocalization, multi-camera rigs): keeps a multi-view foundation model frozen and adds ~32M trainable parameters (resampler, cross-group bridge, pose head).估计两组图像之间的相对 6 自由度位姿（跨序列重定位、多相机里程计）：保持多视图基础模型冻结，只增加约 3200 万可训练参数（重采样器、跨组桥接模块、多帧位姿头）。
30M-parameter monocular depth model for transparent and reflective scenes, trained to undo depth-sensor bias with a small rendered set. Reports beating 300M-scale and multi-view models on transparent benchmarks.面向透明与反光场景的 3000 万参数单目深度模型，用少量渲染数据纠正真实深度传感器带来的监督偏差。报告在透明场景基准上超过 3 亿参数级和多视图模型。
Turns sparse LiDAR intensity into RGB-like video so camera models work in the dark. Reports better segmentation and lane detection than LiDAR baselines and real cameras across lighting.把稀疏 LiDAR 强度转换为类 RGB 视频，让相机模型在黑暗中也能用。报告在不同光照下，分割与车道线检测优于 LiDAR 基线和真实相机。
SLAM that keeps a functional scene graph (which handle opens which drawer) as online state, and uses that topology to help loop closure in repetitive scenes. Reported faster than offline methods at similar accuracy.把功能场景图（如哪个把手开哪个抽屉）作为 SLAM 的在线状态持续维护，并利用其拓扑辅助重复场景的回环检测。报告比离线方法更快且精度相当。
Online 3DGS mapping that integrates each Gaussian over the exposure time analytically to handle motion blur, and prunes by a Hessian score. Matches offline deblurring 3DGS at 15-150x the speed.在线 3DGS 建图：对每个高斯在曝光时间内解析积分以处理运动模糊，并用 Hessian 得分剪枝控制数量。渲染质量与离线去模糊 3DGS 相当，速度快 15–150 倍。
Self-supervised foot-contact estimation from joint encoders and IMU only, fed as a stance belief into a Kalman filter. Reports the lowest leg-odometry error among the compared baselines.仅用关节编码器与 IMU 自监督估计足端接触，并作为支撑置信度输入卡尔曼滤波器。报告腿式里程计误差在对比方法中最低。
Neural inertial odometry whose mean and covariance transform correctly under any IMU mounting rotation; Last-Frame Alignment makes sensor-frame training match world-frame. TLIO ATE: 2.018 m vs 76.389 m for EqNIO.神经惯性里程计：输出的均值与协方差在任意 IMU 安装旋转下都按张量规律变换；Last-Frame Alignment 使传感器坐标系训练等价于世界坐标系训练。在 TLIO 上 ATE 为 2.018 米，EqNIO 为 76.389 米。
Treats spoken descriptions ('my bag is on the table') as sensor readings: a model outputs a calibrated spatial distribution that is fused with onboard perception in one belief map. VLA-3D benchmark + real robot.把自然语言描述（如“包在桌上”）当作传感器观测：模型输出校准过的空间概率分布，并与机载感知融合到统一的信念地图中。在 VLA-3D 基准和真机上验证。
CoRL 2026 Oral. Argues simulation can be enough: large, diverse synthetic data gives zero-shot real transfer for static and mobile manipulation. MolmoBot-Engine is an open procedural data pipeline.CoRL 2026 Oral。挑战“仅靠仿真不够”的观点：足够大规模、多样的合成数据即可零样本迁移到真实静态与移动操作。MolmoBot-Engine 是开源的程序化数据管线。
Infinigen-Articulated: procedural generators for 18 articulated object categories plus Blender tools and an export pipeline into common robot simulators, with physical properties.Infinigen-Articulated：涵盖 18 类常见铰接物体的程序化生成器，附带 Blender 工具，并可连同物理属性导出到常用机器人仿真器。
Synthetic data engine that turns one real mobile-manipulation demo into many: Gaussian-splat scenes, rearranged contact segments chained by whole-body planning, and generated worlds from photos. Also cross-embodiment.合成数据引擎：把一条真实移动操作演示扩展成大量数据——高斯泼溅重建场景、重排接触片段并用全身运动规划串联、再在由照片生成的 3D 世界中合成。也支持跨本体。
CoRL 2026 Oral. Measures diversity of imitation datasets directly from demonstrations, with entropy defined through a signature kernel, so no model has to be trained first.CoRL 2026 Oral。直接从示范数据度量模仿学习数据集的多样性：基于签名核定义熵，无需先训练模型。
Mobile-manipulation agent with two loops: a VLM plans and composes flow-matching skills at run time, and an outer loop mines rollouts to find and fine-tune skills. Reports +22.5 pts over π0.5 on BEHAVIOR-1K.双循环移动操作智能体：内循环由 VLM 规划并组合流匹配技能，外循环从部署数据中自动发现并微调技能。报告在 BEHAVIOR-1K 上比 π0.5 高 22.5 个百分点。
Agentic generator of diverse, feasible manipulation tasks for data collection, with self-reflection and human-verified memory; 18k real trajectories collected. VLAs trained on its data generalize better.为数据采集自动生成多样且可执行的操作任务的智能体框架，结合自我反思与人工验证记忆；据此采集 1.8 万条真机轨迹，用其训练的 VLA 成功率与泛化更好。
Harness that lets coding agents improve real robot policies on their own: automatic reset and verification, parallel rollouts and agent-driven code changes. Reports 99% success on tasks like fastening a zip tie.让编程智能体自主改进真实机器人策略的框架：自动复位与结果验证、多机并行评测，并由智能体分析日志、修改训练代码。报告在扎扎带等灵巧任务上成功率 99%。
No end-to-end training: a VLM coding agent writes robot programs from a library of perception, planning and control tools in a plan-act-replan loop; reports matching or beating VLAs zero-shot on tabletop tasks.不训练端到端策略，而让 VLM 作为编程智能体，调用感知、规划与控制工具库编写机器人程序，按“规划-执行-重规划”循环运行；在桌面任务上零样本达到或超过 VLA 模型，并可完成移动操作。
DeepCORD unrolls a parallel accelerated Riemannian optimizer and learns a self-supervised policy that tunes solver parameters by phase and communication status. Tested on SE(3) pose graphs and SL(4).DeepCORD 把并行加速的黎曼优化器展开为可微迭代，并自监督学习一个依据优化阶段与通信状态调节求解参数的策略。在真实 SE(3) 位姿图与 SL(4) 子图对齐上测试。
Discrete-time neural HJ reachability without PDE gradients: bang-bang structure and gradient-free value probes give teacher actions, turning adversarial actor learning supervised; a windowed curriculum adds stability.无需 PDE 梯度的离散时间神经 HJ 可达性：利用最优安全干预的 bang-bang 结构和免梯度的价值探测构造教师动作，把对抗式策略学习变成监督学习；分窗口的时间课程稳定训练。
Learns neural CBFs from an HJ value approximation; with control-affine dynamics and polytope inputs the optimum is at a vertex, so GPU vertex tree search makes labels. 15 systems + a robot among pedestrians.通过近似 HJ 价值函数学习神经控制障碍函数；在控制仿射动力学和多面体输入集下最优解位于顶点，于是用 GPU 并行顶点树搜索生成监督点。在 15 个系统和行人避让真机实验中验证。
A causal transformer predicts a rollout (in place of a model) and a residual fix to a nominal controller; a barrier critic scores CBF violations along the rollout, so safety looks ahead like MPC.因果 Transformer 预测未来推演（代替动力学模型）并给名义控制器输出残差修正；障碍评估器沿推演检查 CBF 约束违例，从而像 MPC 一样前瞻安全，却无需在线求解优化。
Flow-matching planner whose sampling steps follow the execution dynamics, so safety (high-order CBFs) is enforced only on the executed step. Lower latency and better safety on navigation, racing and locomotion.流匹配规划器让采样过程与系统执行动力学对齐，只需对实际执行的一步施加高阶控制障碍函数约束。在导航、竞速和运动任务上降低规划延迟并提升安全性。
Distributed depth sensors on the robot form an obstacle-aware point cloud; a flow-matching policy trained on 100K+ simulated safety-critical trajectories plans collision-free motion and transfers zero-shot to hardware.机器人身上的分布式深度传感器汇成障碍感知点云；在 10 万多条仿真安全关键轨迹上训练的流匹配策略据此生成无碰撞运动，零样本迁移到真机。
Learns an invertible map plus convex polytopes in its latent space, so convex-set planners work in high-DoF spaces and paths decode back feasible. Tested from 2D to 14-DoF; no false positives after refinement.联合学习可逆映射及其潜空间中的凸多面体并集，使基于凸集的规划可用于高自由度空间，解码回原空间的路径仍可行。在 2D 至 14 自由度场景测试，细化后未观察到误报。
Replaces the sampling noise in diffusion's backward process with an optimized correction, so reachability, collision or trackability constraints hold at sampling time without retraining.把扩散反向过程中的采样扰动替换为经优化得到的修正，从而在采样时满足可达性、避碰或可跟踪性约束，无需重新训练扩散模型。
Theory: gives conditions under which sampled reachable sets are accurate in Hausdorff distance, and shows the sample need grows exponentially in both state dimension and horizon, with a matching lower bound.理论工作：给出采样可达集在 Hausdorff 距离意义下准确的条件，并证明所需样本数随状态维度和时间跨度都呈指数增长，且给出相应下界。
Uses constrained IK, where solution manifolds are known exactly, to test what geometry diffusion models learn. On UR5 and Franka the intrinsic dimension read from the score matches the analytic degrees of freedom.以解流形可精确知道的约束逆运动学为试验场，检验扩散模型学到的几何结构。在 UR5 和 Franka 上，从得分函数估计的内在维度与解析自由度一致。
Diffusion transformer that treats each robot as a token and generates coordinated multi-robot trajectories in one pass, meeting safety and connectivity constraints; trained across team sizes.扩散 Transformer 把每台机器人表示为一个 token，一次前向生成满足安全与连通约束的多机协同轨迹；可在不同团队规模的数据上混合训练。
Splits a VLA into a key-action model (next subgoal) and a motion model (dense actions toward it) in a shared camera-centric action space across datasets and robots. Reports 91% on SimplerEnv Google Robot visual matching.把 VLA 拆为关键动作模型（预测下一个子目标）与运动规划模型（生成通往它的稠密动作），并统一到以相机为中心的动作空间。在 SimplerEnv Google Robot 视觉匹配上报告 91%。
Runtime wrapper for frozen VLAs: conformalized quantile regression ranks sampled actions, and a Mahalanobis anomaly score on latent states warns about abnormal execution. No retraining.冻结 VLA 的运行时外壳：用共形分位数回归为多个候选动作打分择优，并在潜在状态上计算马氏距离异常分，预警异常执行。无需重训。
Adds a residual dynamics network and a learned Q critic to iLQR MPC, with a GPU batched iLQR that solves thousands of problems in parallel, for model-based RL when contacts are poorly modeled.在 iLQR 模型预测控制中加入残差动力学网络和学习得到的动作价值评估器，并实现 GPU 批量 iLQR 并行求解数千个问题，用于接触难以建模时的基于模型强化学习。
For assistive devices with costly user trials: a forward model predicts user outcomes from control policies and drives active queries to learn each user's reward, then picks the final policy without more trials.面向用户试验代价高的辅助设备：先训练前向模型从控制策略预测用户结果，用它主动选择查询来学习个体奖励，再在不增加试验的情况下选出最终策略。
Mines a Signal Temporal Logic constraint from demos (e.g. social-navigation comfort) by evolutionary search to shield policies, and runs an evolutionary PDDL planner where a local LLM proposes and repairs plans.通过进化搜索从示教数据中挖掘信号时序逻辑约束（如社交导航舒适度）来为策略加安全护盾，并用进化式 PDDL 规划器让本地部署的大模型提出和修复计划。
Discovers what a given robot can actually learn: a GFlowNet builds symbolic task skeletons, rewarded by how well low-level policies learn them in sim, then the tasks are distilled into one language-conditioned policy.自动发现某台机器人“学得会”的任务：用 GFlowNet 生成符号化任务骨架，以仿真中底层策略的学习效果作为奖励，再把发现的任务蒸馏进一个语言条件策略。
Picks at runtime which low-level policy to run and with which goal, using a learned energy per policy. New policies join by training their own energy model, with no retraining of others. Deployed on a real humanoid.运行时为每个底层策略学习能量函数，据此选择执行哪个策略、给它什么目标。新增策略只需训练自己的能量模型，无需重训其他策略或中央调度器；已部署到真实人形的移动操作任务。
With a limited budget for re-estimation and replanning under drifting dynamics, a dynamic-regret bound gives an online rule for when to update. Tested on simulated rover slip and a Crazyflie quadrotor.在动力学漂移、重估与重规划次数有限时，用动态遗憾界推导出何时更新的在线规则。在模拟的打滑漫游车和 Crazyflie 四旋翼上验证。
LLM-driven evolutionary search over mode-switching controllers written as programs, so cheap modes run in easy task phases. Reports designs beyond the Pareto front of one-size-fits-all controllers.用 LLM 驱动的进化搜索，把多模式切换控制器写成可解释程序并迭代优化，使简单阶段使用低开销模式。报告找到的设计超越单一结构控制器的帕累托前沿。
Tests whether agent architecture search, so far used on text agents, transfers to perceptual embodied agents; AgentCanvas makes perception, memory, planning and action wiring searchable.研究原用于文本智能体的智能体架构搜索（AAS）能否迁移到有感知的具身智能体；提出 AgentCanvas，把感知、记忆、规划与动作模块的组合方式变成可搜索的设计空间。
ASIMOV v2: physical-safety benchmark built from real injury narratives turned into generated images and videos, testing whether foundation models notice danger and intervene, plus safety-reasoning post-training.ASIMOV v2：把真实伤害事故描述转成生成的图像与视频，构建可持续扩展的物理安全基准，测试基础模型能否察觉危险并干预；并提出安全推理后训练方法。
Tue 11/10, 8:30–9:00 AM · Hilton Austin · 9 papers
CoRL 2026 Spotlight. A large-scale bimanual dexterous grasp dataset, built with a two-stage synthesis pipeline, together with a learning-based grasp generation framework.CoRL 2026 Spotlight。大规模双手灵巧抓取数据集（由两阶段合成流程生成），并配套基于学习的抓取生成框架。
CoRL 2026 Spotlight. Augments contact-rich teleoperation with corrective behaviors learned in simulation, using a simple kNN surrogate of the human operator.CoRL 2026 Spotlight。为富接触遥操作叠加在仿真中学到的纠正行为，并用简单的 kNN 模型替代人类操作员进行训练。
CoRL 2026 Spotlight. A pipeline that turns egocentric human manipulation videos into robot training data, testing whether this helps VLA pretraining at scale.CoRL 2026 Spotlight。把第一视角人类操作视频转换为机器人训练数据的流程，并检验其能否大规模助力 VLA 预训练。
CoRL 2026 Spotlight. Adapts frozen generative policies from human interventions in latent space, as a cheaper alternative to mass data collection or online RL on hardware.CoRL 2026 Spotlight。在隐空间中根据人类干预适配冻结的生成式策略，替代大规模采集或实机在线强化学习。
Generates dexterous grasps as 3D hand keypoints with a plain Transformer flow model, with no contact losses or refinement; reports 76.3% success on Dexonomy and 0.032 s per grasp.以 3D 手部关键点表示灵巧抓取，用普通 Transformer 流模型生成，无需接触损失或后处理优化；报告 Dexonomy 成功率 76.3%，每次抓取 0.032 秒。
Robots that never trained together share value functions, so a robot finishing a task leaves the workspace in a state the next robot prefers; a KNN uncertainty filter blocks overconfident values. Up to +48.5 pts.从未共同训练的机器人通过共享价值函数来协作，前一个机器人会把工作空间留在后一个机器人偏好的状态；KNN 不确定性过滤避免分布外的过度自信。成功率最高提升 48.5 个百分点。
Tue 11/10, 9:00–9:30 AM · Hilton Austin · 3 papers
TraCS adds interpretable, probabilistic traffic-rule logic to black-box motion predictors, generated from natural-language regulations; reports consistent gains on Argoverse 2. Oral.TraCS 将由自然语言交规生成的可解释概率逻辑加到黑盒运动预测模型上；报告在 Argoverse 2 上稳定提升。Oral 论文。
CoRL 2026 Oral. Makes flow matching policy gradients work for robot control, bypassing action likelihoods, so policies need not be Gaussian. Applied to legged locomotion, humanoid motion tracking and manipulation.CoRL 2026 Oral。让流匹配策略梯度可用于机器人控制：绕过动作似然计算，策略不再限于高斯分布。应用于足式运动、人形动作跟踪与操作。
CoRL 2026 Oral. Adapts a VLA to expose a compact RL token, an interface for sample-efficient online RL, so a pretrained policy gains precision and speed from a few hours of real-world practice.CoRL 2026 Oral。让 VLA 输出紧凑的 RL token 作为在线强化学习接口，预训练策略只需几小时真实练习即可提升精度与速度。
Tue 11/10, 10:30–11:00 AM · Hilton Austin · 3 papers
CoRL 2026 Oral. Trains a VLA on over 20,854 hours of action-labeled egocentric human video to test whether large-scale human data can support fine, high-DoF dexterous manipulation.CoRL 2026 Oral。用超过 20,854 小时带动作标注的第一视角人类视频训练 VLA，检验大规模人类数据能否支撑高自由度精细灵巧操作。
CoRL 2026 Oral. Generates executable dexterous grasps from free-form language, inferring the function a task implies from multi-view images and geometry, supported by the OpenDexVerse data.CoRL 2026 Oral。根据自由语言生成可执行的灵巧抓取：结合多视角图像与几何推断任务所需的功能，由 OpenDexVerse 数据支撑。
CoRL 2026 Oral. Asks what dexterous play pretraining actually needs before precise assembly, a contact-rich, sparse-reward task that is hard for both imitation and direct RL.CoRL 2026 Oral。探究灵巧“玩耍”预训练究竟需要什么，才能胜任精密装配这类富接触、稀疏奖励、模仿与直接 RL 都难以解决的任务。
Tue 11/10, 11:00–11:30 AM · Hilton Austin · 3 papers
CoRL 2026 Oral. NEXT estimates external joint torques without force sensors, training in 1 minute from 10 minutes of free-motion data, enabling force-feedback teleoperation and better policy learning on low-cost arms.CoRL 2026 Oral。NEXT 无需力传感器即可估计外部关节力矩，用 10 分钟自由运动数据训练 1 分钟，为低成本机械臂带来力反馈遥操作与更好的策略学习。
DISCO runs progressive-widening MCTS in GPU-parallel simulation, with a learned policy proposing continuous actions, then distills plans into a policy-value model used without search. Reports wins over BC, RL and search.DISCO 在 GPU 并行物理仿真中用渐进展宽 MCTS 搜索长程重排计划，由学习的策略提出连续动作、价值函数打分，并周期性把搜索结果蒸馏回策略-价值模型；部署时无需仿真搜索。报告在仿真与真机重排任务上优于行为克隆、强化学习与搜索基线（CoRL 2026 Oral）。
CoRL 2026 Oral. Splits one tactile stream into three temporal roles, persistent states, abrupt contact events and future consequences, instead of a single fusion path in chunk-based insertion policies.CoRL 2026 Oral。不再用单一融合路径，而是把同一触觉流拆成三种时序角色：持续接触状态、突发接触事件与未来接触后果，用于分块式插接策略。
Tue 11/10, 11:30 AM–12:00 PM · Hilton Austin · 3 papers
CoRL 2026 Oral. A fully virtual pipeline for humanoid loco-manipulation data: grounds video world models in 3D assets and simulator-ready scenes, with no teleoperation or physical setups.CoRL 2026 Oral。全虚拟的人形移动操作数据管线：把视频世界模型落到 3D 资产与可仿真场景中，无需遥操作或搭建实体环境。
CoRL 2026 Oral. Three-stage curriculum for humanoid soccer shooting: a single human-kick reference serves as a scaffold, bridging stable motion tracking and goal-driven RL that struggles to discover kicks.CoRL 2026 Oral。人形足球射门的三阶段课程：以单条人类踢球参考为脚手架，兼顾动作跟踪的稳定性与任务奖励 RL 的目标适应能力。
Rather than distilling per-motion experts or brute-forcing one RL policy, trains a humanoid tracking policy jointly with one differentiable world model of shared dynamics, fit from simulator logs (CoRL 2026 oral).SuperMimic 不再蒸馏逐动作专家或用大算力硬训单一 RL 策略，而是把人形机器人动作跟踪策略与一个从仿真日志拟合的、共享动力学的可微世界模型联合训练（CoRL 2026 口头报告）。
Tue 11/10, 2:00–2:30 PM · Hilton Austin · 9 papers
Helps RL explore contact-rich tasks by resetting episodes with a contact-aware constrained state sampler, covering diverse contact modes; trains universal in-hand and whole-body policies.用感知接触的约束状态采样器重置回合，覆盖多样接触模式，帮助 RL 探索接触丰富任务；可训练通用的手内和全身操作策略。
CoRL 2026 Spotlight. Addresses impedance gain selection for articulated tasks such as doors, bolts and drawers, where neural policies may violate kinematic constraints.CoRL 2026 Spotlight。针对开门、拧螺栓、拉抽屉等铰接任务解决阻抗增益选择问题，这类任务中神经策略可能违反运动学约束。
CoRL 2026 Spotlight. Studies why VLA models struggle in contact-rich manipulation and separates two failure modes: precision failures, where policies struggle to produce small corrective actions, and force failures.CoRL 2026 Spotlight。分析 VLA 模型为何难以完成接触丰富的操作任务，并区分两类失败模式：难以输出接触时所需微小修正动作的精度失败，以及力失败。
CoRL 2026 Spotlight. Dual-branch generative framework that extracts clean shading and marker images at once from one marker-based tactile frame, avoiding serial mask-based inpainting.CoRL 2026 Spotlight。双分支生成框架，可从单帧带标记点的触觉图像同时分离出干净的明暗图与标记图，避免串行掩码修复。
CoRL 2026 Spotlight. Fingertip sensing that combines binocular RGB cameras with a compliant contact surface, giving feedback from pre-contact approach through post-contact control.CoRL 2026 Spotlight。指尖传感将双目 RGB 相机与柔顺接触面结合，从接触前接近到接触后控制全程提供反馈。
LIFT: grafts a reactive force-driven action expert onto a pretrained, vision-driven VLA, keeping its output unchanged at initialization, so post-training gains contact reactivity; CoRL 2026 Spotlight.LIFT：在以视觉为主的预训练 VLA 旁接入由力觉驱动的反应式动作专家，初始化时保持原策略输出不变，使后训练获得接触反应能力，应对遮挡、深度歧义等接触阶段难题；CoRL 2026 Spotlight。
CoRL 2026 Spotlight. A module for contact-rich motion transfer across sim-to-real gaps, payloads and robot instances, as an alternative to domain randomization or system ID.CoRL 2026 Spotlight。面向富接触动作迁移的模块，应对仿真到现实、负载与个体差异，用于替代域随机化或系统辨识。
CoRL 2026 Spotlight. Brings contact into image-conditioned policies by rendering it, targeting the gap left by approaches that need tactile hardware, separate tactile encoders or non-image backbones.CoRL 2026 Spotlight。通过渲染接触信息把触觉带入基于图像的视觉运动策略，针对现有方法需要专用触觉硬件、独立触觉编码器或非图像主干，难以与预训练 2D 视觉表征策略兼容的问题。
Tue 11/10, 2:30–3:00 PM · Hilton Austin · 9 papers
MoMa-SG builds 3D scene graphs that encode how objects articulate, from RGB-D sequences, and adds the Arti4D-Semantic dataset; demonstrated on a quadruped and a mobile manipulator.MoMa-SG 从 RGB-D 序列构建包含物体铰接方式的 3D 场景图，并发布 Arti4D-Semantic 数据集；在四足和移动操作机器人上演示。
Refines cloth actions online from one RGB image: a real-to-sim module estimates the cloth state, then simulator rollouts with prior-guided MPPI correct the policy's trajectory.仅凭单张 RGB 在线修正布料动作：先由 real-to-sim 模块估计布料状态，再用仿真器并行推演和先验引导的 MPPI 修正策略轨迹。
PGRD simulates deformables with an optimizable spring-mass backbone plus learned residuals, reporting better accuracy than either alone; used for MPC and action-conditioned video prediction.PGRD 以可优化的弹簧质点模型为骨架，加上学习得到的残差来仿真可变形物体，报告精度优于单用任一方法；可用于 MPC 和动作条件视频预测。
NISE targets the sharp failure of 3D policies under realistic depth noise, carrying position uncertainty into attention; tested on transparent and reflective objects with commodity RGB-D.NISE 针对 3D 策略在真实深度噪声下的急剧失效，把位置不确定性带入注意力计算；在透明和反光物体上用普通 RGB-D 相机测试。
CoRL 2026 Spotlight. A morphology-conditioned implicit neural model of pose reachability, differentiable, self-collision aware and able to generalize to unseen morphologies.CoRL 2026 Spotlight。以形态为条件的隐式神经位姿可达性模型，可微、考虑自碰撞，并能泛化到未见形态。
CORD-SLS pairs a GPU-parallel differentiable simulator with robust MPC and conformal bounds for rope and cloth, reporting millisecond planning with safety margins on hardware.CORD-SLS 将 GPU 并行可微仿真与鲁棒 MPC、共形预测误差界结合，用于绳索和布料操作，报告在实机上以毫秒级规划并保持安全裕度。
CoRL 2026 Spotlight. Conditions a mobile manipulation policy on a spatiotemporal map of neural points that represents environment and robot body in one latent space, updated online.CoRL 2026 Spotlight。用时空神经点地图作为移动操作策略的条件：环境与机器人本体处于同一隐空间，并在线更新。
Reuses a frozen multi-robot value function as a differentiable surrogate for robot design, avoiding a new co-design RL loop per robot; handles over 1100 design parameters.把冻结的多机器人价值函数作为可微代理模型来优化机器人设计，无需为每台机器人重跑协同设计 RL；可处理 1100 多个设计参数。
CoRL 2026 Spotlight. Simulation benchmark for wires and cables in industrial assembly, aimed at the gap left by benchmarks tied to specific hardware or lacking modular, customizable task assets.CoRL 2026 Spotlight。面向工业装配中线缆等可变形线状物体（DLO）的仿真基准，针对现有基准绑定特定硬件、缺少模块化可定制任务资产等问题。
Tue 11/10, 5:00–6:30 PM · JW Marriott · 231 papers · Lone Star (3rd floor): posters 1–116 · Griffin (2nd floor): posters 117–232
World action model without video generation: predicts future features in a vision foundation model's latent space as policy subgoals. Reports 98.6% on LIBERO and up to 24× lower latency than pixel WAMs.不生成视频的世界动作模型：在视觉基础模型的特征空间中训练潜动作模型，预测未来观测特征作为“潜在视觉子目标”来条件化动作生成。报告 LIBERO 98.6%、RoboTwin 91.22%，每个动作块推理 187 ms，比像素级 WAM 延迟低最多 24 倍。
A router picks a VLM planner per subtask from scene context (CoT depth, model size, memory). On a Franka DROID setup it matches a stronger model's success at up to 65% lower average latency.路由器根据场景上下文为每个子任务选择 VLM 规划器（思维链深度、模型规模、记忆长度）。在 Franka DROID 平台上以最多降低 65% 的平均延迟达到更强模型的成功率。
Learns, with RL, a meta-policy that picks per state how many denoising steps (sequential) and candidate samples (parallel) a frozen diffusion/flow policy gets, trading success against compute.用强化学习训练一个元策略，按状态决定冻结的扩散/流策略在每步分配多少去噪步数（串行）和多少候选采样（并行），在成功率与计算量之间权衡。
Distills intermediate denoising features of a video diffusion model into an action model during training and drops the video model at test time. Reports 95.9% on LIBERO and about 21x faster inference than a WAM baseline.训练时把视频扩散模型的中间去噪表示蒸馏进动作模型，测试时完全丢弃视频模型。在 LIBERO 上报告 95.9% 平均成功率，推理比代表性世界动作模型快约 21 倍。
Keeps the last denoising step of a diffusion VLA as a fast feedback layer, so each action is corrected with the newest image. Reports dynamic sim tasks 27.5% to 85.0% and real tasks 51% to 73%.把扩散 VLA 的最后一步去噪保留为快速反馈层，用最新图像修正每个动作。报告动态仿真任务从 27.5% 升到 85.0%，真机任务从 51% 升到 73%。
FreeBar: training-free plugin that keeps consecutive action chunks of flow-based VLAs consistent with a CBF-style constraint solved in closed form, with little extra inference cost.FreeBar：无需训练的推理插件，用闭式求解的类 CBF 一致性约束让流式 VLA 的相邻动作块与已执行动作衔接，额外推理开销很小。
Shows discrete diffusion policies do real-time chunking natively, since inpainting is just unmasking: no fine-tuning, ~0.7x compute. Reports 65% higher success than flow-matching RTC on a real hockey task.指出离散扩散策略天然支持实时分块异步执行（补全即解掩码），无需微调、计算量约为 0.7 倍。报告在真机冰球防守任务上成功率比流匹配 RTC 高 65%。
Keeps a streaming buffer of action chunks at different noise levels with chunk-wise causal attention, so each inference step emits one executable chunk and asynchronous execution stays smooth.维护一个包含多个不同噪声水平动作块的流式缓冲区，并使用块级因果注意力，使每次推理产出一个可执行动作块，异步执行时动作保持连贯。
PAINT keeps action chunks consistent under async execution by picking the initial noise (backward Euler inversion), not by guiding the flow. No training; reports RTC-level task success with better prefix consistency.PAINT 通过选择初始噪声（后向欧拉反演）而非引导流场，让异步执行时的动作块首尾一致；无需训练，报告任务成功率与 RTC 相当、前缀一致性更好。
Predicts object flow to synthesize future observations, aligned to real future frames by a contrastive loss, so an asynchronous policy can act ahead of its own latency on moving objects.利用预测的物体流合成未来观测，并以对比学习目标使其特征与真实未来帧对齐，让异步策略能提前补偿自身延迟，处理运动中的物体。
Shows autoregressive policies can run in real time by tuning the tokenization horizon and constrained decoding under strict latency bounds; reports they beat an equivalent flow-matching policy.说明通过调整分词时域并在严格延迟上限下做约束解码，自回归策略也能实时执行；报告其表现优于同级别的流匹配策略。
Lets each modality update at its own sensor rate through latent buffers read by the action head, adding fast signals via gated cross-attention. Reports 95.2% vs 40.95% on seven contact-rich real tasks at 100 Hz.各模态按各自传感器频率刷新潜在缓存，供动作头持续读取，高频信号通过门控交叉注意力接入。报告在七个富接触真机任务上成功率 95.2%，对比基线 40.95%，控制频率 100 Hz。
Steers a flow-matching VLA in real time from a keyboard by writing the user input into the action expert's initial noise; no retraining. Fine-tuning on steered runs also improves the autonomous policy.把键盘等简单用户输入写入流匹配动作专家的初始噪声，即可实时引导 VLA，无需重训；用引导后的轨迹微调还能提升自主策略。
Serving one robot policy to a fleet from a remote GPU, posed as scheduling. Naive batching fails when robots consume action chunks at different rates; their scheduler reports up to 18% higher real-world throughput.把从远程 GPU 为多台机器人提供策略推理建模为调度问题。当各机器人消耗动作块的速度不同时，简单批处理失效；所提调度算法在真机上报告吞吐量最高提升 18%。
Finds that shift equivariance, local receptive fields and inference choices let diffusion planners stitch sub-trajectories without TD learning; locality matters most. Proposes the Eq-Net architecture.发现平移等变性、局部感受野和推理设置使扩散规划器无需 TD 学习就能拼接子轨迹，其中局部性最关键；据此提出 Eq-Net 架构。
Pretrain the action expert first as a language-free vision-action prior, then inject language through gated fusion. Targets VLAs that ignore unseen instructions; works with π-style and GR00T-style designs.先把动作专家预训练为不含语言的视觉-动作先验，再通过门控融合注入语言。针对 VLA 对未见指令泛化差的问题，适用于 π 系和 GR00T 系架构。
Feeds a grounded 3D target point through a two-layer MLP straight into a VLA's action head. On LIBERO-PRO it lifts GR00T-N1.6 from 31.2 to 77.5 under task perturbation; π0.5 gains similarly.把一个已定位的 3D 目标点经两层 MLP 直接注入 VLA 的动作头。在 LIBERO-PRO 任务扰动下，GR00T-N1.6 成功率从 31.2 提升到 77.5，π0.5 也有类似提升。
Turns a VLM into a VLA without architecture changes by writing a short language action description before the numeric action tokens. 2B model reports 90.8% on LIBERO after 4–5 hours of training; 0.8B–4B weights.不改结构直接把 VLM 变成 VLA：在数值动作 token 之前先生成一句自然语言动作描述。2B 模型训练 4–5 小时，报告 LIBERO 90.8%；发布 0.8B–4B 开放权重。
Adapts a frozen diffusion or flow policy at test time: a VLM writes differentiable rewards that steer denoising when obstacles or clutter shift the scene. Reports +31% on CALVIN and +13% on LIBERO-PRO; no retraining.在测试时适配冻结的扩散或流匹配策略：当障碍或杂物改变场景时，由 VLM 生成可微奖励来引导去噪。报告 CALVIN 提升 31%、LIBERO-PRO 提升 13%，无需重新训练。
Steers autoregressive VLAs at inference by injecting low-dimensional user input into the action-token space. Reports success from 10.0% to 72.5% and 16.7% to 93.8% on two household tasks.推理时把低维用户输入注入自回归 VLA 的动作 token 空间来引导轨迹，无需训练。报告两个家务任务成功率分别从 10.0% 升到 72.5%、从 16.7% 升到 93.8%。
ReStruct splits a policy into a state-machine skeleton and residual controller; a VLM turns a language preference into a temporal-logic automaton that reshapes the skeleton. 92.5% preference success on a real robot.ReStruct 把视觉运动策略分解为状态机骨架和残差控制器；VLM 把语言偏好编译成时序逻辑自动机来重构骨架。真机上偏好遵循成功率 92.5%。
Adapts a frozen generalist policy at inference: two small proxies (one copies base behaviour, one is fine-tuned on the task) and their velocity difference steers each denoising step. Base weights are not needed.推理时适配冻结的通用策略：训练两个小代理策略（一个模仿基座行为，一个在目标任务上微调），以二者的速度场差值引导每一步去噪。无需访问基座参数。
Improves a frozen one-step generative policy offline by learning a residual on its input noise, trained with advantage-weighted reconstruction instead of critic gradients, so the base behavior does not drift.离线改进冻结的一步生成策略：学习对输入噪声的残差修正，用优势加权的动作重建来训练而不是用评论家梯度，从而避免偏离原有行为。
CoRL 2026 Spotlight. Adapts frozen generative policies from human interventions in latent space, as a cheaper alternative to mass data collection or online RL on hardware.CoRL 2026 Spotlight。在隐空间中根据人类干预适配冻结的生成式策略，替代大规模采集或实机在线强化学习。
Fine-tunes diffusion and flow policies with automatic preference labels: rollouts are ranked by progress measured against a few expert demos, no human labels or reward design. Reports beating SFT on long-horizon tasks.用自动生成的偏好标签微调扩散与流匹配策略：以少量专家示教为参照衡量自主回放的顺序进度并据此排序，无需人工标注或手工奖励；报告在长时程操作任务上优于监督微调基线。
Turns expert demos into negative supervision by pairing actions with mismatched instructions, training an instruction-grounded reward for offline RL of VLAs. More robust on LIBERO-PRO and a real AgiBot-G2-type robot.把专家动作与不匹配的指令配对生成反事实样本，训练与指令对齐的奖励模型，用于 VLA 离线强化学习，无需额外采集。在 LIBERO-PRO 和真机上提升鲁棒性。
Adds an auxiliary branch that predicts future 3D tracks of robot keypoints, computed for free by forward kinematics, with gradient isolation from the VLM. The branch is dropped at inference.增加一个辅助分支，预测机器人关键点未来的 3D 轨迹（由正运动学免费得到），并与视觉语言主干做梯度隔离。推理时丢弃该分支，策略接口不变。
Masks the end-effector in the wrist camera (rendered from known geometry, plus mask augmentation). A VLA trained on one parallel gripper then transfers zero-shot to another gripper, another arm and a five-finger hand.在腕部相机画面中遮住末端执行器（依据已知几何实时渲染遮罩，并做遮罩增强）。只用单一平行夹爪数据训练的 VLA 可零样本迁移到其他夹爪、其他机械臂和五指手。
A VLA conditioned on one demo video of an unseen task predicts visibility-aware future end-effector traces for precise grounding. Releases RoboCasa-DC; reports +9.5 pts real-world success from human demos.以未见任务的一段示教视频为条件的 VLA，通过预测考虑可见性的未来末端轨迹实现精确定位。发布 RoboCasa-DC 基准；在以人类示教为条件的真机实验中成功率提升 9.5 个百分点。
Fluent expert demos rush through alignment and recovery. Slowing down there and resampling those segments helps; STAIR adds a compact short-horizon motion feature between the VLM and the action expert.熟练的示教会把对准与纠错一带而过，导致策略在关键处缺少监督。在这些片段放慢、重采样有帮助；STAIR 进一步在 VLM 与动作专家之间加入紧凑的短时运动特征。
FLEX lets language corrections edit a robot's task features via add, generalize and mask (not add-only). Reports higher normalized reward than other editing methods in 2D nav, Franka tabletop and a real user study.FLEX 用自然语言纠正来迭代修改机器人的任务特征集合，支持添加、泛化、屏蔽三种操作（而非只能添加）；在二维导航、Franka 桌面操作和真实用户研究中报告比其他特征编辑方法获得更高的归一化奖励。
Trains sparse autoencoders on VLA activations and finds features for motion primitives and concepts that transfer across episodes and can be amplified to steer behaviour. LIBERO + real DROID tests.在 VLA 隐层激活上训练稀疏自编码器，发现对应运动基元和语义概念、可跨回合迁移的特征，放大这些特征即可引导行为。在 LIBERO 和真实 DROID 平台验证。
Runs rough actions (from a human or a VLM) backwards through a flow policy to their noise, then forward again to land on a nearby good action mode. Distilling this reports up to 95-point success gains in under a minute.把粗糙动作（来自人或 VLM 的指引）反向送入流策略得到其噪声，再正向生成，落到附近的优质动作模态。将其蒸馏后，不到一分钟训练即可报告最高 95 个百分点的成功率提升。
Adds skill-conditioned LoRA mixture-of-experts adapters to the LLM backbone and turns a predicted affordance heatmap into a spatial prior for the action head. Reports Avg. Len 4.65 on CALVIN ABC→D.在大模型主干中加入以技能为条件的 LoRA 混合专家适配器，并把预测的可供性热图转为动作头的空间先验。在 CALVIN ABC→D 上报告平均完成长度 4.65。
Aligning VLA features point-wise with 3D foundation models causes rank collapse; aligning pairwise relations instead adds 3D awareness without losing expressivity. Adds a new spatial reasoning benchmark.将 VLA 特征逐点对齐到 3D 基础模型会导致秩坍塌；改为对齐成对关系，可在不损失表达力的前提下注入 3D 空间感知。并提出新的空间推理基准。
Post-trains an RGB geometry branch with RGB-D supervision, then lets state-generated queries pull eight geometry tokens for action prediction. Reports 99.0% on LIBERO and 78.8% on eight real ALOHA tasks.先用 RGB-D 监督后训练 RGB 几何分支，再由机器人状态生成查询，提取 8 个几何 token 用于动作预测。报告 LIBERO 99.0%，八个真实 ALOHA 任务 78.8%。
Info-VLA fights forgetting in continual VLA learning by keeping cross-modal structure: replay anchors from a frozen teacher for contrastive alignment, plus a vision-language mutual-information term.Info-VLA 通过保持跨模态信息结构来缓解 VLA 持续学习中的遗忘：以冻结教师模型构造回放锚点做对比对齐，并加入视觉-语言互信息约束。
PGRD simulates deformables with an optimizable spring-mass backbone plus learned residuals, reporting better accuracy than either alone; used for MPC and action-conditioned video prediction.PGRD 以可优化的弹簧质点模型为骨架，加上学习得到的残差来仿真可变形物体，报告精度优于单用任一方法；可用于 MPC 和动作条件视频预测。
World model for deformable objects that couples a differentiable MPM simulator with networks that infer per-particle elasticity from video and correct the simulator's residual errors. Real manipulation sequences.可变形物体世界模型：可微 MPM 仿真器配合两个网络，一个从视觉推断逐粒子弹性，一个学习修正仿真器的残差误差。在真实操作序列上验证。
CORD-SLS pairs a GPU-parallel differentiable simulator with robust MPC and conformal bounds for rope and cloth, reporting millisecond planning with safety margins on hardware.CORD-SLS 将 GPU 并行可微仿真与鲁棒 MPC、共形预测误差界结合，用于绳索和布料操作，报告在实机上以毫秒级规划并保持安全裕度。
Refines cloth actions online from one RGB image: a real-to-sim module estimates the cloth state, then simulator rollouts with prior-guided MPPI correct the policy's trajectory.仅凭单张 RGB 在线修正布料动作：先由 real-to-sim 模块估计布料状态，再用仿真器并行推演和先验引导的 MPPI 修正策略轨迹。
Sim-to-real RL for multi-stage cable routing: GPU-parallel cable simulation, localized policies and a simulation-in-the-loop executor. Reports higher success and 2x shorter cycle times than prior learning methods.多阶段线缆布线的仿真到真机强化学习：GPU 并行线缆仿真、局部化策略，加上执行时“仿真在环”的部署框架。报告成功率更高、节拍时间缩短一半。
CoRL 2026 Spotlight. Simulation benchmark for wires and cables in industrial assembly, aimed at the gap left by benchmarks tied to specific hardware or lacking modular, customizable task assets.CoRL 2026 Spotlight。面向工业装配中线缆等可变形线状物体（DLO）的仿真基准，针对现有基准绑定特定硬件、缺少模块化可定制任务资产等问题。
In-context imitation for folding: from a single human demo, a flow-matching transformer infers and executes the shown folding mode without gradient updates. Trained only in sim, transfers zero-shot to real.面向折叠的上下文模仿学习：只看一条人类演示，流匹配 Transformer 即可推断并执行其中的折叠方式，无需梯度更新。完全在仿真训练，零样本迁移到真机。
Synthetic T-shirt folding and unfolding data: 6 robots, 1K shirts, 120K episodes. Policies trained only on it are reported above 90% success on unseen real shirts.合成 T 恤折叠与展开数据：6 种机器人、1000 件 T 恤、12 万条轨迹。报告仅用其训练的策略在未见过的真实 T 恤上成功率超过 90%。
A short safe wiggle reveals rope parameters that condition a trajectory optimiser for zero-shot dynamic throws. Reports 3.55 cm average error on real 3D target striking vs 15.29 cm for uninformed baselines.先做一次短暂安全的“抖绳”动作来辨识绳子参数，再以此驱动轨迹优化，零样本完成动态甩绳。报告真机 3D 目标击打平均误差 3.55 cm，无辨识基线为 15.29 cm。
CoRL 2026 Oral. Trains a VLA on over 20,854 hours of action-labeled egocentric human video to test whether large-scale human data can support fine, high-DoF dexterous manipulation.CoRL 2026 Oral。用超过 20,854 小时带动作标注的第一视角人类视频训练 VLA，检验大规模人类数据能否支撑高自由度精细灵巧操作。
CoRL 2026 Spotlight. A pipeline that turns egocentric human manipulation videos into robot training data, testing whether this helps VLA pretraining at scale.CoRL 2026 Spotlight。把第一视角人类操作视频转换为机器人训练数据的流程，并检验其能否大规模助力 VLA 预训练。
CoRL 2026 Spotlight. Augments contact-rich teleoperation with corrective behaviors learned in simulation, using a simple kNN surrogate of the human operator.CoRL 2026 Spotlight。为富接触遥操作叠加在仿真中学到的纠正行为，并用简单的 kNN 模型替代人类操作员进行训练。
Converts egocentric human videos into robot data: a video with the human replaced by a robot plus a feasible robot action trajectory. Claims the first zero-shot dexterous visuomotor policy with no real-robot demos.把第一人称人类视频转为机器人数据：生成把人替换成机器人的视频，以及满足可行性约束的机器人动作轨迹。声称首次在无真机演示下零样本训练出灵巧视觉运动策略。
Converts human videos from five sources into robot-aligned observations and actions (about 630K episodes). Pretraining a VLA on more of it raises real-task completion from 51.5% to 80.3% and OOD from 34.9% to 72.2%.把五个来源的人类视频转换成与机器人对齐的观测和动作（约 63 万条轨迹）。报告用更多此类数据预训练 VLA，真机任务完成率从 51.5% 升到 80.3%，分布外从 34.9% 升到 72.2%。
Continuous (not vector-quantized) latent actions learned from 23.3K hours of egocentric human video serve as pre-training labels for a VLA; reports consistent scaling gains.从 2.33 万小时第一人称人类视频中学习连续（非向量量化）的潜动作，结合感知、语义与光流对齐信号，作为 VLA 预训练的动作监督；报告随规模持续提升。
Joint VLA pretraining on 4.53K h of robot/sim data plus 1.48K h of egocentric human video with pseudo-actions, camera-space actions and a reliability-weighted loss. Reports SOTA on RoboCasa GR1 TableTop and RoboTwin 2.0.用 4530 小时机器人/仿真数据加 1480 小时带伪动作的第一视角人类视频联合预训练 VLA，采用相机空间动作与按可靠性加权的损失。报告在 RoboCasa GR1 TableTop 和 RoboTwin 2.0 上达到最优。
Controlled study of which world-prediction target best transfers egocentric human data to robots: pixels transfer weakly, DINO features give up to 4x OOD generalization, 3D flow gives +20–30% in domain.对照研究哪种世界预测目标最能把第一人称人类数据迁移到机器人：像素迁移较弱，DINO 特征带来最高 4 倍的分布外泛化，3D 运动流在域内提升 20–30%。
CoRL 2026 Oral. Makes flow matching policy gradients work for robot control, bypassing action likelihoods, so policies need not be Gaussian. Applied to legged locomotion, humanoid motion tracking and manipulation.CoRL 2026 Oral。让流匹配策略梯度可用于机器人控制：绕过动作似然计算，策略不再限于高斯分布。应用于足式运动、人形动作跟踪与操作。
CoRL 2026 Oral. Adapts a VLA to expose a compact RL token, an interface for sample-efficient online RL, so a pretrained policy gains precision and speed from a few hours of real-world practice.CoRL 2026 Oral。让 VLA 输出紧凑的 RL token 作为在线强化学习接口，预训练策略只需几小时真实练习即可提升精度与速度。
Weights each action chunk by how much it shortens the path to success on a reachability graph built over demos, rollouts and interventions. Reports 91% on precise tasks after a few offline rounds; no reward model.在示教、rollout 与人工干预构成的可达图上，按每个动作块缩短到成功的步数来加权行为克隆。报告经几轮离线迭代在精细任务上达到 91%，无需奖励模型。
Advantage-conditioned BC for VLAs: predicts pairwise conditions directly and injects them into the DiT action expert via AdaRMS. On real retry-heavy data: 90.0% vs 56.7% (bowl stacking), 56.7% vs 33.3% (key hanging).面向 VLA 的优势条件行为克隆：直接预测成对条件，并通过 AdaRMS 注入 DiT 动作专家。在含大量重试的真机数据上，叠碗成功率 90.0%（基线 56.7%），挂钥匙 56.7%（基线 33.3%）。
One-step flow policy that puts observation, timestep and action tokens in one sequence for a single transformer. Reports 86.4% average success (+7.5 points) at 3.0 ms latency, 43.9x faster than DP3.一步生成的流策略：把观测、时间步与动作 token 拼成一个序列，由同一 Transformer 联合建模。报告平均成功率 86.4%（提升 7.5 个百分点），推理延迟 3.0 ms，比 DP3 快 43.9 倍。
Models robot flow as dense velocity fields in a flow-matching formulation rather than sparse keypoint displacements, about 24x faster than standard flow matching; tested in 13 real tasks with 260 trials per method.在流匹配框架中把机器人流建模为稠密速度场，而非稀疏关键点位移，生成速度比标准流匹配快约 24 倍；在 13 个真实任务上每种方法测试 260 次。
Single-camera diffusion policy: points along intermediate end-effector trajectories in the denoising loop sample multi-scale image features, acting as attention anchors. Comparable to multi-camera baselines in sim.单相机扩散策略：在去噪循环中沿中间末端轨迹的点采样多尺度图像特征，作为视觉注意力锚点。仿真中表现与多相机基线相当。
Feynman-Kac correctors with a new guiding potential push diffusion-policy samples toward rare but promising modes; trajectory optimization repairs them and they are added back to training.用 Feynman-Kac 校正器配合新的引导势能，把扩散策略的采样推向少见但有前景的模态；再以基于采样的轨迹优化修复这些轨迹，并加回训练集重训策略。
LeaP replaces the standard Gaussian source of a flow/diffusion-bridge policy with a proprioception-conditioned Gaussian over action chunks from a small MLP. 81.6% average on 15 RoboTwin tasks.LeaP 用一个小型 MLP 预测、以本体感觉为条件的动作块高斯分布，替换流匹配/扩散桥策略的标准高斯起点。在 15 个 RoboTwin 任务上平均成功率 81.6%。
Replaces a policy MLP's first layer with learned multiplicative (bilinear) features, plug-and-play for RL and imitation. Reports humanoid survival up from 51.0% to 92.5%, holding under friction and payload shifts.把策略 MLP 的第一层替换为可学习的乘性（双线性）特征，即插即用于强化学习与模仿学习。报告人形机器人存活率从 51.0% 提升到 92.5%，在摩擦、负载变化下依然有效。
Behavioral INR represents a policy as a state-to-action function modulated by an episode latent (FiLM), so policy identity in unlabeled multi-policy data can be inferred without labels.Behavioral INR 把策略表示为由回合级潜变量（FiLM）调制的“状态→动作”函数，从而在无标签的多策略行为数据中无监督地推断策略身份。
Tests why action chunking helps BC: not temporal consistency or horizon reduction, but non-Markovian expressivity and less compounding error (also matched by delayed policies), plus an 'implicit ensembling' effect.检验动作分块为何有助于行为克隆：并非时间一致性或缩短时域，而是非马尔可夫表达力与更少的误差累积（延迟策略也能做到），另外还有一种“隐式集成”效应。
Composed diffusion policies often learn identical components. A drop-in score-decorrelation loss makes each factor specialize, with no loss in multitask performance.组合式扩散策略的各个分量常常学成相同的副本。即插即用的得分去相关损失让各分量各自专门化，且不损失多任务性能。
MoRE 'unclones' unwanted behavior modes (e.g. handing a knife blade-first) by distilling a temporary mode classifier into the policy weights, with a retain loss. +44 pts success across eight tasks, no inference cost.MoRE 通过把临时模式分类器的重定向信号蒸馏进策略权重，并配合保持损失，“去克隆”不想要的行为模式（如刀刃朝前递刀）。八个任务平均成功率提升 44 个百分点，推理零开销。
Sleeping Robots: learn a new skill while 'awake', then consolidate one shared policy offline during 'sleep' from frozen critics or actor snapshots, combining gradients by Nash bargaining. +64% success on Meta-World MT5.Sleeping Robots：“清醒”时学习新技能，“睡眠”时利用冻结的评估器或策略快照离线巩固单一共享策略，并以纳什议价合并梯度。在 Meta-World MT5 上平均成功率提升 64%。
ReTVL learns a mistake-sensitive value function from imperfect demos, using sparse retry keypoints as pairwise preferences on top of global progress, then reweights demo chunks for behavior cloning.ReTVL 从不完美示教中学习对失误敏感的价值函数：在全局进度校准之上，以稀疏标注的“重试”关键点构造局部成对偏好，再用该价值为行为克隆的示教片段加权。
New manipulation tasks where the target must be found by sound, plus an imitation framework that adds acoustic spatial and spectral features to ACT, Diffusion Policy, VQ-BeT and π0. Sim + real.提出需要靠声音确定操作目标的新任务，以及把声源空间信息与频谱特征接入 ACT、Diffusion Policy、VQ-BeT 和 π0 的多模态模仿学习框架。仿真与真机验证。
Diffusion policy with an obstacle-aware encoder that splits the scene into target, obstacle and background. 61 real greenhouse trials per method: 75.41% success and 8.20% collision rate.带障碍感知编码器的扩散策略，把场景分解为目标、障碍和背景。每种方法在温室真机测试 61 次：成功率 75.41%，碰撞率 8.20%。
CoRL 2026 Spotlight. A morphology-conditioned implicit neural model of pose reachability, differentiable, self-collision aware and able to generalize to unseen morphologies.CoRL 2026 Spotlight。以形态为条件的隐式神经位姿可达性模型，可微、考虑自碰撞，并能泛化到未见形态。
Benchmark showing VLA success often hides weak instruction following: scenes allow several valid tasks so language matters. Nine VLA and WAM policies drop sharply beyond the base level.揭示 VLA 的高成功率常掩盖薄弱的指令遵循：场景支持多种合法任务，必须依赖语言。九个 VLA 与 WAM 策略在基础级以外的测试中明显下降。
Four-level benchmark that widens the gap between a human demo and the robot's scene, plus IG-10K (20,000+ paired human-robot episodes). Nine models collapse at the hardest level; all stay below 13% zero-shot.四级基准逐步拉大人类演示与机器人场景的差异，并发布 IG-10K（2 万余条人-机配对片段）。九个模型在最高级别全面失效，未见任务零样本成功率均低于 13%。
HelpMeOut: 20,000 paired household activity episodes with nine neuropsychology-based error types; agents must spot when a person's task execution goes wrong. Learned, symbolic and LLM monitors all do poorly.HelpMeOut：2 万对家务活动片段，涵盖九类基于神经心理学的执行错误，要求智能体尽早发现人做事出错。学习型、符号型与大模型监控方法表现都较差。
Two cheap failure detectors for fine-tuned VLAs that need no failure data: last-layer Mahalanobis distance and consistency between overlapping action chunks, plus a metric that also scores detection time.为微调后的 VLA 提出两种无需失败数据的轻量失败检测器：末层特征马氏距离与相邻动作块一致性，并提出同时衡量检测时间的指标。
Self-supervised failure detection from nominal demos only, using object-centric spatio-temporal scene graphs; both of its scores beat image-based detectors, especially on semantic and interaction failures.仅用正常演示做自监督失败检测：把操作场景建模为以物体为中心的时空图，所提两种失败评分都优于基于图像的检测器，尤其擅长语义和交互层面的失败。
Learns recovery skills with RL and, from them, discovers new relational predicates for an abstract planner, turning local recoveries into failure avoidance. Reports >50% gains over baselines in four sim domains.用强化学习学习失败恢复技能，并据此发现新的关系谓词、完善抽象规划模型，把局部恢复转化为全局避免失败。报告在四个仿真领域比基线高 50% 以上，并迁移到真机。
Retries that learn from failure: preference pairs from failed attempts steer the policy away from them, plus small action perturbations, and feeds recoveries back into training. Reports +17.6% sim, +11.7% real.会从失败中学习的重试机制：用失败尝试构造偏好数据，引导策略避开失败行为，并加入小幅动作扰动做局部探索，成功恢复的轨迹再回流训练。报告仿真提升 17.6%，真机提升 11.7%。
B2FF: before execution the VLA imagines a bank of future milestone images; after a deviation a selector picks a reachable one as a fixed visual goal. Failure-injected LIBERO: 56.3% → 74.0%, no fine-tuning.B2FF：执行前让 VLA 想象一组未来里程碑图像；偏离轨迹后，由选择器挑出可恢复的那个作为固定视觉目标。在注入失败的 LIBERO 上成功率从 56.3% 升至 74.0%，无需微调。
Turns ambiguous failed rollouts into recovery training data: a privileged coach proposes a corrected high-level decision, kept only if re-executing from the same state improves progress.把含义不明的失败推演转化为恢复训练数据：由特权教练推测早期的高层决策错误并提出修正，只有从同一状态重新执行确实提升进度时才保留。
Shows a prompt that still reads like the intended task can redirect a VLA's rollout to an attacker's target; an on-policy prompt search finds such near-benign prompts. Sim + hardware.揭示 VLA 的轨迹级漏洞：看起来仍在描述原任务的指令，可以把机器人执行结果引向攻击者指定的目标；作者用在线提示搜索找到这类近乎无害的提示。仿真与真机验证。
FLARE, a physical spotlight attack, drops VLA success to zero. Naive augmentation defenses make models ignore color (≤47.5% on color tasks); ChromaGuard keeps color and reaches 97.5% and 92.5% on a real arm.FLARE 物理聚光灯攻击可让 VLA 成功率降为零。朴素的增强防御会让模型忽略颜色（颜色相关任务至多 47.5%）；ChromaGuard 保留色彩信息，在真机上达到 97.5% 与 92.5%。
Measures which instruction factors policies over-rely on: across six foundation policies, color ≥ object ≥ spatial ≥ verb ≥ size. Collecting more data for under-grounded factors wins with half the demos.量化策略过度依赖哪些指令要素：六个基础策略普遍呈现颜色 ≥ 物体 ≥ 空间 ≥ 动词 ≥ 尺寸的顺序。把采集预算倾斜给欠落地的要素，只用一半示教即可胜出。
Suites isolate generalisation across spatial layout, object instance, object category and subtask composition, plus bimanual BiMESA. The best of seven baselines reaches only 52% and 49%.五组评测分别隔离空间布局、物体实例、物体类别和子任务组合上的泛化能力，并提供双臂版 BiMESA。七种基线中最好的也只有 52% 和 49%，仍有很大提升空间。
RoboFQE scores visual manipulation policies from offline trajectories, no rollouts: fitted Q-evaluation on a pretrained vision encoder plus learned success labels. Keeps rankings where video world models drift.RoboFQE 无需推演、直接用离线轨迹评估视觉操作策略：在预训练视觉编码器上做拟合 Q 评估并学习终止成功标注。在视频世界模型会累积误差的长时任务上仍能保持策略排序。
Mobile-manipulation benchmark with 70 task scenarios combining navigation and manipulation in continuous spaces. Finds leading VLMs still struggle with spatial localization.移动操作评测基准：70 个在连续空间中结合导航与操作的任务场景。发现主流 VLM 在空间定位上仍有明显困难。
RoboReel: a unified benchmark for policies learned from human videos, bundling real human demos, simulated robot trajectories and ten tasks with four test suites; sim trends hold on real-robot recreations.RoboReel：从人类视频学习策略的统一基准，打包真实人类演示视频、仿真机器人轨迹与十个任务及四套测试；在真机复现中仿真结论保持一致。
IMBench: 35 tasks and 14K trajectories testing reasoning and execution together under explicit constraints. VLMs reason partly but fail to plan executable steps; VLAs struggle with the constraints.IMBench：35 个任务、1.4 万条轨迹，在显式约束下同时考察物理推理与动作执行。VLM 有部分推理能力但给不出可执行计划，VLA 难以满足任务约束。
600+ hours of whole-body human motion and object interaction from optical mocap with sub-millimetre marker accuracy, organised by FrameNet primitives, with a benchmark suite for humanoid learning.600 余小时的高精度全身人体运动与物体交互数据，由光学动捕采集（亚毫米级标记精度），按 FrameNet 动作基元组织，并配套面向人形机器人学习的评测套件。
Finds real-system failures on a small budget by correcting cheap proxy (e.g. simulator) failure signals with residual modeling and choosing diverse, well-supported scenarios. Up to 2x the failures of random sampling.在有限测试预算下寻找真实系统的失败：用残差建模校正廉价代理（如仿真器）的失败信号，并挑选多样且有数据支撑的场景。发现的失败数最多为随机采样的 2 倍。
SCAPE predicts per-scenario real-world policy performance from a few paired sim/real runs plus many sim runs, correcting sim bias and calibrating with conformal prediction. Tested on driving and quadrupeds.SCAPE 利用少量成对的仿真/真机数据和大量仿真推演，按场景预测策略在真实世界的表现：先校正仿真标签偏差，再用共形预测校准不确定性。在自动驾驶与四足速度跟踪上验证。
AnyRank ranks robot policies with sequential tests: trials go to hard-to-separate pairs, a (partial) ranking holds at any stopping time, and imperfect priors (sim, world model) save trials safely.AnyRank 用序贯检验为多个机器人策略排序：把试验分给难以区分的策略，任何时刻停止都能给出有误差保证的（部分）排序，可借助不完美的先验（仿真、世界模型）节省试验而不破坏保证。
DLS finds where VLA success breaks via on-policy digital-twin rollouts, locates the crossing with privileged progress signals, and shapes the flow toward success-producing denoising directions.DLS 通过数字孪生中的在策略推演大规模发现 VLA 成功与失败的边界，用特权进度信号定位越界位置，并据此强化导向成功的去噪方向、抑制导向失败的方向。
Reuses a frozen multi-robot value function as a differentiable surrogate for robot design, avoiding a new co-design RL loop per robot; handles over 1100 design parameters.把冻结的多机器人价值函数作为可微代理模型来优化机器人设计，无需为每台机器人重跑协同设计 RL；可处理 1100 多个设计参数。
Diffusion transformer over a shared tokenization of robot bodies, states and actions that generates robot designs to track target motions. A fabricated optimised ALOHA design cut tracking error by over 70%.基于统一的机器人本体/状态/动作 token 表示的扩散 Transformer，按目标末端轨迹生成机器人设计。实际制作的优化版 ALOHA 跟踪误差降低 70% 以上。
A simulation benchmark of about 0.5B grasp labels over eight dexterous hands and ~50K objects, plus a diffusion model for cross-embodiment grasps in clutter: 88% average success, 0.3 s inference.构建约 5 亿条抓取标注、覆盖 8 种灵巧手和约 5 万个物体的仿真基准，并提出跨本体的杂乱场景灵巧抓取扩散模型：平均成功率 88%，推理 0.3 秒。
Diffusion grasp generator conditioned on a graph of the hand's kinematics, so grasps transfer zero-shot to unseen hand designs; comes with a cross-morphology grasp benchmark.以手部运动学结构图为条件的扩散抓取生成模型，可零样本迁移到未见过的机器人手型；同时提出跨手型抓取评测基准。
Multi-GraspSet, a large multi-hand grasp dataset with automatic contact annotations, and an LLM that generates language-guided grasps for different robot hands in one model.发布带自动接触标注的大规模多手型抓取数据集 Multi-GraspSet，并用一个基于大模型的统一框架，根据语言指令为不同机器人手生成抓取。
CoRL 2026 Oral. Generates executable dexterous grasps from free-form language, inferring the function a task implies from multi-view images and geometry, supported by the OpenDexVerse data.CoRL 2026 Oral。根据自由语言生成可执行的灵巧抓取：结合多视角图像与几何推断任务所需的功能，由 OpenDexVerse 数据支撑。
CoRL 2026 Spotlight. A large-scale bimanual dexterous grasp dataset, built with a two-stage synthesis pipeline, together with a learning-based grasp generation framework.CoRL 2026 Spotlight。大规模双手灵巧抓取数据集（由两阶段合成流程生成），并配套基于学习的抓取生成框架。
Learns an object-conditioned human grasp prior from 1.8K human grasps to pick contact modes and wrist starts for force-closure optimisation; synthesises 3.2M grasps from pinches to bimanual, 2–30 cm objects.从 1800 次人类抓取中学习物体条件的人类先验，用来选择接触模式和手腕初值以指导力封闭优化；合成 320 万个抓取，覆盖两指捏取到双手抓取、2–30 cm 物体。
Pretrains a low-level dexterous controller on 355K trajectories built from grasp datasets, then fine-tunes for articulated tool use. Introduces the DexCraft benchmark; reports +33.3 points over DP3 on a real robot.用抓取数据集构造的 35.5 万条轨迹预训练灵巧手底层控制器，再微调到需要操作工具活动部件的任务；提出 DexCraft 基准。报告真机完整任务成功率比 DP3 高 33.3 个百分点。
Hierarchical real-world RL for functional dexterous tasks: a hand skill handles contact, an arm-hand policy coordinates and triggers it. Reports 98.3% average success on six tasks within one hour of online training.面向功能性灵巧操作的分层真机强化学习：手部技能负责精确接触，臂-手协调策略负责手臂运动与技能触发。报告在六个任务上一小时在线训练即达 98.3% 平均成功率。
CoRL 2026 Oral. Asks what dexterous play pretraining actually needs before precise assembly, a contact-rich, sparse-reward task that is hard for both imitation and direct RL.CoRL 2026 Oral。探究灵巧“玩耍”预训练究竟需要什么，才能胜任精密装配这类富接触、稀疏奖励、模仿与直接 RL 都难以解决的任务。
Grasps several objects at once with a dexterous hand by learning one-step residual corrections to a reference trajectory, with a strategy that adapts to how the objects are laid out.让灵巧手一次抓取多个物体：学习对参考轨迹的单步残差修正，并根据物体的空间布局调整抓取策略。
AutoDex runs real dexterous-grasp trials, labels outcomes and resets objects with little human help: 3,500+ trials on 100 objects (Allegro, Inspire); about 0.2 human-hours per 500 trajectories vs 9.3 by teleop.AutoDex 自动执行真机灵巧抓取、标注结果并复位物体，几乎无需人工：在 100 个物体上用 Allegro 与 Inspire 手采集 3500 多次试验；每 500 条轨迹约需 0.2 人时，遥操作需 9.3 人时。
Category-level in-hand manipulation of articulated objects: procedural objects and functional grasps, then two-stage training with articulation physics randomization, reward curriculum and latent distillation.类别级的铰接物体手内操作：程序化生成物体与功能性抓取，再用两阶段训练（铰接物理随机化、奖励课程、潜表示蒸馏）学习策略，可泛化到未见实例。
End-to-end in-hand reorientation from RGB and proprioception, with no object pose estimation: distilled from privileged-state teachers, and the target is given as an image. One policy transfers to five real objects.仅凭 RGB 与本体感觉、无需物体位姿估计的端到端手内重定向：由特权状态教师蒸馏而来，目标姿态直接用图像给出。单一策略直接迁移到五个真实物体。
Uses hand-object proximity from reconstructed point clouds as a hardware-agnostic contact signal, for VR teleop feedback and a policy with learned proximity dynamics. More robust under occlusion.以重建点云得到的手-物距离作为与硬件无关的接触信号，用于 VR 遥操作反馈，并训练带接近度动力学的策略；在遮挡下更稳健。
Puts miniature cameras in each fingertip of a multi-fingered hand and trains a whole-body diffusion policy on third-view plus fingertip views. Reports 56.6% relative gain over external and wrist cameras.在多指手的每个指尖装入微型摄像头，用第三视角加多指尖视角训练全身扩散策略。报告比外部+腕部相机方案平均成功率相对提升 56.6%。
Generates dexterous grasps as 3D hand keypoints with a plain Transformer flow model, with no contact losses or refinement; reports 76.3% success on Dexonomy and 0.032 s per grasp.以 3D 手部关键点表示灵巧抓取，用普通 Transformer 流模型生成，无需接触损失或后处理优化；报告 Dexonomy 成功率 76.3%，每次抓取 0.032 秒。
6-DoF grasping that keeps learning on the job through memory in an embedding space: outcomes rescore future grasps and user demos are recalled in new scenes. Reports over 90% on hard categories after 50 tries.通过嵌入空间中的记忆在部署中持续学习的 6-DoF 抓取：抓取结果更新后续评分，用户示范可在新场景中被召回。报告在困难类别上仅 50 次尝试后成功率超过 90%。
Dual-arm grasps for large objects from partial point clouds: a convolutional occupancy network reasons about local thickness and clearance, producing force-closure grasp pairs refined by sampling-based optimization.从局部点云为大型物体生成双臂抓取：卷积占据网络推断局部厚度与夹爪间隙等几何，生成满足力封闭的抓取对，再用基于采样的优化修正几何不完整带来的歧义。
Simulation platform for visuo-tactile data supporting three common sensors, plus an encoder trained on its synthetic data and an 8-task benchmark. Reports +17.1% on the benchmark and +25% in real tasks.视触觉仿真数据合成平台，支持三种常用传感器，并基于合成数据训练视触觉编码器、发布 8 个任务的基准。报告基准成功率提升 17.1%，真机提升 25%。
Learns a structured latent space of behaviour styles (e.g. how much force to apply), and lets language reason over it instead of generating actions directly, for personalized tactile policies like wiping or folding.学习按行为风格（如施力大小）组织的结构化潜空间，让语言在其中推理用户偏好而非直接生成动作，从而实现个性化的触觉策略，如擦拭、叠衣。
Encoders for resistive, magnetic and vision-based tactile sensors are trained on paired contacts into one latent space, so a policy trained with one sensor transfers zero-shot to another on four manipulation tasks.为电阻式、磁式和视觉触觉传感器分别训练编码器，利用成对接触数据映射到同一潜空间；在四个操作任务中，用一种传感器训练的策略可零样本迁移到另一种。
Shared tactile representation across sensors: sensor-specific encoders plus a common transformer, pretrained on a new 1.6M-frame paired dataset from four sensors; adapts to new tasks and unseen sensors.跨传感器的统一触觉表征：各传感器独立编码器加共享 Transformer 主干，在新的 160 万帧四传感器配对数据集上预训练，可迁移到新任务和未见传感器。
Tactile policy pretrained on ~3,000 hours from 26 sources and 21 sensors, mapping image, array and state signals into shared tokens. Reports +17.2% on seen setups and +31% on two unseen sensor setups.触觉基础策略：在 26 个数据源、21 种传感器约 3000 小时数据上预训练，把图像型、阵列型与状态型触觉信号映射到统一 token。报告已见配置提升 17.2%，两种未见传感器提升 31%。
Fuses an RGB camera, a GelSight Mini and a high-rate event-based tactile sensor for a flow-matching policy. Reports 80% average success on five contact-rich tasks vs 35.5% vision-only.融合 RGB 相机、GelSight Mini 与高频事件型触觉传感器的多分辨率表示，驱动流匹配策略。报告五个富接触任务平均成功率 80%，纯视觉仅 35.5%。
DreamTacVLA feeds high-res tactile images with wrist and third-person views, aligns them spatially, then fine-tunes with a tactile world model that predicts future touch signals.DreamTacVLA 把高分辨率触觉图像与腕部、第三人称视角一起输入并做空间对齐，再用预测未来触觉信号的触觉世界模型微调。
FAWAM uses 6-axis force/torque in perception, jointly predicts future actions and wrenches, and corrects actions online against the predicted wrench. Reports +36.25 pts over vision-only baselines on real tasks.FAWAM 在感知、预测与闭环执行三层使用六维力/力矩：联合预测未来动作与末端力旋量，并以预测的力轨迹为参考在线修正动作。真机任务上比纯视觉基线成功率高 36.25 个百分点。
FR-VLA uses learnable force query tokens to let vision-language context decide how much measured force should enter the action pathway, rather than feeding force in through a fixed path. Five real tasks, two robots.FR-VLA 通过可学习的力查询 token，由视觉语言上下文判断测得的力应在多大程度上进入动作生成通路，而不是走固定的注入路径。在两种机器人的五个接触密集任务上验证。
Gives a VLA memory through compact force-history tokens from a pretrained VAE, for tasks like pressing a button a set number of times. Reports over 80% success on three memory-dependent tasks.用预训练 VAE 把力信号历史压缩成记忆 token 输入 VLA，解决按指定次数按按钮、擦盘子等依赖记忆的接触任务。报告三个记忆型任务成功率超过 80%。
LIFT: grafts a reactive force-driven action expert onto a pretrained, vision-driven VLA, keeping its output unchanged at initialization, so post-training gains contact reactivity; CoRL 2026 Spotlight.LIFT：在以视觉为主的预训练 VLA 旁接入由力觉驱动的反应式动作专家，初始化时保持原策略输出不变，使后训练获得接触反应能力，应对遮挡、深度歧义等接触阶段难题；CoRL 2026 Spotlight。
Adds touch to a VLA via sim-real co-training and simulated RL in a real-aligned simulator, learning how contact should change actions near failure. Reports 72.5% vs 50.0% on four bimanual tasks.通过仿真-真机协同训练和在对齐真实的模拟器中做强化学习，为 VLA 加入触觉，学习接近失败时触觉应如何调整动作。报告四个双臂任务成功率 72.5%，基线 50.0%。
Real-world RL that adds a lightweight tactile residual on top of any pretrained visual policy, bootstrapped from the base policy's own rollouts. Reports 5–40% success raised to 85–100% within 40–80 minutes.真实世界强化学习：在任意预训练视觉策略之上学习轻量的触觉残差修正，先用基础策略自主执行的数据冷启动。报告四个任务成功率在 40–80 分钟内从 5–40% 提升到 85–100%。
Tactile-only blind grasping on a LEAP Hand: calibrate a digital twin to reproduce real touch signals, pretrain a layout-aware tactile encoder, and distill object-specific RL experts into a diffusion policy.在 LEAP 手上实现只凭触觉的盲抓：先构建能复现真实触觉信号的校准数字孪生，再自监督预训练感知传感器布局的触觉编码器，最后把各物体的强化学习专家蒸馏为扩散策略。
Learns tactile dexterous skills without tactile simulation: privileged sensors in the real world collect data to distil a tactile state estimator for a sim-trained policy. Reports 182% gain on in-hand rotation.无需触觉仿真即可学习触觉灵巧操作：在真实环境中借助特权传感器采集数据，蒸馏出基于触觉的状态估计器，用于仿真训练的策略。报告手内旋转任务提升 182%。
Uses motor current as a touch signal on low-cost dexterous hands: a model predicts a compliant PD target position whose error yields the right grasp force, with no tactile or force sensors.把电机电流当作触觉：在低成本灵巧手上预测一个“柔顺参考位置”，让 PD 控制器的位置误差产生合适的抓握力，无需触觉或力传感器。
CoRL 2026 Oral. NEXT estimates external joint torques without force sensors, training in 1 minute from 10 minutes of free-motion data, enabling force-feedback teleoperation and better policy learning on low-cost arms.CoRL 2026 Oral。NEXT 无需力传感器即可估计外部关节力矩，用 10 分钟自由运动数据训练 1 分钟，为低成本机械臂带来力反馈遥操作与更好的策略学习。
CoRL 2026 Spotlight. Dual-branch generative framework that extracts clean shading and marker images at once from one marker-based tactile frame, avoiding serial mask-based inpainting.CoRL 2026 Spotlight。双分支生成框架，可从单帧带标记点的触觉图像同时分离出干净的明暗图与标记图，避免串行掩码修复。
Dexterous visuotactile VLA on a mixture-of-transformers: contact-aware gating decides when touch matters; the model jointly imagines future visual and tactile latents. Reports 71% average success on contact-rich tasks.基于混合 Transformer 的灵巧视触觉 VLA：接触感知门控决定何时引入触觉，并联合想象未来的视觉与触觉潜变量。报告在富接触任务上平均成功率 71%。
Inference-time steering with vision and touch: visual sampling picks the behaviour, tactile-guided diffusion editing fixes local contact, scored by a visuo-tactile latent world model. Reports +51% over the base policy.视觉+触觉的推理时策略引导：视觉采样验证决定执行何种行为，触觉引导的扩散编辑修正局部接触，并由视触觉潜在世界模型打分。报告比基础策略提升 51%。
CoRL 2026 Spotlight. Fingertip sensing that combines binocular RGB cameras with a compliant contact surface, giving feedback from pre-contact approach through post-contact control.CoRL 2026 Spotlight。指尖传感将双目 RGB 相机与柔顺接触面结合，从接触前接近到接触后控制全程提供反馈。
CoRL 2026 Oral. Splits one tactile stream into three temporal roles, persistent states, abrupt contact events and future consequences, instead of a single fusion path in chunk-based insertion policies.CoRL 2026 Oral。不再用单一融合路径，而是把同一触觉流拆成三种时序角色：持续接触状态、突发接触事件与未来接触后果，用于分块式插接策略。
CoRL 2026 Spotlight. Brings contact into image-conditioned policies by rendering it, targeting the gap left by approaches that need tactile hardware, separate tactile encoders or non-image backbones.CoRL 2026 Spotlight。通过渲染接触信息把触觉带入基于图像的视觉运动策略，针对现有方法需要专用触觉硬件、独立触觉编码器或非图像主干，难以与预训练 2D 视觉表征策略兼容的问题。
CoRL 2026 Spotlight. Studies why VLA models struggle in contact-rich manipulation and separates two failure modes: precision failures, where policies struggle to produce small corrective actions, and force failures.CoRL 2026 Spotlight。分析 VLA 模型为何难以完成接触丰富的操作任务，并区分两类失败模式：难以输出接触时所需微小修正动作的精度失败，以及力失败。
Helps RL explore contact-rich tasks by resetting episodes with a contact-aware constrained state sampler, covering diverse contact modes; trains universal in-hand and whole-body policies.用感知接触的约束状态采样器重置回合，覆盖多样接触模式，帮助 RL 探索接触丰富任务；可训练通用的手内和全身操作策略。
CoRL 2026 Spotlight. Addresses impedance gain selection for articulated tasks such as doors, bolts and drawers, where neural policies may violate kinematic constraints.CoRL 2026 Spotlight。针对开门、拧螺栓、拉抽屉等铰接任务解决阻抗增益选择问题，这类任务中神经策略可能违反运动学约束。
CoRL 2026 Spotlight. A module for contact-rich motion transfer across sim-to-real gaps, payloads and robot instances, as an alternative to domain randomization or system ID.CoRL 2026 Spotlight。面向富接触动作迁移的模块，应对仿真到现实、负载与个体差异，用于替代域随机化或系统辨识。
Robots that never trained together share value functions, so a robot finishing a task leaves the workspace in a state the next robot prefers; a KNN uncertainty filter blocks overconfident values. Up to +48.5 pts.从未共同训练的机器人通过共享价值函数来协作，前一个机器人会把工作空间留在后一个机器人偏好的状态；KNN 不确定性过滤避免分布外的过度自信。成功率最高提升 48.5 个百分点。
CoRL 2026 Spotlight. Conditions a mobile manipulation policy on a spatiotemporal map of neural points that represents environment and robot body in one latent space, updated online.CoRL 2026 Spotlight。用时空神经点地图作为移动操作策略的条件：环境与机器人本体处于同一隐空间，并在线更新。
MoMa-SG builds 3D scene graphs that encode how objects articulate, from RGB-D sequences, and adds the Arti4D-Semantic dataset; demonstrated on a quadruped and a mobile manipulator.MoMa-SG 从 RGB-D 序列构建包含物体铰接方式的 3D 场景图，并发布 Arti4D-Semantic 数据集；在四足和移动操作机器人上演示。
NISE targets the sharp failure of 3D policies under realistic depth noise, carrying position uncertainty into attention; tested on transparent and reflective objects with commodity RGB-D.NISE 针对 3D 策略在真实深度噪声下的急剧失效，把位置不确定性带入注意力计算；在透明和反光物体上用普通 RGB-D 相机测试。
DISCO runs progressive-widening MCTS in GPU-parallel simulation, with a learned policy proposing continuous actions, then distills plans into a policy-value model used without search. Reports wins over BC, RL and search.DISCO 在 GPU 并行物理仿真中用渐进展宽 MCTS 搜索长程重排计划，由学习的策略提出连续动作、价值函数打分，并周期性把搜索结果蒸馏回策略-价值模型；部署时无需仿真搜索。报告在仿真与真机重排任务上优于行为克隆、强化学习与搜索基线（CoRL 2026 Oral）。
A latent motion prior distilled from human soccer MoCap lets one skill-conditioned policy move, dribble, trap and kick with human-like motion; transfers zero-shot to a Unitree G1.从人类足球动捕数据蒸馏出潜在运动先验，使单一技能条件策略能完成移动、带球、停球和射门，动作自然流畅；零样本迁移到 Unitree G1。
One command-conditioned policy for humanoid soccer skills (dribbling, trapping, shooting) with per-skill motion priors and critics; deployed on a real Noetix E1.单一指令条件策略实现人形足球多种技能（带球、停球、射门），每个技能有独立的运动先验与评论家；部署在真实 Noetix E1 上。
Three-stage hierarchical RL for humanoid badminton: sparse human strokes are augmented into a latent skill space, and a high-level planner composes skills online from the observed shuttle.三阶段分层强化学习实现人形机器人打羽毛球：把少量人类击球动作扩充为连续的潜技能空间，高层规划器再根据观测到的羽毛球在线组合技能。
Learns professional tennis serve and rally styles from broadcast video with a planner-plus-tracker design and speed adaptation for sim-to-real. Deployed on a G1 and a 1.7 m Dobot Atom serving without mocap.从比赛转播视频学习职业网球发球与对打风格，采用“规划器+跟踪器”分层结构并加入速度自适应缩小仿真-真机差距。部署在 G1 和 1.7 m 的 Dobot Atom 上，无需动捕即可发球。
CoRL 2026 Oral. Three-stage curriculum for humanoid soccer shooting: a single human-kick reference serves as a scaffold, bridging stable motion tracking and goal-driven RL that struggles to discover kicks.CoRL 2026 Oral。人形足球射门的三阶段课程：以单条人类踢球参考为脚手架，兼顾动作跟踪的稳定性与任务奖励 RL 的目标适应能力。
Carries several free-rolling balls on a wheel-legged robot's back with no fences: per-frame Deep Sets respect that ball order can change every frame, which standard history encoders miss and exploit as a shortcut.在轮腿机器人背上无围栏地同时运送多个自由滚动的球：逐帧 Deep Sets 编码器考虑到球的顺序每帧都可能变化，而常规历史拼接编码器忽略这一点并把槽位当作捷径。
Compiles human video into generalizable real-world humanoid interaction skills, without task-specific design.把人类视频转化为可泛化的真实人形交互技能，无需按任务设计。
PRISM turns a few real videos into hundreds of counterfactual human-object clips via video generation, rebuilds them in sim, and trains one humanoid carrying policy that handles unseen objects with no real fine-tuning.PRISM 用视频生成把少量真实视频扩增为数百段反事实人-物交互视频，在仿真中重建并训练单一人形搬运策略，无需真机微调即可泛化到未见物体。
MimicGen-style data generation for humanoids: adapts a few whole-body demos to new object poses and links them with locomotion planning, plus a 9-task benchmark. Co-training reports +20% over real data only.面向人形机器人的 MimicGen 式数据生成：把少量全身演示适配到新物体位姿，并用移动与操作规划串联，另提出 9 任务基准。协同训练比仅用真机数据提升 20%。
Humanoid loco-manipulation testbed pairing MuJoCo physics with Isaac Sim rendering: 16 whole-body tasks, 50 scenes, 1,000+ objects, motion-planning and VR data pipelines; benchmarks imitation, VLA and WAM policies.人形移动操作仿真平台，结合 MuJoCo 物理与 Isaac Sim 渲染：16 个全身任务、50 个室内场景、1000 多个物体资产，提供运动规划与 VR 数据管线，并评测模仿、VLA 与世界动作模型策略。
Teleop-free humanoid interaction: a fine-tuned VLM generates human whole-body motion from an egocentric image and instruction, retargeted and run by an interaction-aware tracker on a Unitree G1.无需遥操作的人形交互：微调后的 VLM 根据第一人称图像和语言指令生成人体全身运动 token，重定向到机器人后由注重交互的跟踪策略在 Unitree G1 上执行。
Generates 48,000 vision-language-kinematics trajectories in 3DGS-reconstructed rooms with no human input, trains a policy that outputs whole-body kinematics, and runs it on a real G1 via a tracker.在 3D 高斯泼溅重建的室内场景中无人工合成 4.8 万条“视觉-语言-运动学”轨迹，训练输出全身运动学轨迹的策略，并通过全身跟踪器在真实 G1 上运行。
Trained in simulation with no real demos. Controller-in-the-loop filtering rejects 73% of kinematically valid grasps that would destabilize the robot; reports 87.5% real-world success on a multi-level shelf task.完全在仿真中训练，无真实示范。把控制器放进回路筛选数据，剔除 73% 运动学可行但会使机器人失稳的抓取；作者报告多层货架任务真实成功率 87.5%。
Uses 'contact flow' (key body trajectories plus binary contact signals) as the interface between a skill-tracking policy and a planner that generates future contact flows; releases a MoCap humanoid-object dataset.以“接触流”（关键身体轨迹 + 二值接触信号）作为接口：底层策略跟踪统一的移动操作技能库，高层模块生成未来接触流来串联技能；并发布基于动捕的人-物交互数据集。
Humanoid box-rearranging challenge that chains pickup, walking and placing skills. A shared whole-body controller hardened with closed-loop skill data composes best; controller handoffs cause much of the loss.人形机器人“汉诺塔”式搬箱挑战，需要反复串联拾取、行走与放置技能。发现用技能闭环数据加固的共享全身控制器组合最稳健，控制器切换是性能损失的重要来源。
Cross-embodiment humanoid whole-body controller: shared human-motion command tokens and a causal transformer trained on mixed multi-robot rollouts, with small robot-specific modules. Nine sim bodies, four real robots.跨本体的人形全身控制基础模型：共享的人体运动指令 token 与在多机器人混合数据上训练的因果 Transformer，加上轻量的机器人专属模块。覆盖 9 种仿真本体与 4 台真机。
Humanoid that reacts to a human partner: a token model generates G1 whole-body motion conditioned on the partner, and a tracker executes it. Adds a 260-hour human-humanoid interaction benchmark.让人形机器人对人类伙伴做出回应：token 化序列模型以伙伴动作为条件生成 G1 全身动作，再由跟踪控制器执行。并提供 260 小时的人-人形交互基准。
REACT learns humanoid-human contact interactions from human-human mocap: VLM-guided contact retargeting, then onboard-depth policies that also predict the occluded partner. A real G1 does all five interactions.REACT 从人-人动捕中学习人形与人的接触交互：VLM 引导接触重定向，再蒸馏为只用机载深度的策略，并预测被遮挡伙伴的状态。真实 G1 完成全部五种交互。
Box-carrying humanoid policy conditioned on a desired root trajectory, with reference motion used only in the reward; a persistent object estimator fuses detection with a digital twin to spot slips and re-grasp.人形机器人搬箱策略以期望的根轨迹为条件，参考动作只用于奖励；持续物体估计模块融合实时检测与数字孪生，发现滑落后自动重新抓取。
Discovers long-horizon, contact-rich humanoid loco-manipulation from scratch: LLM-guided evolutionary search over interaction sequences plus kinodynamic trajectory optimization, then RL tracking on a real humanoid.从零发现长时域、多接触的人形机器人移动操作动作：由大模型引导的进化搜索交互序列，配合运动动力学轨迹优化，再训练强化学习跟踪策略部署到真实人形机器人。
Hybrid simulator (mesh foreground over a calibrated 3DGS background) that generates humanoid loco-manipulation demos without teleop. On a G1, VLAs trained only on it match or beat teleop-trained ones.混合模拟器：网格前景叠加校准过的 3D 高斯泼溅背景，无需遥操作即可生成人形移动操作演示。在 G1 上，仅用其数据训练的 VLA 达到或超过遥操作数据训练的效果。
POT-VLA keeps role-indexed 3D object records from RGB-D as tokens for the whole-body action expert and for checking whether actions succeeded. On a Unitree G1: 71/80 successes vs 39/80 for GR00T-N1.7.POT-VLA 从 RGB-D 维护按角色索引的 3D 物体记录，既作为全身动作专家的输入 token，也用于检验动作是否达成。在 Unitree G1 上成功 71/80 次，GR00T-N1.7 基线为 39/80。
One RL policy for legs, arm and gripper that adds a tactile latent from compliant grippers to proprioception, regulating grasp force and preventing slip. Zero-shot on a Unitree Go2 with a WidowX 250 arm.单一强化学习策略同时控制腿、机械臂与夹爪，把柔顺夹爪的触觉潜表示与本体感觉结合，调节握力并防止物体滑落。零样本部署在装有 WidowX 250 机械臂的 Unitree Go2 上。
Wearable whole-body tactile interface for both human demos and humanoid teleop, with a planner that predicts poses plus contact forces for a tactile admittance controller. Five contact-rich tasks.可穿戴的全身触觉接口，可用于人类演示和人形遥操作；规划器同时预测末端位姿与接触力，作为触觉导纳控制器的参考。在五个富接触任务上验证。
Modular humanoid grasping: large vision models handle the scene, a sim-trained end-effector tracker with IK and a learned forward model handles the arm. Reports 2.44 cm tracking error, 5.5x better than prior work.模块化人形抓取：大视觉模型负责场景理解，仿真训练的末端跟踪策略结合 IK 与学习的正运动学模型负责执行。报告末端跟踪误差 2.44 厘米，比先前最好方法低 5.5 倍。
Adds runtime constraints (collisions, joint limits, CoM stability) to RL humanoid tracking policies via operational-space control and CBFs at over 1 kHz, even for constraints set after training.借助操作空间控制与控制障碍函数，以 1 kHz 以上频率为强化学习人形跟踪策略施加运行时约束（碰撞、关节限位、质心稳定），训练后新增的约束也能满足。
FAST: a general humanoid motion tracker with a small residual adaptation policy under orthogonality and KL constraints for new motions without forgetting, plus center-of-mass-aware terms for balance.FAST：通用人形运动跟踪控制器，在正交与 KL 约束下学习轻量残差策略来快速适配新动作且不遗忘，并加入质心相关观测与目标提升平衡。
Low-latency humanoid teleoperation: scaled velocity feedforward cuts tracking delay by ~70 ms and link-pose control skips retargeting. End-to-end latency down to 50 ms, enough for juggling and ping-pong ball balancing.低延迟人形遥操作：缩放速度前馈让跟踪延迟减少约 70 ms，直接控制末端连杆位姿省去关节重定向。端到端延迟低至 50 ms，可完成抛接与乒乓球平衡等反应型动作。
VR humanoid teleoperation for long tasks: a robot-mounted binocular fisheye, wide stereo display, head-decoupled free-looking and stabilization keep the operator aware of the surroundings.面向长时任务的人形机器人 VR 遥操作：机载双目鱼眼相机、宽视场立体显示、与机器人运动解耦的自由环视及画面稳定，让操作者持续掌握周围空间信息。
Humanoid brachiation with passive wrist hooks: biomimetic keyframes make release-swing-capture reachable early in training, and a recurrent estimator gives position/contact latents. Continuous bar traversal on hardware.带被动腕钩的人形机器人臂荡移动：仿生关键帧让“松手-摆荡-抓杆”在训练早期就能被探索到，循环网络估计位置与接触潜变量。真机实现连续过杆。
Teleoperation for a full-size humanoid carrying heavy loads: privileged motion guidance cleans noisy VR references and a payload curriculum builds robustness. A 175 cm, 65 kg robot tracks motions with up to 24 kg.面向全尺寸人形机器人的重载遥操作：特权运动引导修正带噪 VR 参考，窗口化载荷课程提升鲁棒性。175 cm、65 kg 的机器人可携带最多 24 kg 负载跟踪动作。
Trains the low-level humanoid controller by motion imitation with the object physically in the scene, then a high-level policy on top; on a Unitree G1 it carries varied boxes and recovers from drops.让人形底层控制器在场景中真实带着物体做动作模仿训练，再在其上训练高层策略；在宇树 G1 上可搬运不同箱子，并能从掉落中恢复。
Fault-tolerant whole-body control for legged manipulators: a fault estimator finds failed leg joints from proprioception, and a posture module turns arm-driven base requests into safe ones.面向腿足机械臂的容错全身控制：故障估计器从下肢本体感觉历史中识别失效关节，姿态适配模块把机械臂策略提出的底盘姿态请求转为安全可执行的指令。
Humanoid motion tracking that follows explicit part-level contact commands as well as keypoints, so the robot really sits, wipes or pushes, and contact can be switched on or off. 10 sim motions, 5 on hardware.人形动作跟踪在关键点之外还跟踪部位级接触指令，让机器人真正坐下、擦拭或推动物体，并可按需产生或抑制接触。仿真 10 种动作，真机 5 种。
CoRL 2026 Oral. A fully virtual pipeline for humanoid loco-manipulation data: grounds video world models in 3D assets and simulator-ready scenes, with no teleoperation or physical setups.CoRL 2026 Oral。全虚拟的人形移动操作数据管线：把视频世界模型落到 3D 资产与可仿真场景中，无需遥操作或搭建实体环境。
Rather than distilling per-motion experts or brute-forcing one RL policy, trains a humanoid tracking policy jointly with one differentiable world model of shared dynamics, fit from simulator logs (CoRL 2026 oral).SuperMimic 不再蒸馏逐动作专家或用大算力硬训单一 RL 策略，而是把人形机器人动作跟踪策略与一个从仿真日志拟合的、共享动力学的可微世界模型联合训练（CoRL 2026 口头报告）。
TraCS adds interpretable, probabilistic traffic-rule logic to black-box motion predictors, generated from natural-language regulations; reports consistent gains on Argoverse 2. Oral.TraCS 将由自然语言交规生成的可解释概率逻辑加到黑盒运动预测模型上；报告在 Argoverse 2 上稳定提升。Oral 论文。
RL policies modulate a reference trajectory for compact quadrotor flips with reversible thrust, one policy per direction. In JAX simulation: 32% lower position RMSE and 57% shorter settling than baselines.强化学习策略调制参考轨迹，实现可反向推力四旋翼的紧凑翻转，正翻与回翻各训一个策略。JAX 仿真中位置 RMSE 降低 32%，稳定时间缩短 57%。
Visuomotor quadrotor policy flies through moving narrow gaps, trained in one stage from scratch by differentiable simulation; a signed-distance-gated objective turns the sparse pass-through event into a dense loss.视觉运动四旋翼策略穿越移动的狭窄缝隙，借助可微仿真从零单阶段训练；以到时变缝隙平面的有符号距离门控各项损失，把稀疏的穿越事件变成稠密目标。
Event-camera drone flight: learn the representation offline from a large dataset, then fine-tune the policy online on cheap state only, with no event rendering. Training drops from 52.44 h to 1.86 h.事件相机无人机飞行：先在大规模离线数据上学习表示，再只用轻量状态信息在线微调策略，无需在训练中渲染事件。训练时间从 52.44 小时降到 1.86 小时。
Benchmark for multirotor aerial manipulation: under-, fully and over-actuated platforms, 12 tasks, configurable wind and saturation, standard low-level controllers and learning baselines for system-level comparison.多旋翼空中操作基准：包含欠驱动、全驱动与过驱动平台，12 个任务，可配置气动扰动与执行器饱和，并提供标准底层控制器和策略学习基线，便于系统级比较。
Transfers a pretrained π0 to aerial pick-and-place: payload-aware guidance in flow sampling handles flight dynamics and 3DGS data covers navigation. Reports pick-and-place success up from 23% to 50% in 460 real trials.把预训练 π0 迁移到空中抓放：在流采样中注入载荷约束以应对飞行动力学，并用高斯泼溅合成导航数据。在 460 次真机实验中报告抓放成功率从 23% 提升到 50%。
UAV target following that conditions a flow-based trajectory generator on predicted target futures and repairs risky trajectory prefixes inside the sampling loop. Best safety-tracking trade-off; Future-MPC is smoother.无人机目标跟随：用预测的目标未来轨迹作为条件引导流式轨迹生成，并在采样循环中修复高风险的可执行前段。安全与跟踪的权衡最好，但 Future-MPC 更平滑。
GPU batch-rendering quadrotor simulator; reports 20×+ the speed of existing batch quadrotor sims, with a new ray tracer rendering RGB, depth, segmentation, normals and optical flow (CoRL 2026).以 GPU 批量渲染为核心的四旋翼仿真器；报告比现有批量四旋翼仿真器快 20 倍以上，新的光线追踪引擎（支持 OptiX、Vulkan、Metal、WebGPU）可渲染 RGB、深度、分割、法向和光流（CoRL 2026）。
One VLM acts as teacher (sees the whole trajectory, infers the goal and reasoning) and student (onboard view only), so language-conditioned navigation reasoning is learned with no human labels. 13 real tasks.同一个 VLM 既当教师（看到完整轨迹，推断导航目标与推理过程）又当学生（只看机载观测），无需人工标注即可学会以语言为条件的导航推理。在 13 个真实任务上验证。
Wed 11/11, 8:30–9:00 AM · Hilton Austin · 9 papers
CoRL 2026 Spotlight. Real-world benchmark that factorizes tabletop manipulation into motor and instruction atoms to test atomic skills and how policies recombine them.CoRL 2026 Spotlight。真实世界基准：把桌面操作分解为动作原子与指令原子，测试策略的原子技能及其重新组合能力。
Treats clarifying questions as Bayesian active learning over Signal Temporal Logic specs, so the robot asks fewer questions and hands a formal planner a verifiable task.把澄清提问视为信号时序逻辑规约上的贝叶斯主动学习，机器人提问更少，并向形式化规划器交付可验证的任务。
SCOPE gives diffusion trajectory models calibrated uncertainty tubes without Monte Carlo sampling, by distilling score curvature; tested in crowd navigation and Franka manipulation.SCOPE 通过蒸馏得分函数曲率，让扩散轨迹模型无需蒙特卡洛采样即可给出校准的不确定性管道；在人群导航和 Franka 操作上测试。
CoRL 2026 Spotlight. A multi-agent self-learning harness aimed at Variational Automation, industrial tasks with more object variation than fixed automation, which need sustained reliability.CoRL 2026 Spotlight。面向“可变自动化”的多智能体自学习框架：这类工业任务的物体变化大于固定自动化，且要求持续可靠。
Constrained decoding for LLM planners: grounded 'meta-tokens' guarantee syntax and preconditions by construction and cut decoder memory from over 107 TB to under 2 GB.面向 LLM 规划的约束解码：用接地的“元 token”从结构上保证语法与前置条件，并将解码内存从 107 TB 以上降到 2 GB 以内。
CoRL 2026 Spotlight. RATs, teams of embodied coding agents, propose exploratory tasks and learn skills through self-directed play before any downstream task arrives.CoRL 2026 Spotlight。具身编程智能体团队 RATs 在下游任务到来之前，自行提出探索任务，并通过自主玩耍学习技能。
Makes deep unfolding work for large conic solvers via matrix-free implicit differentiation and a stable PSD-projection gradient; reports up to 50× speedups on SDP-based planning.通过无矩阵隐式微分和稳定的半正定投影梯度，让深度展开适用于大规模锥优化求解器；报告在基于 SDP 的规划上最高提速 50 倍。
Wed 11/11, 9:00–9:30 AM · Hilton Austin · 9 papers
CoRL 2026 Spotlight. Hardens whole-body end-effector tracking for legged manipulators against out-of-distribution inputs such as sensor noise or infeasible commands, using a frame-wise single-step safety manifold.CoRL 2026 Spotlight。提升足式移动操作臂全身末端跟踪对分布外输入（传感器噪声、不可行指令）的鲁棒性：用逐帧安全方案把无限时域安全约束转化为计算高效的单步流形约束。
CoRL 2026 Spotlight. Language-conditioned humanoid traversal of cluttered rooms that treats navigation as whole-body adaptation of arms, torso and gait rather than 2D path planning.CoRL 2026 Spotlight。语言条件下的人形机器人杂乱环境穿行：不把导航当二维路径规划，而是手臂、躯干与步态的全身协同调整。
CoRL 2026 Spotlight. GPU physics solver in Python on NVIDIA Warp, integrated into Newton, for parallel simulation of robots with true kinematic loops rather than tree approximations.CoRL 2026 Spotlight。基于 NVIDIA Warp 的 Python GPU 物理求解器，集成于 Newton，可并行仿真含真实运动学闭环的机器人，而非树形近似。
CoRL 2026 Spotlight. A humanoid system for climbing diverse ladders and manipulating while on them; a two-stage pipeline learns climbing experts through hybrid motion tracking.CoRL 2026 Spotlight。让人形机器人攀爬多种梯子并在梯上操作；两阶段流程用混合动作跟踪学习多个攀爬专家策略。
Uses sample-based MPC in simulation as an automatic expert, then trains off-policy RL on sparse rewards, avoiding reward shaping; deployed on an arm-equipped Spot and a G1 humanoid.在仿真中用采样 MPC 自动生成专家数据，再以稀疏奖励训练离策略 RL，免去奖励塑形；已部署到带臂 Spot 和 G1 人形机器人。
CoRL 2026 Spotlight. A one-variable-at-a-time empirical study of what a whole-body native VLA needs, mapping language and pixels directly to all humanoid degrees of freedom.CoRL 2026 Spotlight。逐变量对照的实证研究：探索直接把语言与图像映射到人形全部自由度的全身原生 VLA 需要哪些条件。
CoRL 2026 Spotlight. Humanoid locomotion dataset from internet video that targets the floating, penetration and foot skating common in earlier video-derived data.CoRL 2026 Spotlight。源自网络视频的人形运动数据集，针对此前视频数据常见的悬空、穿透与脚滑问题。
Separates performance from safety on humanoids: the nominal policy handles the task while a residual RL policy acts as an implicit safety filter, avoiding competing reward terms.在人形机器人上分离性能与安全：名义策略负责任务，残差 RL 策略充当隐式安全过滤器，避免相互竞争的奖励项。
CoRL 2026 Spotlight. Brings sole pressure sensing into humanoid locomotion control, aiming for softer foot-terrain contact than parkour policies that succeed despite harsh landings.CoRL 2026 Spotlight。把足底压力感知引入人形运动控制，目标是比“能完成但落地粗暴”的跑酷策略更柔和的足地接触。
Wed 11/11, 9:30–10:00 AM · Hilton Austin · 9 papers
CoRL 2026 Spotlight. Current-centric streaming 3D reconstruction that predicts window-relative poses and local geometry in the current frame, to limit drift on long runs.CoRL 2026 Spotlight。以当前帧为中心的流式三维重建：在当前帧坐标系预测窗口相对位姿与局部几何，以抑制长时漂移。
MEND plans UAV-UGV recharge teamwork with masked diffusion, enforcing operational rules at execution through a logic-based action mask; tested on real maps and physical robot teams.MEND 用掩码扩散规划无人机与地面车协同充电任务，在执行时通过逻辑动作掩码强制满足运行约束；在真实地图和实体机器人队伍上测试。
Argues VLN success metrics hide a 'last-3-meter grounding gap': agents stop nearby without facing the target. Adds instance-level metrics and REALM, a refinement module for any backbone.指出 VLN 成功率掩盖了“最后 3 米定位缺口”：智能体停在附近却未对准目标。提出实例级指标和可接入任意主干的修正模块 REALM。
Neural inertial odometry across carriers: a router composes carrier-specific experts, and a light residual branch adapts to new ones. Reports 40% lower ATE than learned baselines.面向多种载体的神经惯性里程计：路由器组合各载体专家，新载体通过轻量残差分支适配。报告 ATE 比学习型基线降低 40%。
CoRL 2026 Spotlight. Decides whether to wait or reroute at a blocked route by modeling how long each obstacle class tends to persist.CoRL 2026 Spotlight。通过建模各类障碍物通常停留多久，决定在路线受阻时等待还是改道。
Shows termination-based constrained RL can learn to avoid the goal when constraints tighten near it, and fixes this with a dense success critic; validated on spacecraft docking tasks.发现基于终止的约束强化学习在目标附近约束收紧时会学会回避目标，并用稠密成功信号的辅助价值网络解决；在航天器对接任务上验证。
CoRL 2026 Spotlight. Benchmark for visual navigation on urban sidewalks, built on Isaac Sim with procedurally generated and real-world environments.CoRL 2026 Spotlight。城市人行道视觉导航基准，基于 Isaac Sim，包含程序生成与真实世界环境。
Simulator that models raw measurement physics for event, SPAD, iToF/dToF, LiDAR and thermal sensors, reproducing failure modes like pileup, phase wrapping and refractory drops; releases an indoor dataset.仿真平台对事件相机、SPAD、iToF/dToF、激光雷达和热成像的原始测量物理过程建模，复现堆积、相位缠绕、不应期丢事件等失效模式；并发布室内数据集。
Wed 11/11, 10:30 AM–12:00 PM · JW Marriott · 228 papers · Lone Star (3rd floor): posters 1–116 · Griffin (2nd floor): posters 117–232
Hierarchical MPC on latent visual world models at two time scales, linked by latent subgoals and learned macro-actions. Reports 70% vs 0% on real Franka pick-and-place from one goal image, with up to 3x less compute.在两个时间尺度的潜在视觉世界模型上做分层 MPC，以潜在子目标和学习到的宏动作衔接。报告真实 Franka 单目标图抓放成功率 70%（单层规划为 0%），计算量最多少 3 倍。
Predicts task-relevant future states in a pretrained prediction-oriented feature space (no pixel reconstruction) and conditions actions on them via causal attention. +11.8 pts over reactive policies on RoboTwin.在预训练的面向预测的特征空间中预测与任务相关的未来状态（不做像素重建），并通过结构化因果注意力让动作生成以其为条件。在 RoboTwin 困难长时任务上比反应式策略高 11.8 个百分点。
Action-conditioned world-action model that also learns from failed rollouts, predicting future video and task progress for the action actually taken; progress scores can rank sampled actions.以实际执行动作为条件的世界动作模型，把失败轨迹也作为监督，预测未来视频与任务进度；进度分可在推理时为候选动作排序。
One video-diffusion backbone serves as both a video-action policy and an action-conditioned simulator that scores progress, trained on ~27,300 hours of robot, UMI and human data; samples and ranks actions at test time.同一个视频扩散主干既是视频-动作策略，也是可预测任务进度的动作条件模拟器；训练数据约 2.73 万小时（真机、UMI、人类视频等），推理时采样并排序候选动作。
Turns a pretrained video diffusion transformer into a policy by encoding each action step as one patch in the video latent space; at test time it outputs an action chunk in one DDIM step (84 ms), no video generation.把预训练视频扩散 Transformer 改造成策略：每个时间步的动作编码为视频潜空间中的一个 patch，联合训练视频、视频+动作与纯动作三种扩散目标；测试时一步 DDIM 生成动作块（84 毫秒），无需生成视频。
Steers a frozen world-action model with unlabeled human videos at test time: a small memory inside the model adapts via self-supervised video prediction, after meta-training on paired human-robot data.测试时用无标注的人类视频引导冻结的世界动作模型：模型内部的小型记忆通过自监督视频预测进行适配；此前用成对人-机数据做元训练使其对控制有用。
Factorizes an interaction into action, camera and embodiment so one human video can be re-rendered as many robot rollouts. Generated data helps on RoboCasa GR1 and a real IRON humanoid.把一次交互分解为动作、相机与本体三个因素，使一段人类视频可重新生成为多种机器人视角的数据。生成数据在 RoboCasa GR1 与真实 IRON 人形上带来提升。
Video world model that renders the robot's own motion from its URDF and controller, so the network only predicts how the scene responds. Transfers to unseen robots and can animate retargeted human demos.视频世界模型把机器人自身运动交给控制器和 URDF 渲染，网络只需预测场景的响应。可迁移到未见过的机器人本体，并能由重定向的人类演示生成机器人操作视频。
Probe study of which video world model latents carry action information: temporal video pretraining matters, pixel reconstruction quality does not. V-JEPA-style encoders give the best trade-off.探测哪些视频世界模型的潜变量含有动作信息：关键在于时序视频预训练，而非像素重建质量；V-JEPA 类编码器取得最佳权衡。
Do world action models need video generation? ImageWAM conditions a flow-matching action expert on image-editing KV caches instead. Reports beating VLA baselines at 1/6 the FLOPs and 1/4 the latency of video WAMs.追问世界动作模型是否真需要视频生成：ImageWAM 改用图像编辑模型去噪时的 KV 缓存来条件化流匹配动作专家；报告优于 VLA 基线，计算量为视频 WAM 的 1/6、延迟为 1/4。
World action model that predicts oriented bounding-box trajectories of task objects instead of future video frames: much cheaper, and reported to predict actions better.世界动作模型不再预测未来视频帧，而是预测任务相关物体的有向包围盒轨迹：计算量大幅下降，并报告动作预测更好。
Object-centric world-action model that both takes object masks as prompts and predicts future masks, reducing language ambiguity in clutter and background bias in RGB prediction.以物体为中心的世界动作模型：既把目标物体掩码作为提示输入，又预测未来掩码，从而减少杂乱场景中的语言歧义和 RGB 预测中的背景偏差。
3D world model that encodes each object's point cloud with an SDF-supervised vector-neuron VAE and predicts per-object SE(3) increments applied analytically, so shapes do not drift over long rollouts.3D 世界模型：用 SDF 监督的向量神经元 VAE 编码每个物体的点云，并预测逐物体的 SE(3) 运动增量后解析地施加，长时推演中物体形状不会漂移。
A large pretrained world model plans subtasks and generates goal images; a VLA follows them. Reports the same VLA rising from 14% to 69% in novel scenes with these generated goals.用大规模预训练世界模型做高层规划，把任务拆成子任务并生成目标图像，由底层 VLA 跟随执行。报告同结构 VLA 在新场景中的成功率借此从 14% 提升到 69%。
Defines Temporal Ratio, how much an action head attends to future latents vs the current frame; it predicts compositional generalization of video-action models and drives an inference-time adaptive guidance method.提出 Temporal Ratio：动作头对未来潜变量相对当前帧的注意力依赖程度；它能预测视频动作模型的组合泛化能力，并据此设计推理时的自适应引导方法。
Adds a training-only head that predicts 3D point tracks from VLA features, teaching world dynamics with no inference cost. Reports +10% on LIBERO-Long and +40% on RoboCasa.训练时为 VLA 加一个预测 3D 点轨迹的轻量头，让模型学到场景动力学，推理时去掉、零额外开销。报告 LIBERO-Long 提升 10%，RoboCasa 提升 40%。
AHEAD wraps a frozen VLA with a small latent world model that forecasts future patch tokens from optical-flow velocity and feeds them to the action decoder. +4.9M params on a 7B OpenVLA; 79-97% on 20 dynamic sim scenes.AHEAD 为冻结的 VLA 加上一个小型潜空间世界模型，依据光流速度预测未来图像块 token 并交给动作解码器，以应对运动物体。在 7B OpenVLA 上只增加 490 万参数，20 个动态仿真场景成功率 79–97%。
Before contact, a latent world model rolls the policy's planned action chunk forward in multi-view embedding space, and a small probe aborts if the predicted post-contact state looks like failure.在接触发生前，用潜空间世界模型在多视角嵌入空间中推演策略计划的动作块，若预测的接触后状态呈现失败迹象，轻量探针即中止执行。
EMMA: a diffusion-transformer video editor (DreamTransfer) changes foreground, background and lighting with multi-view consistency, and AdaMix weights real and generated samples by policy performance.EMMA：扩散 Transformer 视频编辑器 DreamTransfer 在保持多视角一致的前提下修改前景、背景与光照，AdaMix 再按策略表现为真实与生成样本自适应加权。
Recalibrates DROID into DROID-3D with metric depth and multi-view extrinsics, then trains a video world model that predicts RGB and depth. Depth supervision reportedly also improves RGB by 1.48 dB PSNR.把 DROID 重新标定为带度量深度与多视角外参的 DROID-3D，并训练同时预测 RGB 与深度的视频世界模型。报告深度监督还让 RGB 预测提升 1.48 dB PSNR。
RL post-training of VLAs inside a video world model, with keyframe-initialised rollouts and world-model–policy co-evolution to keep hallucination in check. Reports strong LIBERO results and real-robot gains.在视频世界模型中对 VLA 做强化学习后训练，通过关键帧初始化的短推演和世界模型-策略协同演化来抑制幻觉。报告 LIBERO 表现优异并在多个真机平台上提升。
Trains an action-conditioned video world model purely from autonomous robot play instead of human demos, capturing failure-rich contact. Reports up to 40% better policy evaluation and +65% real success after RL in it.完全用机器人自主“玩耍”数据（而非人类演示）训练动作条件视频世界模型，覆盖更多接触与失败情形。报告策略评估提升最高 40%，在其中做强化学习后真机成功率提升 65%。
RL inside a learned diffusion world model without a simulator: the big model rolls out trajectories while a light latent surrogate supplies gradients. Beats PPO on sample efficiency on Push-T; quadruped task too.不用物理仿真器，直接在学习到的扩散世界模型中训练策略：大模型负责生成轨迹，轻量潜在代理模型提供一阶梯度。在 Push-T 上样本效率明显优于 PPO，并在四足任务上验证。
Online safe RL for VLAs inside an interactive video world model: a success classifier and a contact-probability safety head are balanced with Lagrangian-constrained GRPO. +8.25 pts safe success on SafeLIBERO Level I.在可交互视频世界模型中对 VLA 做在线安全强化学习：成功分类器与预测接触概率的安全头，经拉格朗日约束的 GRPO 平衡。在 SafeLIBERO Level I 上安全成功率提升 8.25 个百分点。
Names two failure modes of latent safe control, estimation gaps (safety facts not visible) and prediction gaps (failures not foreseeable), gives a diagnostic for each and mitigations.指出潜空间安全控制的两种失效：估计缺口（当前观测看不到安全关键量）与预测缺口（失败无法提前预见），并分别给出诊断指标和缓解方法。
Shared latent action space grounded by predicting the same future across data sources, so action-free and differently-labelled data can train one BC policy. Reports +48% average success on five tasks.通过让不同数据源对未来观测的预测保持一致来“落地”一个共享潜在动作空间，使无动作标签和动作空间不同的数据都能训练同一个行为克隆策略。报告五个任务平均成功率提升 48%。
Learns STRIPS-style operators (preconditions, add/delete effects) over latent predicates from image transitions, then plans classically from start and goal images. Beats visual-rollout and latent-search methods.直接从图像转移数据学习 STRIPS 式符号算子（前提条件、增删效果）和潜在二值谓词，再从起始图与目标图用经典规划求解。优于视觉推演和潜在图搜索等基线。
Inverse RL from observation only (videos, no actions or rewards) made practical on a real robot via online world modeling. Reports pick-and-place learned from scratch to 82% in under 40 minutes.仅从观测（视频，无动作和奖励）做逆强化学习，借助在线世界模型在真机上实现。报告不到 40 分钟从零学会抓放，成功率 82%。
World model that predicts 3D traces of key points (objects, tools, hands) instead of pixels or actions, pretrained from video without action labels. Trace-conditioned policies are reported competitive with π0-style VLAs.不预测像素也不预测动作，而是预测物体、工具、手等关键点的 3D 轨迹，可用无动作标签的视频预训练。报告以轨迹为条件的策略可与 π0 类 VLA 相当。
Training-free speedup for world-action models: reuses denoising residuals across consecutive inference chunks at the same step, since they correlate during smooth motion. Up to 2.5x with a Fast-WAM backbone.无需训练的世界动作模型加速方法：平滑运动时相邻推理块在同一去噪步的残差高度相关，因此跨块缓存复用。以 Fast-WAM 为骨干最高提速 2.5 倍。
SKIP generates only task-relevant keyframes (approach, contact, grasp, release) with a sparse video diffusion model, then fills the gaps with an action-conditioned interpolator. 4.16x faster dense rollouts on LIBERO.SKIP 只用稀疏视频扩散模型生成与任务相关的关键帧（接近、接触、抓取、释放），再用动作条件插值器补全中间帧。在 LIBERO 上生成稠密推演快 4.16 倍。
Steers a video world model toward rare but plausible bad outcomes (e.g. task failures) by optimizing its initial noise with a VLM objective, to evaluate and improve policies. Driving and manipulation.通过优化视频世界模型的初始噪声（以 VLM 为目标），把想象引向罕见但合理的不利结果（如任务失败），用于评估和改进策略；覆盖自动驾驶与操作。
Why action-conditioned video models ignore actions: expert demos carry too little action influence. AIM actively collects data to raise it; 500 such episodes + 3k demos match a 300k-demo model on controllability.解释动作条件视频世界模型为何忽略动作：专家示教的“动作影响力”太低，模型容量再大也无法弥补。AIM 主动采集数据提高这一指标；报告 500 条主动采集数据加 3000 条示教即可在末端可控性上追平 30 万条专家数据。
Detects robot anomalies from external cameras only by estimating dynamics and predicting states. On 240 real crane trajectories: +26% anomaly classification, and 88% accuracy locating the degrading motor.仅凭外部相机，通过估计动力学参数与预测状态来检测机器人异常。在 240 条真实起重机轨迹上异常分类提升 26%，定位退化电机的准确率 88%。
Retargets human whole-body demos to a mobile manipulator offline with a closed-form shoulder-elbow-wrist solver and lazy base control; claims the first zero-shot whole-body mobile manipulation from them.离线把人类全身演示重定向到移动操作机器人：闭式肩-肘-腕几何求解器加“惰性”底盘控制，得到一致的动作；声称首次仅凭离线人类演示实现零样本全身移动操作。
Low-cost upper-limb exoskeleton that records whole-arm poses and joint torques with real-time haptic torque feedback; retargets to OpenArm, Franka and xArm for learning compliant whole-body policies.低成本上肢外骨骼：记录整臂构型与关节力矩，并提供实时力矩触觉反馈；可重定向遥操作 OpenArm、Franka 和 xArm，用于学习主动柔顺的全身策略。
Wearable dexterous UMI: a shared end-effector module (light hand, in-hand camera, fingertip touch) driven by an isomorphic glove, so collected and deployed observations and actions match. Eight real tasks.可穿戴的灵巧 UMI：共享的末端模块（轻量灵巧手、手内相机、指尖触觉）由同构手套驱动，采集与部署时的观测和动作完全一致，无需重定向。在八个真机任务上验证。
Low-calibration hand retargeting for teleop: self-supervised fingertip correspondence plus a few human-guided anchors, and a contact classifier to refine pinch poses; tested on several dexterous hands.面向遥操作的灵巧手动作重定向，几乎无需标定：自监督学习指尖对应关系，再用少量人工引导的锚点约束映射，并以接触分类器细化捏取姿态；在多种灵巧手上验证。
Few-shot calibration of biased mocap gloves or hand trackers: a HyperNet reads a few paired poses and predicts LoRA updates for a frozen MANO estimator, with no per-user fine-tuning.对存在偏差的动捕手套或手部追踪做少样本校准：HyperNet 读取少量成对姿态，为冻结的 MANO 手部估计模块预测 LoRA 式更新，无需逐用户微调。
Teleoperation controller that turns operator intent into learned hand-object contact execution, trained with single-stage RL on human motions. Reports 75% average success on seven in-hand and tool-use tasks.遥操作控制器把操作者意图转化为学习到的手-物接触执行，基于人类动作用单阶段强化学习训练，零样本迁移真机。报告七个手内操作与工具使用任务平均成功率 75%。
From one human demo: retarget hand-object motion to a robot reference, train residual RL in sim to track object keypoints, transfer zero-shot with careful system ID. Scissors and screwdriver use on two hands.从一条人类演示出发：把手-物运动重定向为机器人参考，在仿真中用残差强化学习跟踪物体关键点，并通过细致的系统辨识零样本迁移。在两种多指手上完成用剪刀、拧螺丝刀等任务。
CHORD guides dexterous RL by comparing human and robot motions in object-centric contact wrench space. Also a 4,739-task bimanual benchmark built from MoCap and videos; 82.12% average on 1,831 tasks.CHORD 在以物体为中心的接触力旋量空间中比较人与机器人的动作来引导灵巧操作强化学习，并构建来自动捕与视频的 4739 个双手任务基准；在其中 1831 个任务上平均成功率 82.12%。
Learns object-object interaction priors from human videos and composes them with robot feasibility reasoning and motion planning, so one prior serves a Franka and an RBY-1 mobile humanoid without robot data.从人类视频学习以物体间交互运动表示的先验，再与机器人可行性推理和运动规划组合，同一先验无需机器人数据即可用于 Franka 与 RBY-1 移动人形机器人。
Real-to-sim benchmark for human-to-robot transfer: real human video demos plus simulated robot demos on four tasks, with a fixed protocol to compare how H2R methods scale with more human data.评估“人到机器人”迁移的 Real2Sim 基准：四个任务配有真实人类视频示教和仿真机器人示教，以统一协议比较各方法随人类数据增加的扩展能力。
Treats head-camera motion in egocentric human video as a viewpoint action and pretrains active perception plus manipulation from it. Reports matching state-of-the-art models pretrained on robot data.把第一人称人类视频中的头部相机运动视为“视角动作”，同时预训练主动感知与操作能力。报告效果与使用机器人数据预训练的最先进模型相当。
Reconstructs hand-object interaction from ordinary monocular human videos and retargets it into executable multi-fingered robot actions; beats prior work on hand-object estimation and gives a data-collection playbook.从普通单目人类视频重建手-物交互，并重定向为多指机器人手可执行的动作序列；手-物估计优于已有方法，并给出采集人类数据的实践建议。
Turns monocular egocentric human videos into robot training data: tracks hand and object, retargets to a gripper, and renders photorealistic wrist-camera views with Gaussian Splatting.把单目第一人称人类视频转为机器人训练数据：跟踪手与物体，重定向到机械夹爪，并用高斯泼溅渲染逼真的腕部相机视角。
AdvDex trains a dexterous VLA on human and robot demos: OmniShare human data with touch, a joint-aligned action space (wrist pose + 15 finger joints) shared by hands and grippers, and adversarial removal of body cues.AdvDex 用人类与机器人示教训练灵巧操作 VLA：含触觉的人类示教数据集 OmniShare、人手/灵巧手/夹爪共用的关节对齐动作空间（腕部位姿 + 15 个手指关节），并以对抗学习去除本体相关特征。
Learns a short-horizon 'what should happen next' intent model from internet human videos and pairs it with embodiment-specific policies trained in parallel sim; the same intent model drives a hand and a gripper.从互联网人类视频学习预测“场景接下来该发生什么”的短时意图模型，再与在大规模并行仿真中训练的本体专用策略结合；同一意图模型可驱动灵巧手和平行夹爪。
Makes UMI data usable for VLAs: a fisheye wrist-view VQA dataset aligns the VLM to distorted views, and a physics validation pipeline scores trajectories for kinematic limits, collisions and controller bandwidth.让 UMI 数据可用于训练 VLA：构建腕部鱼眼视角的 VQA 数据集使 VLM 适应畸变画面，并用物理校验流程对轨迹的运动学极限、碰撞与控制带宽进行评分筛选。
Human and robot wear the same hand exoskeleton, so encoders and wrist cameras match between collection and rollout; retargeting becomes paired supervision. 3.0x faster collection, 70.0% success on five tasks.人与机器人佩戴同一套手部外骨骼，采集与执行时关节编码器和腕部相机看到相同的机构，重定向变成成对监督。数据采集效率是外骨骼遥操作的 3.0 倍，五个任务平均成功率 70.0%。
Annotators define their own preference axes in language (speed, safety, care) and compare trajectories per axis; a language-conditioned reward trains a steerable policy. Reports +38 points over sparse and binary rewards.让标注者用自然语言自定义偏好维度（速度、安全、细致等）并逐维比较轨迹，训练语言条件奖励模型和可引导的策略。报告比稀疏奖励与二元偏好方法高 38 个百分点。
WACL learns from human demos to push and hold with the whole arm, choosing where on the arm to make contact while keeping the gripper's constraints.WACL 从人类示范中学习用整条手臂推、抵物体，自主选择手臂上的接触位置，同时满足末端执行器的约束。
Gaze recorded in VR teleop is drawn as crosshairs on training images; at test time a small predictor supplies the gaze. With π0: 26.3% → 56.0% on six real bimanual tasks. Releases GazeMani (1,200 trajectories).把 VR 遥操作中记录的注视点画成十字准星叠加在训练图像上，部署时由轻量预测器估计注视点。基于 π0，六个真实双臂任务平均成功率从 26.3% 升至 56.0%；发布 GazeMani 数据集（1200 条轨迹）。
SDPG estimates policy gradients from random perturbations of rollouts, so visual RL needs far fewer batch-rendered envs and trains in hours on one RTX 4080; adds visual benchmarks and sim-to-real demos.SDPG 通过对推演轨迹随机扰动来估计策略梯度，视觉强化学习所需的批量渲染环境大幅减少，单张 RTX 4080 几小时即可训练；并提供视觉机器人基准与仿真到真实演示。
DDPG-style RL for flow-matching policies without backprop through the ODE: the critic gradient at a one-step action estimate is distilled into the velocity field. Reports 92% on dual-arm AirPods assembly.无需沿 ODE 反向传播的流匹配策略 DDPG 式强化学习：在一步估计的动作上计算 critic 梯度，并蒸馏进速度场。报告双臂 AirPods 装配成功率 92%，炒蛋任务评分 86%。
RL-Co: SFT on mixed real and sim demos, then RL in simulation with an auxiliary supervised loss on real data to avoid forgetting. Tested on four real tabletop tasks with OpenVLA and π0.5.RL-Co：先在真实与仿真混合示教上做监督微调，再在仿真中强化学习，同时加入真实数据的辅助监督损失以防遗忘。用 OpenVLA 和 π0.5 在四个真实桌面任务上验证。
Residual RL on top of a frozen VLA using only object poses, so it transfers from sim to real; a sim twin of the VLA is trained by replaying the same teleop demos. Reports real FR3 success from 42% to 76% zero-shot.在冻结的 VLA 之上做只依赖物体位姿的残差强化学习，从而可直接从仿真迁移到真机；通过回放同一批遥操作数据训练 VLA 的仿真副本。报告真实 FR3 上成功率从 42% 提升到 76%。
PPO-style fine-tuning for diffusion and flow policies that builds a surrogate policy ratio from forward noising and the score/flow loss, avoiding likelihoods. Tested in sim and real.为扩散与流匹配策略做 PPO 式微调：用前向加噪与得分/流匹配损失构造替代的策略比率，绕开似然计算；在仿真与真机上验证。
OTQL post-trains and speeds up a suboptimal flow policy from the robot's own experience with advantage-weighted conditional optimal-transport flow matching, in 50-60 episodes and without distillation.OTQL 用优势加权的条件最优传输流匹配，从机器人自身经验对次优流策略做强化学习后训练并加速推理，只需 50–60 个交互回合，无需蒸馏。
Self-supervised progress reward learned from time-warped demos (speed changes, reversals), used to filter and reweight behaviour-cloning data. Reports up to ~18x T-shirt folding throughput over vanilla BC.用时间扭曲（变速、倒放）增强的演示自监督学习帧级进度奖励，再用它过滤和加权行为克隆数据。报告 T 恤折叠吞吐量最高为普通 BC 的约 18 倍。
CPU-parallel simulation with GPU learning, joined by one runtime for data movement and sync. Reports 3–10x end-to-end training speedups on the same hardware and runs on macOS, ROCm and Intel XPU.CPU 并行仿真 + GPU 策略学习的异构训练架构，用统一运行时处理数据搬运与同步。报告同等硬件下端到端训练效率提升 3–10 倍，并支持 macOS、ROCm 与 Intel XPU。
Shows termination-based constrained RL can learn to avoid the goal when constraints tighten near it, and fixes this with a dense success critic; validated on spacecraft docking tasks.发现基于终止的约束强化学习在目标附近约束收紧时会学会回避目标，并用稠密成功信号的辅助价值网络解决；在航天器对接任务上验证。
Steers a pretrained flow-matching VLA with a spatial cue from a grounding model: a small adapter predicts an affine transform of the action chunk toward the cued target. Up to ~2x task success for open-world delivery.用视觉定位模型给出的空间提示引导预训练流匹配 VLA：轻量适配器预测作用在动作块上的仿射变换，使动作对准目标。开放世界配送任务成功率最高提升近 2 倍。
Uses suboptimal demos only at high and low diffusion noise levels, where their coarse or local structure still helps. Covers noisy, sim, mismatched and mixed data; reports up to 33% over co-training on Open X-Embodiment.只在扩散过程的高噪声与低噪声阶段使用次优演示，此时其粗粒度或局部结构仍然有用。覆盖带噪、仿真、任务不匹配和混合数据；报告在 Open X-Embodiment 上比协同训练基线高最多 33%。
Generated robot videos keep geometry but not control, so GRA uses them only to supervise visual perception (aligned to extracted geometry) and leaves action learning to real demos.生成的机器人视频保留了几何轨迹却丢失了控制信号，因此 GRA 只用它们监督视觉感知（与提取的几何对齐），动作学习仍交给真实示教。
In-context imitation policy trained only on RGB synthetic data, with next-subgoal image prediction. Reports 79% average success on 16 unseen real tasks from a single test-time demo.完全用 RGB 合成数据训练的上下文模仿学习策略，并预测下一个子目标图像。报告仅给一条演示时，在 16 个未见真机任务上平均成功率 79%。
Turns one RGB image into a sim-ready scene: collision-aware foreground assets plus a 3DGS background. Supports synthetic data and sim-real-correlated evaluation; releases DROID-Sim from 564 DROID scenes.把单张 RGB 图像变成可用于仿真的场景：前景为碰撞可靠的资产，背景用 3D 高斯泼溅保持外观。可用于合成训练数据和与真机相关的策略评测；发布由 564 个 DROID 场景构建的 DROID-Sim。
Builds sim-ready digital twins from a video and edits them into 'digital cousins'. Sim scores predict real results (mean Pearson 0.911); task cousins add 40% real success for sim-trained policies.从一段视频自动构建可仿真的数字孪生，并编辑出保持可供性的“数字表亲”场景。仿真评测与真机结果高度相关（平均 Pearson 0.911）；任务表亲让仿真训练策略的真机成功率提升 40%。
Agentic generation of functional object compositions (e.g. parts that must fit): disentangled mesh reconstruction, relative pose refinement under geometric and physical guidance, and a VLM verifier.智能体式生成具备功能关系的物体组合：先生成实例分离的网格，再在几何与物理引导下细化相对位姿，最后由 VLM 验证器挑选最符合描述的候选。
Uses image editing of simulator renders as a middle step: edit in task semantics, recover 3D state from simulator priors, refine in physics, and turn it into robot supervision, e.g. dexterous grasps and goal states.以对仿真渲染图的图像编辑为中间表示：先编辑注入任务语义，再借助仿真先验恢复 3D 状态并在物理中细化，最后转为机器人可执行的监督，如灵巧抓取与目标状态。
ExpertGen trains a diffusion prior on imperfect demos (even LLM-synthesized), then RL steers its initial noise with the policy frozen, keeping behaviour human-like under sparse rewards; experts feed sim-to-real BC.ExpertGen 先在不完美示教（甚至由大模型合成）上训练扩散先验，再冻结策略、用强化学习只优化初始噪声，在稀疏奖励下保持类人行为；得到的专家策略用于仿真到真实的行为克隆。
Compositional simulation: classical sim produces accurate action-video pairs and a neural simulator restyles them as real-looking video, in a real-sim-real loop seeded by a little real data.组合式仿真：经典仿真生成准确的动作-视频对，再由神经模拟器转换成接近真实的画面，形成以少量真机数据为起点的“真-仿-真”数据扩增闭环。
Generative augmentation from a single box-annotated image, plus a region-contrastive loss that steers attention to task regions. Beats augmentation baselines on UR5e, AgileX and Tian Gong 2.0 under scene shifts.只需一张带框标注的图片即可做生成式语义增强，并用区域对比损失把注意力引向任务相关区域。在 UR5e、AgileX 与天工 2.0 上，面对背景、干扰物和光照变化时优于其他增强方法。
Auto-labels robot demos with boxes, object trajectories and phase labels, plus a reliability score from the task's spatio-temporal structure, so fewer noisy labels are kept and fewer good ones are thrown out.为机器人示教自动标注边界框、物体轨迹与操作阶段，并利用任务的时空结构为每个标注给出可靠性分数，从而少留噪声标签、少丢有用样本。
A VLM agent explores simulated tasks by calling reusable RL-trained interaction skills, verifying outcomes and keeping memory in a verifier-guided tree search, to generate robot data without scripts.VLM 智能体在仿真中调用可复用的强化学习交互技能来探索任务，借助验证、反思与记忆进行验证器引导的树搜索，无需脚本即可生成机器人数据。
GUILD decides what to collect next: a VLM proposes environments, object setups and goals, collected data is evaluated for its value to the policy, and the loop repeats, instead of filtering after the fact.GUILD 主动决定下一步采集什么数据：由 VLM 提出环境、物体布置与目标，采集后评估数据对策略的价值，再进入下一轮，而非事后筛选数据集。
Influence functions at billion-parameter VLA scale (about 313x faster via Kronecker structure and a rank-r Hessian approximation) for multi-task data curation; matches full-data fine-tuning with 50% of demos in sim.将影响函数扩展到十亿参数级 VLA（借助 Kronecker 结构与秩 r 的 Hessian 近似提速约 313 倍）用于多任务数据筛选；在仿真中只用 50% 示教即可达到全量数据微调的效果。
MetaGenSim learns which simulated data to generate, using meta-gradients from real-world performance, instead of sampling the simulator uniformly.MetaGenSim 利用来自真实世界表现的元梯度，学习应该生成哪些仿真数据，而不是对仿真器做均匀采样。
Open-source data lifecycle layer for robot policy iteration: rollouts, reviews, datasets, training runs and evaluations become typed lineage artifacts that agents maintain, making iteration faster and auditable.开源的机器人策略迭代数据生命周期管理层：把执行记录、审核、数据集决策、训练与评测都表示为带类型的溯源工件，由智能体维护，让迭代更快、更可追溯。
Simulator that models raw measurement physics for event, SPAD, iToF/dToF, LiDAR and thermal sensors, reproducing failure modes like pileup, phase wrapping and refractory drops; releases an indoor dataset.仿真平台对事件相机、SPAD、iToF/dToF、激光雷达和热成像的原始测量物理过程建模，复现堆积、相位缠绕、不应期丢事件等失效模式；并发布室内数据集。
CoRL 2026 Spotlight. GPU physics solver in Python on NVIDIA Warp, integrated into Newton, for parallel simulation of robots with true kinematic loops rather than tree approximations.CoRL 2026 Spotlight。基于 NVIDIA Warp 的 Python GPU 物理求解器，集成于 Newton，可并行仿真含真实运动学闭环的机器人，而非树形近似。
CoRL 2026 Spotlight. RATs, teams of embodied coding agents, propose exploratory tasks and learn skills through self-directed play before any downstream task arrives.CoRL 2026 Spotlight。具身编程智能体团队 RATs 在下游任务到来之前，自行提出探索任务，并通过自主玩耍学习技能。
CoRL 2026 Spotlight. A multi-agent self-learning harness aimed at Variational Automation, industrial tasks with more object variation than fixed automation, which need sustained reliability.CoRL 2026 Spotlight。面向“可变自动化”的多智能体自学习框架：这类工业任务的物体变化大于固定自动化，且要求持续可靠。
Makes deep unfolding work for large conic solvers via matrix-free implicit differentiation and a stable PSD-projection gradient; reports up to 50× speedups on SDP-based planning.通过无矩阵隐式微分和稳定的半正定投影梯度，让深度展开适用于大规模锥优化求解器；报告在基于 SDP 的规划上最高提速 50 倍。
SCOPE gives diffusion trajectory models calibrated uncertainty tubes without Monte Carlo sampling, by distilling score curvature; tested in crowd navigation and Franka manipulation.SCOPE 通过蒸馏得分函数曲率，让扩散轨迹模型无需蒙特卡洛采样即可给出校准的不确定性管道；在人群导航和 Franka 操作上测试。
Treats clarifying questions as Bayesian active learning over Signal Temporal Logic specs, so the robot asks fewer questions and hands a formal planner a verifiable task.把澄清提问视为信号时序逻辑规约上的贝叶斯主动学习，机器人提问更少，并向形式化规划器交付可验证的任务。
Constrained decoding for LLM planners: grounded 'meta-tokens' guarantee syntax and preconditions by construction and cut decoder memory from over 107 TB to under 2 GB.面向 LLM 规划的约束解码：用接地的“元 token”从结构上保证语法与前置条件，并将解码内存从 107 TB 以上降到 2 GB 以内。
Scores how much a small action error at each demo state would hurt the outcome, using an adversarial disturbance policy. The scores reweight behavior cloning and select smaller, higher-value demo subsets.用对抗扰动策略衡量每个示教状态上的小动作误差会对结果造成多大损害。该分数可用于给行为克隆加权，并挑选更小但价值更高的示教子集。
CoRL 2026 Spotlight. Real-world benchmark that factorizes tabletop manipulation into motor and instruction atoms to test atomic skills and how policies recombine them.CoRL 2026 Spotlight。真实世界基准：把桌面操作分解为动作原子与指令原子，测试策略的原子技能及其重新组合能力。
Learns a neural actuator model from joint trajectories alone, combined with rigid-body sim, to transfer RL policies to a pneumatic-muscle arm. Claims the first sim-to-real transfer for a 4-DoF muscle-actuated arm.仅用关节位置轨迹学习神经执行器模型，结合刚体仿真，把强化学习策略迁移到气动人工肌肉机械臂上。声称首次实现 4 自由度肌肉驱动机械臂的仿真到真机迁移。
Puts tendon-driven continuum robots natively into MuJoCo with a continuum-mechanics-based discretization, validated against Cosserat rods and hardware; sim-trained imitation policies transfer zero-shot to a real robot.以连续介质力学推导的离散化方法，把肌腱驱动连续体机器人原生放进 MuJoCo，并对照 Cosserat 杆模型与真机验证；仿真中训练的模仿策略可零样本迁移到真机富接触任务。
Meta-learned policy and forward model share a belief latent; at deployment the latent is updated from action-outcome mismatches, so the policy adapts to unknown dynamics on the fly.元学习的策略与前向动力学模型共享信念潜变量；部署时根据动作结果的偏差更新该潜变量，使策略能即时适应未知动力学。
Learned dynamics that ignore control changes can hide out-of-distribution actions. A support-conditioned control-sensitivity regularizer keeps predictions responsive, improving OOD detection and safer planning.对控制变化不敏感的学习动力学会掩盖分布外动作。基于数据支撑的控制敏感性正则保持模型对控制的响应，从而提升分布外检测并使闭环规划更安全。
A flow-matching policy and an action-conditioned dynamics model co-sample an action and its predicted future until they agree. Reports 56.5% (sim) and 33.3% (real) relative gains over the strongest baselines.流匹配策略与动作条件动力学模型联合采样动作及其未来观测，相互修正直到二者自洽。报告相对最强基线在仿真提升 56.5%、真机提升 33.3%。
Learns how a known symmetry acts on pixel observations and actions from offline data, then uses it for frame-averaged behavior cloning. Reports average success 80.5% to 85.1% on four Robosuite tasks.从离线交互数据中学习已知对称群如何同时作用于像素观测和动作，再用于帧平均行为克隆。报告在四个 Robosuite 任务上平均成功率从 80.5% 升到 85.1%。
Dexterous food cutting from a handful of real demos: reconstruct objects, randomize physical parameters within category limits and generate augmented demos in sim, then transfer back to the real hand.只用少量真实示教学习灵巧切割食物：重建物体实例，在类别约束下随机化物理参数，在仿真中生成扩增的灵巧示教，再迁移回真实灵巧手。
Task-agnostic 'any pose to any pose' dexterous policy trained in sim on thousands of objects and conditioned on 3D point tracks; prompted with tracks from generated videos, it runs real tasks zero-shot.在仿真中用数千种物体训练与任务无关的“任意位姿到任意位姿”灵巧策略，以 3D 点轨迹为条件；用生成视频提取的点轨迹作提示，即可零样本完成真实任务。
Finger-driven handheld gripper for UMI-style data collection plus an 8,434-hour, 1.20M-episode, 119-task dataset; one policy transfers across UR, Franka and ELEY by mounting the gripper. Hardware and data released.手指驱动的手持夹爪，用于 UMI 式数据采集，并发布 8434 小时、120 万条、119 个任务的数据集；只需装上同款夹爪，同一策略即可迁移到 UR、Franka 与 ELEY。开源硬件与数据。
A cheap drop-in finger module for parallel grippers: two soft belts keep normal open/close and add in-hand translation, pitch and roll; used with MPC, a light 10-DoF teleop interface and learned policies.低成本的平行夹爪升级手指模组：双软皮带结构保留常规开合，并增加手内平移、俯仰与翻滚三个自由度；配合 MPC、轻量的 10 自由度遥操作接口和学习策略使用。
RL for mating two parts within one dexterous hand (capping a marker, a syringe, a bottle) with no second arm. Trained in sim, transfers zero-shot with a single camera.用强化学习在单只灵巧手中完成两个零件的装配（如给马克笔盖帽、注射器、瓶子），不用第二只手臂；仿真训练，单相机零样本迁移到真机。
Pretrains a dexterous RL policy on generic object reposing, then post-trains downstream tasks with BC distillation, critic warm-up and conservative updates. Zero-shot sim-to-real on 23-DoF and 29-DoF hand-arm systems.先在通用物体重定位任务上用强化学习预训练灵巧策略，再用行为克隆蒸馏、critic 预热和保守更新后训练下游任务。在 23 与 29 自由度手-臂系统上零样本仿真到真机。
Models a dexterous skill as a latent linear (Koopman) system in which visual flow and robot motion co-evolve; it replans when observed and predicted flow diverge. Matches strong baselines with much faster inference.把灵巧技能建模为潜在线性（Koopman）系统，视觉特征流与机器人动作流共同演化；当观测与预测的视觉流偏离时自动重规划。性能与强基线相当，推理快得多。
Lifts 2D inputs to 3D with per-pixel vertex maps and expresses camera views and actions in one shared bird's-eye-view frame, aligning data across robots, cameras and datasets.用逐像素顶点图把 2D 输入提升到 3D，并把各相机视角与机器人动作统一表示在同一鸟瞰（BEV）坐标系中，对齐不同机器人、相机与数据集的数据。
Multimodal LLM that reads assembly manuals, point clouds and text to predict 6D assembly poses, plus AssemBench with 900K+ samples. Reports state-of-the-art 6D pose reasoning and real multi-step assembly.读取装配说明书、点云与文字指令来预测 6D 装配位姿的多模态大模型，并发布含 90 万余样本的 AssemBench。报告 6D 位姿推理达到最优，并完成真实多步装配。
Slow system plans sparse long-horizon intent, fast system refines dense motion, linked by a latent Schrödinger bridge. Reports 45.8% lower latency than Cosmos-Policy and 70.8% success on six bimanual tasks.慢系统生成稀疏的长时程意图，快系统细化高频动作，二者以潜空间薛定谔桥耦合。报告延迟比 Cosmos-Policy 低 45.8%，六个双臂灵巧任务成功率 70.8%。
Human-in-the-loop RL whose latent world model also predicts when the operator would take over, steering the policy away from risky states. On a 16-DoF LEAP Hand: 96.7% grasp success, up to 84% less intervention.人在回路强化学习：潜在世界模型额外预测操作者何时会接管，并据此让策略远离高风险状态。在 16 自由度 LEAP 手上抓取成功率 96.7%，人工干预最多减少 84%。
For catching and intercepting moving objects, forecasts the object's future SE(3) trajectory from short pose histories (15 ms) and picks a reachable contact state; reports real success rising from 15.3% to 67.3%.面向抓取运动物体的主动式动态抓取：由短时位姿历史预测物体未来 SE(3) 轨迹（15 毫秒），再选择可达的接触状态；报告真实环境成功率从 15.3% 提升到 67.3%。
Estimates mass, elasticity and friction visuotactilely during the grasp and applies the minimum safe grip force; adds a dataset of fruits and vegetables with measured physical properties.在抓取过程中用视触觉实时估计物体的质量、弹性和摩擦，从而施加刚好足够的最小夹持力；并提供一个测量了物理属性的水果蔬菜数据集。
Active in-hand 3D reconstruction: estimates per-ray reconstruction uncertainty to pick the next view, then an in-hand rotation policy turns the object to expose it. Better quality and efficiency than non-active rotation.主动式手内三维重建：沿候选视线估计各方向的重建不确定性来选择下一最佳视角，再由手内旋转策略把物体转到该视角。重建质量与效率均优于非主动旋转。
PACER fine-tunes a VLA for electronics disassembly on its own rollouts plus human corrections, reweighting action chunks by progress, source and outcome; 87/100 vs 54/100 for plain SFT on motherboard parts.PACER 用策略自身的执行轨迹和人工纠正数据微调用于电子产品拆解的 VLA，按进度、来源和结果为动作块加权；在主板部件拆解上成功 87/100，普通 SFT 为 54/100。
Tactile-reactive VLA: 100-hour tactile dataset built around motor primitives, a variable-rate mixture-of-transformers and a temporal tactile VQ-VAE. Reports 30%+ higher average success than the best baseline on 12 tasks.触觉反应式 VLA：围绕基本动作采集的 100 小时触觉数据集、可变频率的 Mixture-of-Transformers 架构与时序触觉 VQ-VAE 编码器。报告在 12 个任务上平均成功率比最强基线高 30% 以上。
Touch2Trace feeds a cable through a dexterous hand from touch alone: a self-supervised encoder for a 32x32 piezoresistive sensor plus a small transformer BC policy at 60 Hz; 93% success vs 0% without touch.Touch2Trace 仅凭触觉让灵巧手完成线缆穿引：为 32x32 压阻触觉传感器自监督预训练编码器，配合 60 Hz 的轻量 Transformer 行为克隆策略；成功率 93%，无触觉时为 0%。
SlipSense fuses a 240 Hz pressure array with an 8 kHz accelerometer for slip detection: 96.7% macro F1, 76% of slips caught within 23.1 ms; trained on UMI data, it transfers to a dexterous hand.SlipSense 融合 240 Hz 压力阵列与 8 kHz 加速度计进行滑移检测：宏 F1 达 96.7%，76% 的滑移在 23.1 毫秒内被检测到；仅用 UMI 数据训练即可零样本迁移到灵巧手。
Learns object representations from tactile exploration, aligned with text descriptions of properties, and uses them as context for manipulation. Reports task success from 41% to 92% (seen) and 19% to 67% (unseen).通过触觉探索学习物体表征，并与物理属性的文本描述对比对齐，再作为语境指导操作策略。报告任务成功率在已见物体上从 41% 升到 92%，未见物体从 19% 升到 67%。
Generates realistic tactile images from one reference image plus contact force and pose, to scale tactile datasets across sensors. Augmented data improves insertion, imitation learning and weighing in real tests.以一张参考触觉图像加接触力与接触位姿为条件，生成逼真的触觉图像来扩充跨传感器的触觉数据。在插装、模仿学习与称重等真实任务中，增强数据均带来提升。
Uses a 25 Hz tactile reflex controller as a teacher during data collection so demos carry well-regulated grasp forces; tactile-free ACT then reaches 95% stable grasps on fragile cups. Disturbances still need the reflex.数据采集时用 25 Hz 触觉反射控制器当“老师”，让演示自带良好的抓握力；之后不用触觉的 ACT 策略在易碎杯子上达到 95% 稳定抓取。但抵抗外部扰动仍需保留反射控制器。
Robot-DIFT distills a noisy diffusion feature extractor into a deterministic single-pass backbone with a feature pyramid, so policies see fine correspondence cues for the last millimeters of contact-rich tasks.Robot-DIFT 将依赖噪声的扩散特征提取器蒸馏为确定性的单次前向骨干网络并配合特征金字塔，使策略获得精细的对应关系线索，改善接触密集任务最后几毫米的精度。
VISTA: equivariant visuotactile diffusion policy that fuses touch into spherical visual tokens and rotates them with the gripper pose, for more data-efficient contact-rich imitation.VISTA：等变视触觉扩散策略，把触觉融合进球面视觉 token 并随末端姿态旋转，提升富接触模仿学习的数据效率。
Splits force regulation into a direction part learned in sim (task frame and control-mode outputs of a VLA-based policy) and a magnitude part set by hand in an admittance controller. Sim data only; four real tasks.把力调节拆成两部分：方向分量（任务坐标系与控制模式，由基于 VLA 的策略在仿真数据上学习）和幅值分量（在导纳控制器中人工设定）。只用仿真数据，在四个真机任务上验证。
CompVLA adds a compliance expert to a VLA that predicts time-varying stiffness and virtual displacement alongside motion, executed with geometric impedance control for contact-rich tasks.CompVLA 为 VLA 增加柔顺专家模块，在预测运动的同时输出时变刚度与虚拟位移，并通过几何阻抗控制执行，用于接触密集任务。
DexFLEX treats fingertip motion commands from teleop or a policy as intent: it proposes short motion chunks, predicts their contact outcome and picks the best; real teleop success rises from 29.2% to 78.3%.DexFLEX 把来自遥操作或策略的指尖运动指令视为意图：生成若干短时动作块、预测其接触结果并择优执行；真实遥操作成功率从 29.2% 提升到 78.3%。
Argues policies should condition on both the executed robot state and the leader command while predicting future commands, since their mismatch signals contact; beats either alone with no force sensing.认为策略应同时以机器人实际执行状态和主端指令为输入并预测未来指令，因为两者的偏差蕴含接触信息；在没有力觉输入的低成本平台上优于单独使用任一信号。
Adds force-torque sensing to a pretrained vision-only world-action model with staged fusion, multisensory future prediction and replay of pretraining data; strong on contact tasks without losing original skills.通过分阶段融合、多感官未来预测和预训练数据回放，把力/力矩传感加入预训练的纯视觉世界动作模型；在富接触任务上表现好，同时保住甚至提升原有任务性能。
TacEx points curiosity at touch: it splits model uncertainty by sense so exploration seeks contact, not free-space motion. The collected data trains pick-and-place offline and post-trains VLAs.TacEx 让好奇心聚焦触觉：按感知模态分解模型不确定性，使探索去寻找接触而非空中乱动。采集的数据可离线训练抓放策略，也可用于 VLA 后训练。
An imitation-learning benchmark on the open KlaskTron table game robot: 80k real human demo steps, augmented and synthetic data, and baselines exposing BC and GAIL failure modes; policies run on hardware unchanged.基于开源 KlaskTron 桌面对抗游戏机器人的模仿学习基准：8 万步真实人类示教、对称增强与合成数据，以及揭示 BC 和 GAIL 失效模式的基线；训练出的策略可直接上真机运行。
Low-cost wrist sEMG band whose model predicts per-finger forces, turning video demos into force-labelled ones after a short calibration. Reports >50% lower force error than vision and 87% on force-sensitive tasks.低成本腕戴式肌电（sEMG）设备，经短暂校准即可预测每根手指的力，把视频演示变成带力标注的演示。报告力预测误差比视觉方法低 50% 以上，力敏感任务成功率 87%。
Adds a cheap clutch and 3D-printed spring to each joint of a passive leader arm so redundant joints return to a neutral pose, making demos more consistent; block stacking success 73.3% vs 36.7%.为被动主臂的每个关节加装低成本离合器和 3D 打印螺旋弹簧，使冗余关节回到中立姿态，让示教数据更一致；堆积木成功率 73.3%，未加时为 36.7%。
Uni-Mo generates quadruped motions with an LLM plus a video diffusion model and lifts them to 3D references for tracking policies; the Quad-Imaginarium dataset has 7,488 motions, 96.7% of 392 sampled ran on a Go2.Uni-Mo 用大语言模型和视频扩散模型生成四足动作，并提升为 3D 参考轨迹训练跟踪策略；Quad-Imaginarium 数据集含 7488 个动作，抽样的 392 个在 Go2 真机上 96.7% 成功执行。
MorFiC finds that a shared critic averages incompatible values across robot bodies and fixes it with multiplicative morphology conditioning; trained on one robot, it transfers zero-shot to seven quadrupeds.MorFiC 指出共享价值网络会把不同机身的价值目标混为一谈，并用乘性形态条件化修正该问题；只在一台机器人上训练即可零样本迁移到七种四足机器人。
QWM conditions one latent world model on scale-invariant morphology features and trains policies in imagination; unlike a model-free policy, it transfers zero-shot to unseen quadruped morphologies.QWM 以尺度无关的形态特征作为条件训练单一潜空间世界模型，并在想象中训练策略；与无模型策略不同，它能零样本迁移到未见过的四足形态。
Splits quadruped training into rewards for the task, constraints for limits and energy minimization for gait, with no gait priors; reports 56% lower cost of transport and 96% fewer limit violations, on a Go2.将四足训练拆分为：奖励定义任务、约束保证运行限制、能耗最小化决定步态，且不使用任何步态先验；报告运输代价降低 56%、越限减少 96%，并在 Go2 上部署。
HumoSlope trains humanoid walking on steep slopes with a slope-aware ZMP balance prior and a gait adapter that avoids crouching; the blind policy walks outdoor grass slopes up to 32.1 degrees.HumoSlope 用基于斜坡支撑面的 ZMP 平衡先验和避免蹲伏步态的适配器训练人形机器人陡坡行走；纯本体感知策略可在户外草坡上行走，坡度最高 32.1 度。
RPL distills terrain-specific experts into one multi-depth-camera transformer for multi-directional humanoid walking with payloads, with a ray-casting depth renderer about 5x faster than existing pipelines.RPL 将各地形专家策略蒸馏为一个多深度相机 Transformer 策略，实现可负重的人形多方向行走，并配套比现有管线快约 5 倍的光线投射深度渲染器。
Humanoid parkour over sparse footholds from onboard depth: a saliency-guided gated memory keeps useful depth features across frames, and an alternation loss enforces left-right foot placement.基于机载深度相机的人形跑酷，跨越稀疏落脚点：显著性引导的门控记忆跨帧保留有用的深度特征，交替损失约束左右脚交替落步。
55 scene-aligned 4D human motion sequences (150k+ frames) on rough terrain from egocentric wearables and a portable scanner, with an open pipeline; used to train perceptive locomotion run on a Unitree G1.利用第一人称可穿戴设备与便携式扫描仪采集 55 段与场景对齐的复杂地形 4D 人体运动（超过 15 万帧），并开源重建流程；用于训练感知式运动策略，部署到 Unitree G1。
MeshMimic reconstructs both human motion and terrain geometry from monocular video, then retargets with contact constraints, so humanoids learn motion-terrain interactions without MoCap.MeshMimic 从单目视频中同时重建人体运动与地形几何，再以接触约束进行重定向，使人形机器人无需动捕即可学习动作与地形的交互。
Perceptive BFM grounds human motion references in the robot's local terrain: synthesized terrain-conformal references train a teacher, and a student tracker adds terrain corrections only when needed.Perceptive BFM 让人类动作参考适配机器人所处的局部地形：用合成的贴合地形参考训练教师策略，学生跟踪器仅在必要时加入地形修正。
MARCH guides RL with a safe reference from simplified models and a control-Lyapunov reward, then distills to a vision policy; a Unitree G1 crosses sparse footholds with lateral constraints.MARCH 用简化模型生成的安全参考轨迹和控制李雅普诺夫奖励引导强化学习，再蒸馏为视觉策略；Unitree G1 可通过带侧向约束的稀疏落脚点。
Perceptive humanoid locomotion that keeps sharp terrain edges and assigns credit over time during distillation. Reports 96% stepping-stone success and a 1.5 km outdoor walk with one depth camera.感知式人形行走：保留地形的清晰边缘，并在蒸馏时做跨时间的信用分配。报告踏石通过率 96%，仅用一个深度相机完成 1.5 公里户外路线。
CAP keeps one humanoid policy working as depth degrades: a world-model encoder learns to denoise corrupted depth, alongside a proprioceptive encoder; tested on a G1 indoors and outdoors.CAP 让单一人形策略在深度感知退化时仍能工作：世界模型编码器学习对受损深度去噪，并配合本体感知编码器；在 G1 上进行了室内外测试。
RoboDreamer distills a humanoid teacher into a Mamba student that sees randomly masked recent observations, so it infers missing state from history; tested in IsaacLab, MuJoCo and on a G1.RoboDreamer 将人形教师策略蒸馏为 Mamba 学生策略，训练时随机遮蔽最近观测，使其从历史推断缺失状态；在 IsaacLab、MuJoCo 和 G1 上验证。
HAF-VLA adapts a flow-matching VLA to humanoid loco-manipulation by denoising in stages, locomotion and head first, then waist, then full body, linked by a cross-stage KV cache; six real tasks.HAF-VLA 将流匹配 VLA 适配到人形移动操作：分阶段去噪，先生成移动与头部指令，再细化腰部，最后输出全身操作动作，各阶段通过跨阶段 KV 缓存衔接；在六个真实任务上验证。
Real-time text-driven humanoid control: an autoregressive motion diffusion model streams short trajectories from the current command, which can change mid-motion, and a tracking policy executes them on hardware.实时文本驱动的人形机器人控制：自回归运动扩散模型根据当前（可随时修改的）文字指令持续生成短时轨迹，再由运动跟踪策略在真机上执行，可流畅切换舞蹈、跳跃等动作。
End-to-end text-conditioned humanoid control: a diffusion action prior learned from text-labelled state-action data, plus a light residual RL policy for robustness and smooth switching between commands.端到端的文本条件人形全身控制：从带文本标签的状态-动作轨迹学习扩散动作先验，再叠加轻量残差强化学习策略提升长时鲁棒性和指令切换的平滑度。
Samples many motions from a text-to-motion generator and picks one that a learned verifier says the humanoid can track and that matches the text. Improves tracking and text alignment, incl. on a real G1.从文本生成运动模型采样多个候选，用学习到的动态可行性验证器与语义对齐验证器挑选人形机器人既能跟踪又符合文字的动作。仿真与真实 G1 上跟踪精度与文本一致性均提升。
Fine-tunes humanoid motion-tracking policies online on the real robot: a proprioceptive neural dynamics model learned from real data supplies gradients, no motion capture needed. Shown on Unitree G1.在真机上在线微调人形机器人动作跟踪策略：用真实数据训练的本体感知神经动力学模型提供可反传梯度，无需动捕系统。在 Unitree G1 上验证对多种扰动的快速适应。
Studies adapting closed-weight Gemini Robotics On-Device on a humanoid via a managed fine-tuning API; CLIFT turns deployment reward feedback into SFT data, reaching near-perfect success after two cycles.研究通过托管微调 API 在人形机器人上适配闭源权重的 Gemini Robotics On-Device；CLIFT 把部署时的奖励反馈转为监督微调数据，两轮迭代后接近完美成功率。
Planner plus inverse-dynamics student distilled from a privileged policy; at deployment only the IDM is fine-tuned on about 2 minutes of target-domain rollouts. Beats in-context and end-to-end adaptation baselines.从特权策略蒸馏出“规划器+逆动力学模型”学生策略；部署时冻结规划器，只用约 2 分钟目标域数据微调逆动力学模型。优于上下文自适应与端到端适应基线。
ADAPT feeds a humanoid policy an analytical whole-body disturbance observer that estimates external force/torque from robot dynamics without F/T sensors; more robust to pushes and payloads on a G1.ADAPT 为人形策略提供解析式全身扰动观测器，无需力/力矩传感器即可由机器人动力学在线估计外力与力矩；在 G1 上对推搡和负载更鲁棒。
Walk Like This: give a style clip and a joystick velocity; a style-conditioned autoregressive diffusion model generates the reference and an RL controller tracks it, in simulation and on a robot.Walk Like This：给定风格参考片段和摇杆速度指令，风格条件自回归扩散模型生成运动参考，再由强化学习控制器跟踪，在仿真角色和真实机器人上均可运行。
Recasts divergent components of motion as unstable Koopman eigenfunctions learned from about an hour of real biped data; they improve walking tracking and give viability constraints for MPC.把发散运动分量（DCM）重新表述为不稳定的 Koopman 特征函数，仅用约一小时真实双足数据学习得到；可改善行走跟踪，并为 MPC 提供可行性约束。
Starts RL with stiff joints and anneals stiffness to the identified value as episodes lengthen, which enlarges the viable set; enables a quadruped-to-handstand transition on Spot that otherwise never succeeds.强化学习训练初期使用较高的关节刚度，随回合变长逐步退火到辨识值，从而扩大可行状态集；使原本无法学会的 Spot 由四足站立转为倒立的动作得以完成并上真机。
TACL builds curricula from trajectory tasks on unstructured terrain maps: an evaluator learns difficulty for the current policy and a sampler proposes the next tasks; +18.5% on the hardest terrain vs handcrafted.TACL 直接在非结构化地形图上以轨迹任务构建课程：评估器学习当前策略的任务难度，采样器据此提出下一批任务；在最难地形上比人工课程成功率高 18.5%。
SHOES adapts a frozen sim-trained quadruped policy to footwear by fine-tuning only small identity-initialized modulation modules on the robot, with a workspace-bounded sampler; Go2 tested with shoes.SHOES 冻结仿真训练的四足策略，仅在真机上微调以恒等初始化的轻量调制模块来适应穿鞋带来的形态变化，并用限定工作空间的采样器保证安全；在穿鞋的 Go2 上测试。
Studies locomotion policies via effective rank of the network Jacobian conditioned on gait phase; normalization and residual connections help, giving up to 2x lower joint jitter on a real Spot.以步态相位为条件分析运动策略网络雅可比的有效秩；发现归一化与残差连接更健康，在真实 Spot 上关节抖动最多降低 2 倍。
A full-scale humanoid sprints above 7 m/s and corners with up to 25 degrees of trunk lean, using causal attention over feedback, a heading-direction reward and skill-conditioned motion-prior discriminators.全尺寸人形机器人以超过 7 m/s 冲刺，并以最高 25 度躯干侧倾高速过弯；方法包括对反馈的因果注意力、航向奖励以及按技能条件化的运动先验判别器。
Dynamic routing makes policy-update divergence misestimated in mixture-of-experts policies, so MAPO optimizes router and experts in turn. Reports tracking 182 motions PPO with MoE fails on, against one the other way.动态路由会让混合专家策略的更新散度估计失准，MAPO 因此交替优化路由与专家。报告可跟踪 182 个 PPO+MoE 失败的动作，反过来只有 1 个。
CHIP is a plug-and-play module adding controllable end-effector stiffness to agile tracking policies, for forceful tasks such as wiping and pushing carts.CHIP 是即插即用模块，为敏捷跟踪策略加入可控的末端刚度，用于擦拭、推车等需要施力的任务。
λ-Reachability learns Hamilton-Jacobi safety values with a TD(λ)-like multi-step estimator with a random horizon, proven to be a contraction; better safe-set estimates on simulated and real humanoids.λ-Reachability 用类似 TD(λ)、随机视野的多步估计器学习 Hamilton-Jacobi 安全值函数，并证明其为压缩映射；在仿真与真实人形上安全集估计更准确。
MPC-Injection steers off-policy RL toward a usable gait by inserting MPC-generated transitions into training, with no reward shaping or discriminator; Go2 sim-to-real, and a barrel roll pure RL misses.MPC-Injection 把 MPC 生成的状态转移注入离策略强化学习训练，在不改奖励、不加判别器的情况下把策略引向可用步态；在 Go2 上实现仿真到真实，并完成纯 RL 学不会的侧滚翻。
CoRL 2026 Spotlight. A humanoid system for climbing diverse ladders and manipulating while on them; a two-stage pipeline learns climbing experts through hybrid motion tracking.CoRL 2026 Spotlight。让人形机器人攀爬多种梯子并在梯上操作；两阶段流程用混合动作跟踪学习多个攀爬专家策略。
CoRL 2026 Spotlight. Humanoid locomotion dataset from internet video that targets the floating, penetration and foot skating common in earlier video-derived data.CoRL 2026 Spotlight。源自网络视频的人形运动数据集，针对此前视频数据常见的悬空、穿透与脚滑问题。
CoRL 2026 Spotlight. Brings sole pressure sensing into humanoid locomotion control, aiming for softer foot-terrain contact than parkour policies that succeed despite harsh landings.CoRL 2026 Spotlight。把足底压力感知引入人形运动控制，目标是比“能完成但落地粗暴”的跑酷策略更柔和的足地接触。
Separates performance from safety on humanoids: the nominal policy handles the task while a residual RL policy acts as an implicit safety filter, avoiding competing reward terms.在人形机器人上分离性能与安全：名义策略负责任务，残差 RL 策略充当隐式安全过滤器，避免相互竞争的奖励项。
CoRL 2026 Spotlight. A one-variable-at-a-time empirical study of what a whole-body native VLA needs, mapping language and pixels directly to all humanoid degrees of freedom.CoRL 2026 Spotlight。逐变量对照的实证研究：探索直接把语言与图像映射到人形全部自由度的全身原生 VLA 需要哪些条件。
CoRL 2026 Spotlight. Language-conditioned humanoid traversal of cluttered rooms that treats navigation as whole-body adaptation of arms, torso and gait rather than 2D path planning.CoRL 2026 Spotlight。语言条件下的人形机器人杂乱环境穿行：不把导航当二维路径规划，而是手臂、躯干与步态的全身协同调整。
Uses sample-based MPC in simulation as an automatic expert, then trains off-policy RL on sparse rewards, avoiding reward shaping; deployed on an arm-equipped Spot and a G1 humanoid.在仿真中用采样 MPC 自动生成专家数据，再以稀疏奖励训练离策略 RL，免去奖励塑形；已部署到带臂 Spot 和 G1 人形机器人。
CoRL 2026 Spotlight. Hardens whole-body end-effector tracking for legged manipulators against out-of-distribution inputs such as sensor noise or infeasible commands, using a frame-wise single-step safety manifold.CoRL 2026 Spotlight。提升足式移动操作臂全身末端跟踪对分布外输入（传感器噪声、不可行指令）的鲁棒性：用逐帧安全方案把无限时域安全约束转化为计算高效的单步流形约束。
CoRL 2026 Spotlight. Benchmark for visual navigation on urban sidewalks, built on Isaac Sim with procedurally generated and real-world environments.CoRL 2026 Spotlight。城市人行道视觉导航基准，基于 Isaac Sim，包含程序生成与真实世界环境。
CoRL 2026 Spotlight. Decides whether to wait or reroute at a blocked route by modeling how long each obstacle class tends to persist.CoRL 2026 Spotlight。通过建模各类障碍物通常停留多久，决定在路线受阻时等待还是改道。
Argues VLN success metrics hide a 'last-3-meter grounding gap': agents stop nearby without facing the target. Adds instance-level metrics and REALM, a refinement module for any backbone.指出 VLN 成功率掩盖了“最后 3 米定位缺口”：智能体停在附近却未对准目标。提出实例级指标和可接入任意主干的修正模块 REALM。
Low-level VLN action head that predicts a language-conditioned flow field over the visible sector and rolls it out into trajectories. Adds a continuous-time Habitat benchmark.低层视觉语言导航动作头：在可见扇区上预测语言条件的流场并展开为连续轨迹；并提供连续时间的 Habitat 评测基准。
VIGOR-MPC distills a privileged long-horizon MPC into a short-horizon MPC whose cost weights are inferred online from noisy vision; in vision-based racing it nears the teacher's lap times with 90%+ less compute per step.VIGOR-MPC 将具有特权信息的长视野 MPC 蒸馏为短视野 MPC，其代价参数由含噪视觉在线推断；在视觉自主竞速中接近教师圈速，每步计算量减少 90% 以上。
Neural inertial odometry across carriers: a router composes carrier-specific experts, and a light residual branch adapts to new ones. Reports 40% lower ATE than learned baselines.面向多种载体的神经惯性里程计：路由器组合各载体专家，新载体通过轻量残差分支适配。报告 ATE 比学习型基线降低 40%。
CoRL 2026 Spotlight. Current-centric streaming 3D reconstruction that predicts window-relative poses and local geometry in the current frame, to limit drift on long runs.CoRL 2026 Spotlight。以当前帧为中心的流式三维重建：在当前帧坐标系预测窗口相对位姿与局部几何，以抑制长时漂移。
MEND plans UAV-UGV recharge teamwork with masked diffusion, enforcing operational rules at execution through a logic-based action mask; tested on real maps and physical robot teams.MEND 用掩码扩散规划无人机与地面车协同充电任务，在执行时通过逻辑动作掩码强制满足运行约束；在真实地图和实体机器人队伍上测试。
From the来自 official workshop page官方研讨会页面. Rooms and times are on each workshop site.地点与时间见各研讨会网站。