RECAP trains a VLA with RL via advantage conditioning on demos, on-robot rollouts and teleop corrections; reports it more than doubles throughput on hard tasks like laundry folding and espresso making.RECAP 通过优势条件化，用演示、机器人自主运行数据和遥操作纠正对 VLA 做强化学习；论文报告在叠衣服、做意式咖啡等难任务上吞吐量提升一倍以上、失败率约减半。
@misc{intelligence2025vla,
title = {{π*0.6: a VLA That Learns From Experience}},
author = {Physical Intelligence and Ali Amin and Raichelle Aniceto and Ashwin Balakrishna and Kevin Black and Ken Conley and Grace Connors and James Darpinian and Karan Dhabalia and Jared DiCarlo and Danny Driess and Michael Equi and Adnan Esmail and Yunhao Fang and Chelsea Finn and Catherine Glossop and Thomas Godden and Ivan Goryachev and Lachy Groom and Hunter Hancock and Karol Hausman and Gashon Hussein and Brian Ichter and Szymon Jakubczak and Rowan Jen and Tim Jones and Ben Katz and Liyiming Ke and Chandra Kuchi and Marinda Lamb and Devin LeBlanc and Sergey Levine and Adrian Li-Bell and Yao Lu and Vishnu Mano and Mohith Mothukuri and Suraj Nair and Karl Pertsch and Allen Z. Ren and Charvi Sharma and Lucy Xiaoyang Shi and Laura Smith and Jost Tobias Springenberg and Kyle Stachowicz and Will Stoeckle and Alex Swerdlow and James Tanner and Marcel Torne and Quan Vuong and Anna Walling and others},
year = {2025},
eprint = {2511.14759},
archivePrefix = {arXiv},
primaryClass = {cs.LG},
url = {https://arxiv.org/abs/2511.14759},
}
A 7B open VLA built on Llama 2 + DINOv2/SigLIP; the de-facto open baseline for VLA research and LoRA fine-tuning.基于 Llama 2 与 DINOv2/SigLIP 的 7B 开源 VLA，是 VLA 研究与 LoRA 微调的事实开源基线。
@misc{kim2024openvla,
title = {{OpenVLA: An Open-Source Vision-Language-Action Model}},
author = {Moo Jin Kim and Karl Pertsch and Siddharth Karamcheti and Ted Xiao and Ashwin Balakrishna and Suraj Nair and Rafael Rafailov and Ethan Foster and Grace Lam and Pannag Sanketi and Quan Vuong and Thomas Kollar and Benjamin Burchfiel and Russ Tedrake and Dorsa Sadigh and Sergey Levine and Percy Liang and Chelsea Finn},
year = {2024},
eprint = {2406.09246},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2406.09246},
}
A VLM backbone with a flow-matching action expert, trained across many robots, doing long dexterous tasks like laundry folding.VLM 主干加流匹配动作专家，跨多种机器人训练，能完成叠衣服等长时程灵巧任务。
@misc{black2024vision,
title = {{$π_0$: A Vision-Language-Action Flow Model for General Robot Control}},
author = {Kevin Black and Noah Brown and Danny Driess and Adnan Esmail and Michael Equi and Chelsea Finn and Niccolo Fusai and Lachy Groom and Karol Hausman and Brian Ichter and Szymon Jakubczak and Tim Jones and Liyiming Ke and Sergey Levine and Adrian Li-Bell and Mohith Mothukuri and Suraj Nair and Karl Pertsch and Lucy Xiaoyang Shi and James Tanner and Quan Vuong and Anna Walling and Haohuan Wang and Ury Zhilinsky},
year = {2024},
eprint = {2410.24164},
archivePrefix = {arXiv},
primaryClass = {cs.LG},
url = {https://arxiv.org/abs/2410.24164},
}
@misc{brohan2023rt,
title = {{RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control}},
author = {Anthony Brohan and Noah Brown and Justice Carbajal and Yevgen Chebotar and Xi Chen and Krzysztof Choromanski and Tianli Ding and Danny Driess and Avinava Dubey and Chelsea Finn and Pete Florence and Chuyuan Fu and Montse Gonzalez Arenas and Keerthana Gopalakrishnan and Kehang Han and Karol Hausman and Alexander Herzog and Jasmine Hsu and Brian Ichter and Alex Irpan and Nikhil Joshi and Ryan Julian and Dmitry Kalashnikov and Yuheng Kuang and Isabel Leal and Lisa Lee and Tsang-Wei Edward Lee and Sergey Levine and Yao Lu and Henryk Michalewski and Igor Mordatch and Karl Pertsch and Kanishka Rao and Krista Reymann and Michael Ryoo and Grecia Salazar and Pannag Sanketi and Pierre Sermanet and Jaspiar Singh and Anikait Singh and Radu Soricut and Huong Tran and Vincent Vanhoucke and Quan Vuong and Ayzaan Wahid and Stefan Welker and Paul Wohlhart and Jialin Wu and Fei Xia and Ted Xiao and others},
year = {2023},
eprint = {2307.15818},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2307.15818},
}
1M+ trajectories from 22 robot types pooled by 30+ labs; RT-X models trained on it beat single-robot baselines.30 多个实验室汇集 22 种机器人的 100 多万条轨迹；在其上训练的 RT-X 优于单机器人基线。
@misc{collaboration2023open,
title = {{Open X-Embodiment: Robotic Learning Datasets and RT-X Models}},
author = {Open X-Embodiment Collaboration and Abby O'Neill and Abdul Rehman and Abhinav Gupta and Abhiram Maddukuri and Abhishek Gupta and Abhishek Padalkar and Abraham Lee and Acorn Pooley and Agrim Gupta and Ajay Mandlekar and Ajinkya Jain and Albert Tung and Alex Bewley and Alex Herzog and Alex Irpan and Alexander Khazatsky and Anant Rai and Anchit Gupta and Andrew Wang and Andrey Kolobov and Anikait Singh and Animesh Garg and Aniruddha Kembhavi and Annie Xie and Anthony Brohan and Antonin Raffin and Archit Sharma and Arefeh Yavary and Arhan Jain and Ashwin Balakrishna and Ayzaan Wahid and Ben Burgess-Limerick and Beomjoon Kim and Bernhard Schölkopf and Blake Wulfe and Brian Ichter and Cewu Lu and Charles Xu and Charlotte Le and Chelsea Finn and Chen Wang and Chenfeng Xu and Cheng Chi and Chenguang Huang and Christine Chan and Christopher Agia and Chuer Pan and Chuyuan Fu and Coline Devin and others},
year = {2023},
eprint = {2310.08864},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2310.08864},
}
A 2.4B VLM + flow-matching VLA pretrained only on public data; reports beating all but ~2× larger models, plus cheap register-token memory.仅用公开数据预训练的 2.4B VLM + 流匹配 VLA；报告称只输给约两倍大的模型，并用少量寄存器 token 实现低成本记忆。
@misc{marrie2026reki,
title = {{REKI: Registering Knowledge in VLAs}},
author = {Juliette Marrie and Paul Pacaud and Adrien Ramanana Rahary and Shizhe Chen and Moritz Böhle},
year = {2026},
eprint = {2610.22178},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.22178},
}
Gives π0.5 memory by caching compact latent transitions between observations, aimed at tasks where key objects go out of view.通过缓存相邻观测之间的紧凑潜在转移，为 π0.5 加入记忆，面向关键物体移出视野的任务。
@misc{sun2026ltm,
title = {{LTM: Latent Transition Based In-Context Memory for Vision-Language-Action Models}},
author = {Qixin Sun and WenLong Gao and Xiansheng Chen and Zhongyuan Wang and Si Liu and Pengwei Wang},
year = {2026},
eprint = {2610.21362},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.21362},
}
Speeds up VLA inference by keeping only the visual tokens the current task needs, with a token budget that changes from step to step.只保留当前任务所需的视觉 token、并让 token 预算随时间步变化，以加快 VLA 推理。
@misc{xu2026fastvla,
title = {{FastVLA: Task-Conditioned and Temporally Adaptive Visual Token Compression for Efficient Vision-Language-Action Models}},
author = {Jiuyi Xu and Chengyuan Qian and Zhengzhong Tu and Yang Sui and Yangming Shi},
year = {2026},
eprint = {2610.22146},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.22146},
}
Learns skills separately, then consolidates them after the fact with a nested hypernetwork that generates VLA parameters instead of merging weights.先独立学习各项技能，再用嵌套超网络“生成”VLA 参数来事后整合技能，而不是直接合并权重。
Argues VLAs overfit observation→action pairs and adds explicit supervision of the physical state change an action should cause, rather than predicting future states in language or pixels.认为 VLA 容易过拟合“观测→动作”的对应关系，因此显式监督动作应引起的物理状态变化，而不是在语言或像素空间里预测未来状态。
@misc{lu2026vla,
title = {{VLA-$Δ$: Learning Physical State Transition for Vision-Language-Action Models}},
author = {Qi Lu and Jiarun Li and Ce Zhang and Miao Pan and Yi Pan and Xu Wang and Wenqi Zhang and Xuhong Zhang},
year = {2026},
eprint = {2610.21306},
archivePrefix = {arXiv},
primaryClass = {cs.CV},
url = {https://arxiv.org/abs/2610.21306},
}
Keeps π0.5's pretrained vision-language encoder but swaps its Transformer action expert — run many times per denoising loop — for a compact Mamba expert to cut inference cost.保留 π0.5 预训练的视觉语言编码器，把在去噪循环中被反复调用的 Transformer 动作专家换成紧凑的 Mamba 专家，以降低推理开销。
Instead of feeding a VLA a longer frame history, learns to condense past frames into a few tokens that keep short-lived cues such as brief contacts and occlusions.不靠简单加长历史帧窗口，而是学习把过去的帧压缩成少量 token，保留短暂接触、遮挡等转瞬即逝的线索。
@misc{hong2026condensevla,
title = {{CondenseVLA: Learnable History Condensation for Efficient Multi-Frame VLA}},
author = {Feiyang Hong and Yujie Wei and Bo Zhao and Xiu Su and Hongxun Yao and Shuo Yang},
year = {2026},
eprint = {2610.25927},
archivePrefix = {arXiv},
primaryClass = {cs.CV},
url = {https://arxiv.org/abs/2610.25927},
}
Post-hoc kernel-variance heads (Gaussian-process variance over training-demo embeddings) estimate one VLA policy's epistemic uncertainty without an ensemble, so the robot knows when to ask for help.在训练完成后为单个 VLA 策略加上轻量的核方差头（基于训练示教嵌入的高斯过程方差）来估计认知不确定性，无需训练和运行策略集成，让机器人知道何时该求助或补充示教。
@misc{keller2026post,
title = {{Post-hoc Epistemic Uncertainty for Vision-Language-Action Policies}},
author = {Johannes Keller and Albrecht M. Bloße and Adrian Lindenmeyer and Thomas Neumuth and Daniel Schneider},
year = {2026},
eprint = {2610.29730},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.29730},
}
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）。
@misc{sujit2026when,
title = {{When One Step Is Enough: An Empirical Study of Euler Integration in Flow-Matching Policies}},
author = {Shivakanth Sujit and Luca Nunziante and Rousslan Fernand Julien Dossa and Kai Arulkumaran},
year = {2026},
eprint = {2610.30735},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.30735},
}
Reuses the unexecuted tail of a flow-matching VLA's action chunk to start the next one, with a small confidence model setting how much to re-denoise; reports lower inference cost on π0.5, SmolVLA and others.流匹配 VLA 每次只执行动作块的前几步，剩余部分通常丢弃；本文把这段"动作尾巴"作为下一块的起点，并用小型置信度模型决定还需去噪多少步；论文报告在 π0.5、SmolVLA 等模型上降低推理开销。
@misc{he2026vla,
title = {{VLA-Rolling: Fewer Denoising Steps, Better Control via Adaptive Action Reuse}},
author = {Sicheng He and Yi Pan and Qiuting Yu and Jiaming Huang and Qi Lu and Shuyan Wu and Hongyan Feng and Xuanyu Liu and Jun Zhang and Jie Zhang and Wenqi Zhang and Xuhong Zhang and Jianwei Yin},
year = {2026},
eprint = {2610.25043},
archivePrefix = {arXiv},
primaryClass = {cs.AI},
url = {https://arxiv.org/abs/2610.25043},
}
Drives robots with a frozen frontier LLM plus in-context 'knowledge packages' on 12 sim benchmarks + 8 real robots; reports beating prior bests on 9 of 12, but weak real-robot results (competence, not yet intelligence).用冻结的前沿大模型加上下文“知识包”直接控制机器人，覆盖 12 个仿真基准和 8 台真机；报告在 12 个基准中的 9 个超过此前最好结果，但真机表现较弱——结论是具备了具身“能力”，尚未达到具身“智能”。
@misc{sensenova2026has,
title = {{Has GPT-6-Astra Achieved Embodied Intelligence?}},
author = {SenseNova and Joint Research Team and Zhongang Cai and Ziang Cao and Tailai Chen and Han Chong Chua and Linjun Dai and Hanming Deng and Hokin Deng and Haiwen Diao and Chenjing Ding and Weichen Fan and Xiao Fu and Zerui Gong and Chenyang Gu and Zexin He and Zhi Hou and Bo Li and Qinyue Li and Quanzhou Li and Xiaowan Li and Zhichao Liao and Dahua Lin and Zilong Lin and Bo Liu and Chunxiao Liu and Pengfei Liu and Tianqi Liu and Ziwei Liu and Chen Change Loy and Lewei Lu and Jiapeng Luo and Yuexiao Ma and Liang Pan and Xingang Pan and Oscar Qian and Ziyi Shang and Yuyao Shao and Dawei Su and Dacheng Tao and Qingyi Tao and Quan Wang and Ruisi Wang and Xiaogang Wang and Ziyi Wang and Beichen Wen and Donghao Wu and Jiaxu Wu and Silei Wu and Xiaoming Wu and others},
year = {2026},
eprint = {2610.22471},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.22471},
}
Adds a 257M latent world model to a pretrained VLA so predicted future features condition the action expert, with no pixel generation or video pretraining; reports 98.7% on LIBERO and real-robot gains over π0.5.在预训练 VLA 上加一个 2.57 亿参数的潜空间世界模型，用预测的未来表征去条件化动作专家，无需像素生成或大规模视频预训练；报告 LIBERO 成功率 98.7%，真机任务上明显优于 π0.5。
@misc{wang2026wave,
title = {{WAVE-0: Bridging VLA Policies and WAMs with Lightweight Latent Future Prediction}},
author = {Haomin Wang and Anqiong Ji and Ke Wang and Zhongshan Liu and Ru Ying and Ran Zheng and Zhipeng Zhang},
year = {2026},
eprint = {2610.18568},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.18568},
}
Instead of one fixed set of VLA weights, generates context-dependent low-rank updates for both the vision-language and action modules; reports gains up to 16.3% on LIBERO/Meta-World and up to 14% over π0.5 on real tasks.不再让 VLA 在所有交互情境下共用固定参数，而是按当前情境为视觉语言模块和动作模块生成耦合的低秩参数更新；报告在 LIBERO/Meta-World 上最高提升 16.3%，真机上相对 π0.5 最高提升 14%。
@misc{liu2026morphact,
title = {{MorphAct: Robot Policies Morphing with the Worlds They Act In}},
author = {Ziyan Liu and Jiting Liu and Weihong Zhong and Wenda Yu and Yeqiu Chen and Jason Li and Xiaojiang Peng and Shuyuan Hu and Kai Wang and Bo Zhao},
year = {2026},
eprint = {2610.20019},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.20019},
}
Wraps a frozen VLA in two loops — one keeps user preferences and scene memory for planning, one blocks hazardous actions and calls recovery skills; reports violations down from 52.6% to 6.1%, recovery up from 13% to 80%.给冻结的 VLA 套上双环框架：规划环保存用户偏好与场景记忆以做个性化任务分解，执行环拦截危险接触并调用可复用的恢复技能；报告安全违规率从 52.6% 降到 6.1%，失败恢复成功率从 13% 提升到 80%。
@misc{wang2026carryon,
title = {{CarryOn: A Dual-Loop Harness for Personalized and Reliable Robot Assistance}},
author = {Zixuan Wang and Yuqi Liu and Yuxin Chen and Sengqiao Yang and Jinhui Ye and Pengguang Chen and Changsheng Lu and Shu Liu and Bei Yu and Jiaya Jia},
year = {2026},
eprint = {2610.21453},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.21453},
}
Survey of 'agentic robots' — foundation-model agents that plan, collect data, build environments and recover from errors — organised into 17 categories under model, data, environment and harness; companion paper list.关于"智能体化机器人"的综述：基础模型智能体负责规划、采集与筛选数据、构建环境与测试并自主纠错；把相关研究归为模型、数据、环境、执行框架四大领域下的 17 个类别，附配套论文列表。
@misc{qin2026towards,
title = {{Towards Agentic Robots: Model, Data, Environment, and Harness}},
author = {Chonghan Qin and Dongming Wu and Kaituo Feng and Jiaxu Wang and Xinlei Yu and Peiwen Sun and Yijia Fan and Xilin He and Bingao Chen and Weitong Lian and Ryan Ji and Wenqi Ren and Xiaobin Hu and Dahua Lin and Yu-Gang Jiang and Xiangyu Yue},
year = {2026},
eprint = {2610.20442},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.20442},
}
Survey framing computer-use and robot agents with five functions — perceive, anticipate, plan, act, verify; across 64 benchmarks it finds only 5 test anticipation and 3 test verification. Companion paper list.用"感知、预判、规划、执行、验证"五种能力统一看待电脑操作智能体和机器人智能体的综述；统计 64 个基准后发现只有 5 个评测预判、3 个评测验证。附配套论文列表。
@misc{chen2026multimodal,
title = {{Multimodal Embodied Agents: A Unified Capability-centric Perspective from Computer-Use to Robot-Use}},
author = {Yanzhe Chen and Ziyi Yang and Jifeng Zhu and Qiming Huang and Ruihe An and Peiyao Xu and Hesen Yang and Runda Liu and Chang Gong and Zhijun Cao and Zechen Bai and Wenzheng Zeng and Yiqi Lin and Guoqiang Liang and Kevin Yuchen Ma and Kevin Qinghong Lin and Mike Zheng Shou},
year = {2026},
eprint = {2610.30108},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.30108},
}
Open spatiotemporal embodied foundation models (2B, 8B, 30B-A3B MoE) for egocentric understanding, localization, grounded reasoning and planning, plus post-trained Nav, Plan, VLA and CoP variants.开源时空具身基础模型（2B、8B 与 30B-A3B MoE），覆盖第一视角理解、时空定位、物理落地推理与规划，并提供导航、规划、VLA、CoP 等后训练版本。
@misc{dang2026rynnbrain,
title = {{RynnBrain: Open Embodied Foundation Models}},
author = {Ronghao Dang and Jiayan Guo and Bohan Hou and Sicong Leng and Kehan Li and Xin Li and Jiangpin Liu and Yunxuan Mao and Zhikai Wang and Yuqian Yuan and Minghao Zhu and Xiao Lin and Yang Bai and Qian Jiang and Yaxi Zhao and Minghua Zeng and Junlong Gao and Yuming Jiang and Jun Cen and Siteng Huang and Liuyi Wang and Wenqiao Zhang and Chengju Liu and Jianfei Yang and Shijian Lu and Deli Zhao},
year = {2026},
eprint = {2602.14979},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2602.14979},
}
Practical warning: fine-tuning VLA weights in bf16 without an fp32 master copy rounds away AdamW updates below 1/512 of a weight; finds this recipe in 9 of 12 open VLA codebases and 79%+ of updates lost on LIBERO.实用警示：VLA 权重以 bf16 微调且不保留 fp32 主副本时，小于权重 1/512 的 AdamW 更新会被舍入抹掉；报告 12 个开源 VLA 代码库中 9 个的默认或文档配方如此训练，LIBERO 实验中至少 79% 的更新被抹除。
Six frontier general-purpose models drive a LIBERO-10 arm through tool calls (end-effector poses, gripper), no robot training; best reported success is 44%. More reasoning effort usually helps but costs more.让六个前沿通用大模型通过工具调用（指定末端位姿与夹爪开合）直接控制 LIBERO-10 机械臂，无任何机器人专门训练；报告的最高成功率为 44%。提高推理强度通常有帮助，但推理成本更高。
Harness where a frontier LLM reviews a frozen π0.5 policy less often as experience builds up and focuses on recovery after errors; reports ~57% fewer reviewer tokens and higher success on four sim tasks.让前沿大模型监督冻结的 π0.5 策略的执行框架：随着经验积累逐步减少审查频率，并把监督集中在出错后的恢复上；报告在四个仿真任务上审查 token 平均减少约 57%，成功率也更高。
@misc{lin2026fastharness,
title = {{FastHarness: Experience-Driven Review Scheduling for Vision-Language-Action Policies}},
author = {Tao Lin and Yuxin Du and Haoxiao Wang and Zile Zhong and Bo Zhao},
year = {2026},
eprint = {2610.31427},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.31427},
}
Between direct action prediction and full video prediction: VLDA learns a compact latent 'Delta' of task-relevant state change and uses it to seed action denoising. Built on Baidu's open LoongForge.介于直接预测动作与完整预测未来画面之间：VLDA 学习一个紧凑的隐变量 Delta 表示与任务相关的状态变化，并用它引导动作去噪；实现于百度开源训练框架 LoongForge。
@misc{zhang2026vision,
title = {{Vision-Language-Delta-Action: Learning Compact Deltas as Actionable Transition Interfaces for Robotic Manipulation}},
author = {Lingsen Zhang and Wei Li and Jie He and Henghua Zhang and Kai Mo and Shiyong Li and Yanpeng Wang and Rui Shao},
year = {2026},
eprint = {2610.25874},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.25874},
}
Deployment knob for flow-matching VLAs: fewer solver steps hurt under fast replanning but help with long open-loop chunks. Tuning both together, training-free, reports higher success and lower latency for π0/π0.5/GR00T.流匹配 VLA 的部署调参发现：求解步数少在频繁重规划时会降低精度，但在长时开环执行动作块时反而有利。报告在 π0、π0.5、GR00T 上联合调整两者，无需训练即可提高成功率并降低推理延迟。
@misc{yang2026coupling,
title = {{Coupling Execution Horizon and Sampling Depth in Flow-Matching VLA Policies}},
author = {Jiabing Yang and Yuan Xu and Yixiang Chen and Qisen Ma and Peiyan Li and Kai Wang and Chenhang Cui and Tao Yu and Zichen Wen and Wen Ye and Bowen Fang and Ziheng He and Xiangnan Wu and Yingda Li and Yan Huang and Liang Wang},
year = {2026},
eprint = {2610.32102},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.32102},
}
Report: GPT-6-Astra drives a robot on EBench via waypoints or by calling a VLA/WAM policy. Direct control leads on short tasks, trails on long-horizon and contact-rich ones; calling a policy cuts token cost.评测报告：在 EBench 上让 GPT-6-Astra 直接输出路径点控制机器人，或通过执行框架调用 VLA/WAM 策略。直接控制在短时任务上领先，但在长时程与接触密集任务上落后；调用策略可显著减少 token 开销。
@misc{yao2026profiling,
title = {{Profiling a Frontier Agent with EBench: From Direct Control to Harness with Policy}},
author = {Yifei Yao and Yude Zou and Guanbo Jiao and Xing Gao and Ning Gao and Yukai Wang and Jiantong Chen and Zihou Zhu and Xinyu Li and Ningyu Yan and Shuai Wang and Jianming Ma and Haoyu Guo and Weinan Zhang and Chunhua Shen},
year = {2026},
eprint = {2610.35832},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.35832},
}
Renders wrist-camera views from a 3D Gaussian Splatting scan of the real workspace to fine-tune π0.5 for precise reaching in e-waste disassembly; reports 80% vs. 58.7% with 150 teleop demos.用真实工作台的 3D 高斯泼溅重建渲染腕部相机训练视角，微调 π0.5 以完成电子垃圾拆解中的精确抵达；报告成功率 80%，而 150 条遥操作示教为 58.7%，人工采集不到 10 分钟。
Mixed-precision quantization for VLAs that sets each weight matrix's bit width by how much it affects actions; reports 92.7% LIBERO success at 3 bits (above AWQ and GPTQ) and 70.5% at 2.5 bits.面向 VLA 的混合精度量化：根据每个权重矩阵对输出动作的影响分配比特数；报告在 3 比特下 LIBERO 成功率 92.7%（高于 AWQ 与 GPTQ），2.5 比特下仍有 70.5%。
@misc{qiu2026agq,
title = {{AGQ: Action-Guided Ultra-Low-Bit Quantization for Vision-Language-Action Models}},
author = {Xuhong Qiu and Zhuoyu Wu and Qiawei Zheng and Wenhui Ou and Wenqi Fang and Raphaël C. -W. Phan},
year = {2026},
eprint = {2610.33074},
archivePrefix = {arXiv},
primaryClass = {cs.CV},
url = {https://arxiv.org/abs/2610.33074},
}
Pretrains a 4B VLA on 3D-grounded latent actions mined from 30K hours of human, robot and sim video, then adapts to real robot actions. Reports 98.8% on LIBERO and 88.6% on RoboTwin 2.0.从 3 万小时人类、机器人与仿真视频中学习带 3D 约束的连续潜在动作（RynnLAM），用于预训练 4B 参数 VLA，再适配到具体机器人动作空间。报告 LIBERO 98.8%、RoboTwin 2.0 平均 88.6%。
@misc{zhao2026rynnvla,
title = {{RynnVLA-Latent: Scaling Robotic Manipulation Foundation Models with 3D-Grounded Latent Actions}},
author = {Haoyu Zhao and Xingyue Zhao and Zhian Su and Bohan Hou and Yinzhou Tang and Fengyi Wang and Kehan Li and Yuming Jiang and Siteng Huang and Xin Li and Zhenyu Zeng and Mingxiu Chen and Deli Zhao and Zhongyu Li},
year = {2026},
eprint = {2610.34356},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.34356},
}
Meta-agent that improves a robot's harness by exploring: keeps a lineage graph of harness versions and runs new rollouts to probe failures before editing. Reports more success with fewer rollouts than static reflection.通过主动探索改进机器人 harness 的元智能体：维护 harness 版本谱系图，先主动生成新回放探查失败与恢复行为再修改；报告在 LIBERO-Pro、RoboCasa、RoboTwin 上比静态反思用更少仿真回放取得更高成功率。
@misc{zhang2026vista,
title = {{VISTA: Active Exploration for Self-Improvement in Embodied Agents}},
author = {Zhang Zhang and Shengjun Zhang and Simin Huang and Min Chen and Chong Xia and Xin Fei and Yueqi Duan},
year = {2026},
eprint = {2610.35051},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.35051},
}
Teaches a VLA to know when it is done or failing: stop and failure signals are two extra action dimensions, denoised faster than the actions so they can guide them; no frame-level progress labels. Real-robot tests.让 VLA 自己判断任务完成或失败：把停止与失败信号作为动作块的两个额外维度，与动作一起去噪且去噪更快，从而引导后续动作生成；无需逐帧进度标注，在真机操作实验中验证。
@misc{liu2026echo,
title = {{Echo: Coupled Action and Task Status Signal Denoising for Vision Language Action Models}},
author = {Yufeng Liu and Bohao Fan and Juntu Zhao and Xinyi Wang and Hang Yu and Wenzhuan Luo and Yingzhi Wang and Qinghai Li and Yingdong Hu and Junliang Guo and Richard Wang and Dequan Wang and Yang Gao and Cheng Ma},
year = {2026},
eprint = {2610.31569},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.31569},
}
Latent action model that fuses synchronized camera views to resolve depth and occlusion. Reports 17% better 6D action decodability than single-view and +13.5% policy success (vs +6.2%); human-hand demos help on hardware.融合多路同步相机视角的潜在动作模型（LAM），用交叉注意力解决深度与遮挡问题；报告 6D 动作可解码性比单视角高 17%，下游策略成功率提升 13.5%（单视角为 6.2%），真机上用人手演示也比纯动作策略高 20%。
@misc{hausdorfer2026multi,
title = {{Multi-View Latent Action Pretraining for Cross-Embodiment Robot Learning from Videos}},
author = {Oliver Hausdörfer and Felix Minzenmay and Mani Majd and Maximilian P. Christof and Shiyuan Zhang and Angela P. Schoellig},
year = {2026},
eprint = {2610.27794},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.27794},
}
Splits the VLA vs WAM debate into backbone, future loss and inference machinery: a foresight loss predicts frozen future features, no video generation. Reports beating FastWAM on LIBERO-Plus (64.5% vs 55.3%).把 VLA 与世界动作模型（WAM）之争拆成骨干初始化、未来预测损失和推理时是否保留预测模块三个选择：用预测冻结未来特征的前瞻损失，训练与推理都不生成视频；报告在 LIBERO-Plus 上 64.5%，FastWAM 为 55.3%。
@misc{yu2026vision,
title = {{Vision-Language-Action or World-Action? Just Action Models}},
author = {Yonghao Yu and Lang Huang and Runyi Li and Toshihiko Yamasaki},
year = {2026},
eprint = {2610.37219},
archivePrefix = {arXiv},
primaryClass = {cs.CV},
url = {https://arxiv.org/abs/2610.37219},
}
Lets a VLA pick, at each step, between fine and temporally coarser action decoders: redundant motion runs faster while precision phases stay fine-grained. Tested on 3 sim benchmarks and 6 real tasks.让 VLA 在每一步自行选择细粒度或时间上更粗的动作解码器：冗余动作阶段加速执行，精细接触阶段保持细粒度控制；在 3 个仿真基准和 6 个真实任务上验证效率提升且成功率不降。
@misc{jeon2026pacevla,
title = {{PaceVLA: Adaptive Temporal Action Quantization in Vision-Language-Action Models}},
author = {Hojin Jeon and Jimin Lee and Taekwan Kim and Nayoung Oh and Taeyoung Kim and Myungkyu Koo and Jinmo Koo and Joonwoo Ahn and Jinwoo Shin},
year = {2026},
eprint = {2610.25961},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.25961},
}
Xiaomi's compact embodied agent that plans, tracks progress and recovers from failures across manipulation and navigation, distilled from stronger teachers with a rubric-driven self-improvement loop.小米具身智能团队的轻量统一具身智能体：在操作与导航中负责任务理解、规划、进度跟踪和失败恢复，由更强的教师模型多教师在线蒸馏而来，并配有基于评分细则的自我改进数据飞轮。
@misc{team2026omnibrain,
title = {{OmniBrain-0: A Unified Agentic Model for Embodied Intelligence}},
author = {Xiaomi Embodied Intelligence Team and Yuechen Luo and Chong He and Fang Li and Shaoqing Xu and Jingkai Wang and Menglun Su and Jiaxin Liu and Shengyin Jiang and Zhiyi Lai and Long Chen and Longlong Wang and Liang Hu and Haochen Liu and Yuncheng Jiang and Hongwei Xie and Jieming Cui and Hanbing Li and Junjie Huang and Zhixiang Duan and Hexin Zhang and Weixiang Liang and Ying Gong and Yuhan Wang and Yong Pan and Pengna Li and Kangyi Wu and Ken Deng and Kun Ma and Fuxi Wen and Zhi-xin Yang and Hangjun Ye},
year = {2026},
eprint = {2610.36013},
archivePrefix = {arXiv},
primaryClass = {cs.CL},
url = {https://arxiv.org/abs/2610.36013},
}
VLA that reasons in discrete motion tokens (a chain of dynamics) shared across embodiments; predicted futures are scored against what actually happened, so action-free human videos give RL signal. Sim and real tests.VLA 用跨本体共享的离散运动 token 链（Chain-of-Dynamics）表达动作思考；预测每条链导致的未来并与真实结果比对来强化，使没有动作标注的人类视频也能提供监督。在仿真与真机上验证人到机器人的迁移。
@misc{kim2026thinkworld,
title = {{ThinkWorld: World Imagination Reinforces Chain-of-Dynamics Reasoning}},
author = {Chaehyun Kim and Chi-Pin Huang and Jisoo Kim and Min-Hung Chen and Yu-Chiang Frank Wang and Seungryong Kim and Fu-En Yang},
year = {2026},
eprint = {2610.37666},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.37666},
}
Test-time add-on for frozen VLAs: a GNN retrieves future motion primitives from a graph of training demos and gates them into the actions. Reports up to +12% in sim and +40% real (GR00T N1.7, π0.5, StableVLA).面向冻结 VLA 的测试时插件：把训练示范组织成运动基元图，用 GNN 定位当前执行并检索后续基元，再按维度门控融合到预测动作中；报告在 GR00T N1.7、π0.5 等上仿真最高 +12%、真机 +40%。
@misc{quang2026retrieve,
title = {{Retrieve to Act: Motion Primitive Graph Retrieval for Robust VLA Execution}},
author = {Pham Tri Quang and Bao-Ngoc Dao and Minh Duc Nguyen and Duc-Duy Nguyen and Hung T. Ho and Linh Dang Le and Trong-Bao Ho and Thien-Loc Ha and Tan Q. Nguyen and Tuan Dam and Vu N. Duong and Daniel Sonntag and Ngan Le and Khoa D. Doan and Jan Peters and An Thai Le and Duy M. H. Nguyen and Vien Anh Ngo},
year = {2026},
eprint = {2610.37215},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.37215},
}
@misc{wang2026pdrvla,
title = {{PDRVLA: A Vision-Language-Action Model with Physical Dynamics Representations}},
author = {Hanqing Wang and Zhaotian Cai and Ruichuan An and Bohan Zeng and Tao Zhang and Yilin Chen and Xu Jiang and Kaiquan Chen and Chongyang Wu and Haoran Li and Fan Yang and Mingan Lin and Wentao Zhang and Zenan Zhou and Youzheng Wu},
year = {2026},
eprint = {2610.37381},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.37381},
}
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%。
A coding agent drives a bimanual robot through one command line; after each failure it reads the logs and writes or fixes small reusable tool programs.编程智能体通过一条命令行驱动双臂机器人；每次失败后读取日志，编写或修复可复用的小工具程序。
@misc{gao2026hexashell,
title = {{HexaShell: A Thin Harness for Exploring the Limits of VLM Agents as Robot Coding Agents}},
author = {Hongcheng Gao and Alex Zhou and Jay Zhu and Zelin Zheng and Shijia Ge and Xiping Li and Di Wu and Yazhe Wang and Jianshu Zeng and Xuan Shangguan and Yongli Wang and Yexing Wan and Lingyu He and Zhiqi Jia and Sihang Wu and Xiao He},
year = {2026},
eprint = {2610.23479},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.23479},
}
SERA turns the high-level VLM into an agent with an editable harness that tracks task progress for a fixed VLA executor, and revises the harness from rollouts; reports 51.4% vs 44.4% overall success on RoboMME.SERA 把高层 VLM 规划器做成带可编辑执行框架的智能体，为固定的底层 VLA 记录任务进度与执行反馈，并利用环境回放自动改进框架；报告在 RoboMME 上总体成功率 51.4%，未自我演化时为 44.4%。
@misc{du2026from,
title = {{From VLM Planners to Robotic Agents: Self-Evolving Harnesses for History-Dependent Manipulation}},
author = {Tingwen Du and Haoyu Guo and Hanxue Zhang and Xuanxu Wang and Junting Dong and Chao Gao and Mingfei Han and Tao Lu and Mulin Yu and Chunhua Shen and Xiaojun Chang and Zhihui Li},
year = {2026},
eprint = {2610.29560},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.29560},
}
World action model that brings scene-level 3D flow (tracked 3D points) inside the policy rather than as a separate module or extra head, so 4D geometry and visual semantics are modelled together with actions.世界动作模型：把场景级三维流（跟踪的三维点）纳入策略内部，而不是作为独立模块或额外输出头，让 4D 几何、视觉语义与动作一起建模。
@misc{tian2026uniflowwam,
title = {{UniFlowWAM: A Unified World-Action Model with Visual Semantics and 4D Geometry}},
author = {Jialin Tian and Zhuofan Li and Xinkai Wang and Bokai Lin and Chenyi Wang and Xinyu Zhan and Cewu Lu and Lixin Yang},
year = {2026},
eprint = {2610.30991},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.30991},
}
One world model that is also the policy, across 20 robot families mapped to a shared joint-space action format; reports strong WorldArena, RoboTwin 2.0 and LIBERO-plus results and a one-step distilled humanoid policy.同一个世界模型兼作策略，用共享的关节坐标动作空间覆盖 20 类机器人；论文报告其在 WorldArena、RoboTwin 2.0、LIBERO-plus 上表现突出，并蒸馏出可实时运行的单步人形机器人策略。
@misc{sidorov2026greenwam,
title = {{GreenWAM: A World Foundation Model for Robotics}},
author = {Gennady Sidorov and Valerii Kornilov and Denis Grankin and Polina Fedotova and Aleksey Postnikov},
year = {2026},
eprint = {2610.18774},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.18774},
}
A 16B model interleaving video prediction and language reasoning for embodied understanding and long-horizon visual planning, plus a world-action extension for control; reports 51.5% on 20 RoboCasa365 tasks.16B 统一具身模型，交替进行视频预测与语言推理，兼顾具身理解与长程视觉规划，并扩展出用于机器人控制的世界-动作版本；另提出 Rynn-VP、Rynn-GEU 评测，报告在 20 个 RoboCasa365 组合任务上成功率 51.5%。
@misc{zhu2026rynnunify,
title = {{RynnUnify: Towards Unified Embodied Brain}},
author = {Minghao Zhu and Bohan Hou and Kun Xiang and Zixuan Wang and Yinzhou Tang and Yifei Yang and Zhikai Wang and Tianyi Zhang and Kehan Li and Haoyu Zhao and Siteng Huang and Leigang Qu and Sicong Leng and Dongchi Huang and Mingxiu Chen and Shijian Lu and Yue Wang and Xiaodan Liang and Zhenyu Zeng and Deli Zhao and Chengju Liu and Jianfei Yang and Xin Li},
year = {2026},
eprint = {2610.32965},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.32965},
}
Auxiliary training task for world action models: an EMA teacher sees the clean frame, the online model sees a perturbed one and aligns its action features. Reports better robustness for four WAMs.为世界-动作模型加一个自监督辅助任务：EMA 教师看原始画面，在线模型看外观扰动后的画面并对齐其动作表征；报告在 RoboTwin 2.0、LIBERO-Plus 等基准上提升四种 WAM 对视觉扰动的鲁棒性，且不损失原有性能。
@misc{zou2026pica,
title = {{PICA-WAM: Visual Perturbation Invariance via Cross-Appearance Self-Alignment in World Action Models}},
author = {Yude Zou and Yifei Yao and Guanbo Jiao and Jianming Ma and Guanjie Zheng and Xing Gao},
year = {2026},
eprint = {2610.26813},
archivePrefix = {arXiv},
primaryClass = {cs.CV},
url = {https://arxiv.org/abs/2610.26813},
}
World action model that acts in one diffusion step: 'imagination tokens' replace random-noise initialisation as an action prior. Reports ~3× speedup over direct-policy WAMs, 40 ms latency on an RTX 5090.单步扩散即可生成动作的世界动作模型：用“想象 token”代替随机噪声初始化，为动作提供先验。报告比直接策略式 WAM 快约 3 倍，编译后在 RTX 5090 上延迟低至 40 毫秒，覆盖自动驾驶与具身控制。
@misc{zhang2026oswam,
title = {{OSWAM: An Effective World Action Model with One-Step Action Generation}},
author = {Guowen Zhang and Liyi Chen and Xu Hu and Pengfei Wang and Jingfan Chen and Chenhang He and Lei Zhang},
year = {2026},
eprint = {2610.35850},
archivePrefix = {arXiv},
primaryClass = {cs.CV},
url = {https://arxiv.org/abs/2610.35850},
}
Finds a VLA at 80% on atomic skills drops to 19% when composed into mobile manipulation, diagnoses why, and builds an agentic system around it.发现 VLA 在原子技能上 80% 成功率、组合成移动操作后仅 19%，分析失败原因并据此构建智能体式系统。
@misc{wang2026cracking,
title = {{Cracking Mobile Manipulation: From Failure Diagnosis to Agentic Robot System}},
author = {Zixuan Wang and Jia-Ning Li and Yinzhou Tang and Yanzhao Yu and Biao Wu and Meng Fang and Xin Li and Yiwu Zhong},
year = {2026},
eprint = {2610.19600},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.19600},
}
Adapts a language-guided navigation policy to a new building from one iPhone scan, compiled into a training environment with synthetic instructed episodes. Reports 85% real success vs 49% for a demo-fine-tuned baseline.只需用 iPhone 扫描一次新建筑：扫描被编译成训练环境并自动合成带语言指令的导航回合，再用无评论家的闭式 Pareto 效用适配策略；报告真机成功率 85%，用人类示教后训练的基线为 49%。
@misc{yu2026one,
title = {{One Scan is Enough: Demonstration-Free Adaptation for Language-Guided Navigation}},
author = {Zhuoyuan Yu and Yuxing Long and Du Wang and Junyang Wang and Hongwei Fan and Hao Dong},
year = {2026},
eprint = {2610.33794},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.33794},
}
Fine-tunes Qwen3.5-9B on ~1.5M grounding samples auto-built from robot recordings plus other visual data; reports a 75.41 average over seven grounding benchmarks vs 73.68 for GPT-6 Astra.用自动标注的机器人操作录像加多来源视觉数据构建约 150 万条可验证的定位监督，微调 Qwen3.5-9B；报告在七个公开视觉定位基准上平均 75.41 分，GPT-6 Astra 为 73.68。
@misc{wang2026ace,
title = {{ACE-Brain-0.6: Scaling Spatial Grounding for Physical Foundation Models}},
author = {Yinhe Wang and Tao Tao and Zexin He and Shuang Luo and Wei Mao and Yao Mu and Feng Pan and Hanming Deng and Lewei Lu and Zhi Hou and Xingang Pan and Dahua Lin and Dacheng Tao and Xiaogang Wang},
year = {2026},
eprint = {2610.27040},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.27040},
}
Continual off-policy RL for VLAs — small actor/critic networks carry over between tasks, distilled on a few kept episodes to avoid forgetting; reports 31–68.5% less real-robot time on new assembly tasks.面向 VLA 的持续离策略强化学习：轻量 actor/critic 在任务间延续，用少量保留的旧任务轨迹做蒸馏以防遗忘。论文报告新的接触式装配任务所需真机训练时间减少 31–68.5%，旧任务几乎不遗忘。
@misc{kim2026accrue,
title = {{ACCRUE: Continual Off-Policy Reinforcement Learning for Robot Foundation Models}},
author = {Changyeon Kim and Kimin Lee and Yashraj Narang and Yijie Guo},
year = {2026},
eprint = {2610.18596},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.18596},
}
Post-trains an open 9B robot planner (SFT + multi-turn RL) across harnesses that call VLA skills or write control code; reports LIBERO-PRO success rising from 18.1% to 71.0% without external memory.像训练编程智能体一样对 9B 开源机器人规划器做后训练（SFT + 多轮强化学习），适配调用 VLA 技能和生成控制代码两类执行框架；报告在 LIBERO-PRO 上无外部记忆成功率从 18.1% 提升到 71.0%。
@misc{liu2026agenticroborl,
title = {{AgenticRoboRL: Bringing Coding-Agent Post-Training to Agentic Robotics}},
author = {Xinxin Liu and Bingxin Xu and Wenrui Bao and Ming Li and Wentao Bao and Chen Chen and Yuzhang Shang},
year = {2026},
eprint = {2610.28589},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.28589},
}
Distills a pretrained VLA teacher into a policy without action pretraining via on-policy supervision plus teacher-run recoveries, triggered by a reward model (AutoDAgger); reports up to 4.1x over BC on 3 real tasks.把预训练具身基础模型（教师）蒸馏到未做动作预训练的策略：在学生访问的状态上做在线监督，并由奖励模型监控进度、自动触发教师接管采集纠错轨迹（AutoDAgger，无需人工看护）；报告在 3 个真机任务上比行为克隆最高提升 4.1 倍。
@misc{lu2026tilde,
title = {{TILDE: Teacher-in-the-Loop Distillation from Embodied Foundation Models}},
author = {Yuxiang Lu and Yantong Liu and Jinghua Hou and Xueli Sun and Ruihua Han and Zhe Liu and Hengshuang Zhao},
year = {2026},
eprint = {2610.25376},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.25376},
}
Co-trains a VLA on robot demos and noisy egocentric-video actions by weighting each source by reliability and distilling from clean future frames; reports steeper scaling than the baseline up to 30k hours of data.将机器人示教与动作标签噪声较大的第一视角视频联合训练 VLA：按数据源可靠性分配监督权重，并从干净的未来画面蒸馏表征；报告在最多 3 万小时预训练数据范围内扩展斜率高于基线。
@misc{zhang2026egoforesight,
title = {{EgoForesight: Effectively Scaling Policy Learning with Noisy Egocentric Actions}},
author = {Chubin Zhang and Changyuan Wang and Yirui Wang and Ao Li and Shilin Ma and Yue Su and Cai Zhou and Xiuwei Xu and Tianle Zhang and Yansong Tang},
year = {2026},
eprint = {2610.31845},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.31845},
}
Packages VLA policies, perception models and know-how as 250 reusable "embodied skills" and measures how much they help LLM agents on physical tasks.将 VLA 策略、感知模型与经验知识封装成 250 个可复用的“具身技能”，评测它们对大模型智能体完成物理任务的帮助。
@misc{han2026embodied,
title = {{Embodied Skill Bench: Benchmarking How Well Agents with Embodied Skills Perform on Physical Tasks}},
author = {Zhi Han and Zihan Guo and Chenxi Zeng and Zicai Cui and Yue Xin and Ming Zhou and Dongzhan Zhou and Weinan Zhang and Yang Li},
year = {2026},
eprint = {2610.16238},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.16238},
}
VLA robustness benchmark on RoboCasa365: paired clean/perturbed rollouts differing in one shift, plus stage-level diagnostics of where long-horizon failures start; evaluated on GR00T N1.5.基于 RoboCasa365 的 VLA 鲁棒性诊断基准：成对比较只差一个分布偏移因素的干净/扰动执行，并定位长时序任务中失败从哪个阶段开始，而不只看总成功率；在 GR00T N1.5 上评测。
@misc{wang2026robocasa,
title = {{RoboCasa-Shift: Stage-Aware Robustness Evaluation of Vision-Language-Action Policies in Long-Horizon Household Manipulation}},
author = {Haocheng Wang and Youquan Liu and Weiyu Guo and Aokai Zhang and Donghang Wu and Hongyu Liu and Sha Zhang and Weidong Yang and Ben Fei},
year = {2026},
eprint = {2610.27861},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.27861},
}
Benchmark of 410 long-horizon task environments (manipulation, puzzles, household) for foundation-model robot agents, with a curated EP-100 split and a shared protocol checking goals, rules and physical validity.面向基础模型机器人智能体的长程任务基准：410 个任务环境（基础操作、序列谜题、多阶段家务），精选 EP-100 子集，统一协议同时检查目标完成、任务规则与物理有效性。
@misc{chen2026embodied,
title = {{Embodied Playground: A Long-Horizon Agentic Robotics Benchmark for Physical, Strategic, and Spatial Reasoning}},
author = {Feng Chen and Pei Zhou and Lixuan He and Yunchao Yao and Li Sun and Shenghua Gao and Mingyu Ding and Kaifeng Zhang and Kaichen Zhou and Yi Ma},
year = {2026},
eprint = {2610.27039},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.27039},
}
BEHAVIOR-Spatial, built on BEHAVIOR-1K: spatial QA, point grounding and robot action tests. A baseline policy drops from 41.6% on held-out layouts to 16.1% when the requested target changes; oracle target cues help most.基于 BEHAVIOR-1K 的 BEHAVIOR-Spatial 基准：包含空间问答、指点定位与机器人动作评测。基线策略在未见布局上成功率 41.6%，但在同一布局中更换目标物体指令时降至 16.1%；提供真值目标提示带来的提升最大。
@misc{chen2026behavior,
title = {{Behavior Spatial Benchmark: Evaluating Spatial Understanding for Robotics}},
author = {Sirui Chen and Huang Huang and Qineng Wang and Wenlong Huang and Manling Li and Ruohan Zhang and Li Fei-Fei and Jiajun Wu},
year = {2026},
eprint = {2610.34051},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.34051},
}
Benchmark where a robot must run physics experiments to find an unknown answer: 30 experiments, 5 disciplines, parameters resampled per instance. Reports even the best frontier agent succeeds under half the time.“为求知而行动”的基准：机器人在物理仿真实验室中自行操作仪器、读数，测出每次随机重采样的未知参数。共 30 个实验、5 个学科；报告最强的前沿模型智能体在原生控制下成功率也不到一半，失败多源于与仪器交互。
@misc{gao2026embodiedlab,
title = {{EmbodiedLab: Benchmarking Embodied Agents That Act to Know}},
author = {Hongcheng Gao and Zelin Zheng and Alex Zhou and Jay Zhu and Shijia Ge and Yazhe Wang and Xiping Li and Jianshu Zeng and Di Wu and Xuan Shangguan and Yexing Wan and Lingyu He and Yongli Wang and Xiaohan Yu and Zhiqi Jia and Yue Liu and Bryan Hooi and Yinpeng Dong and Wentao Zhang and Sihang Wu and Xiao He},
year = {2026},
eprint = {2610.35385},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.35385},
}
@misc{mao2026robotagentbench,
title = {{RobotAgentBench: Towards Open-world Robot Autonomy}},
author = {Lingjun Mao and Jinglin Cao and Lukun He and Wenpeng Xu and Yuchen Yan and Hengyu Jin and Zijun Zhang and Jinzhou Tang and Songyao Jin and Ziche Liu and Yifei Shao and Junbo Huang and Fang Nan and Ruobing Han and Ziqiao Xi and Ziming Xu and Shuang Liang and Sibo Zhu and Wenyi Wu and Xinyue Wang and Yiming Huang and Xinkai Zou and Gao Zhu and Kun Zhou and Biwei Huang},
year = {2026},
eprint = {2610.34649},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.34649},
}
OpenVLA-OFT — parallel decoding, action chunking and continuous actions make VLA fine-tuning much faster and more accurate.OpenVLA-OFT：并行解码、动作分块与连续动作让 VLA 微调更快、成功率更高。
@misc{kim2025fine,
title = {{Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success}},
author = {Moo Jin Kim and Chelsea Finn and Percy Liang},
year = {2025},
eprint = {2502.19645},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2502.19645},
}
Co-training on heterogeneous data lets a mobile manipulator clean kitchens and bedrooms in homes it has never seen.通过异构数据联合训练，让移动操作机器人在从未见过的家庭中整理厨房和卧室。
@misc{intelligence2025vision,
title = {{$π_{0.5}$: a Vision-Language-Action Model with Open-World Generalization}},
author = {Physical Intelligence and Kevin Black and Noah Brown and James Darpinian and Karan Dhabalia and Danny Driess and Adnan Esmail and Michael Equi and Chelsea Finn and Niccolo Fusai and Manuel Y. Galliker and Dibya Ghosh and Lachy Groom and Karol Hausman and Brian Ichter and Szymon Jakubczak and Tim Jones and Liyiming Ke and Devin LeBlanc and Sergey Levine and Adrian Li-Bell and Mohith Mothukuri and Suraj Nair and Karl Pertsch and Allen Z. Ren and Lucy Xiaoyang Shi and Laura Smith and Jost Tobias Springenberg and Kyle Stachowicz and James Tanner and Quan Vuong and Homer Walke and Anna Walling and Haohuan Wang and Lili Yu and Ury Zhilinsky},
year = {2025},
eprint = {2504.16054},
archivePrefix = {arXiv},
primaryClass = {cs.LG},
url = {https://arxiv.org/abs/2504.16054},
}
Compresses action chunks with a DCT-based tokenizer so autoregressive VLAs can learn high-frequency, dexterous control.用基于 DCT 的分词器压缩动作块，使自回归 VLA 能学习高频灵巧控制。
@misc{pertsch2025fast,
title = {{FAST: Efficient Action Tokenization for Vision-Language-Action Models}},
author = {Karl Pertsch and Kyle Stachowicz and Brian Ichter and Danny Driess and Suraj Nair and Quan Vuong and Oier Mees and Chelsea Finn and Sergey Levine},
year = {2025},
eprint = {2501.09747},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2501.09747},
}
@misc{nvidia2025gr,
title = {{GR00T N1: An Open Foundation Model for Generalist Humanoid Robots}},
author = {NVIDIA and : and Johan Bjorck and Fernando Castañeda and Nikita Cherniadev and Xingye Da and Runyu Ding and Linxi "Jim" Fan and Yu Fang and Dieter Fox and Fengyuan Hu and Spencer Huang and Joel Jang and Zhenyu Jiang and Jan Kautz and Kaushil Kundalia and Lawrence Lao and Zhiqi Li and Zongyu Lin and Kevin Lin and Guilin Liu and Edith Llontop and Loic Magne and Ajay Mandlekar and Avnish Narayan and Soroush Nasiriany and Scott Reed and You Liang Tan and Guanzhi Wang and Zu Wang and Jing Wang and Qi Wang and Jiannan Xiang and Yuqi Xie and Yinzhen Xu and Zhenjia Xu and Seonghyeon Ye and Zhiding Yu and Ao Zhang and Hao Zhang and Yizhou Zhao and Ruijie Zheng and Yuke Zhu},
year = {2025},
eprint = {2503.14734},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2503.14734},
}
@misc{shukor2025smolvla,
title = {{SmolVLA: A Vision-Language-Action Model for Affordable and Efficient Robotics}},
author = {Mustafa Shukor and Dana Aubakirova and Francesco Capuano and Pepijn Kooijmans and Steven Palma and Adil Zouitine and Michel Aractingi and Caroline Pascal and Martino Russi and Andres Marafioti and Simon Alibert and Matthieu Cord and Thomas Wolf and Remi Cadene},
year = {2025},
eprint = {2506.01844},
archivePrefix = {arXiv},
primaryClass = {cs.LG},
url = {https://arxiv.org/abs/2506.01844},
}
@misc{team2025gemini,
title = {{Gemini Robotics: Bringing AI into the Physical World}},
author = {Gemini Robotics Team and Saminda Abeyruwan and Joshua Ainslie and Jean-Baptiste Alayrac and Montserrat Gonzalez Arenas and Travis Armstrong and Ashwin Balakrishna and Robert Baruch and Maria Bauza and Michiel Blokzijl and Steven Bohez and Konstantinos Bousmalis and Anthony Brohan and Thomas Buschmann and Arunkumar Byravan and Serkan Cabi and Ken Caluwaerts and Federico Casarini and Oscar Chang and Jose Enrique Chen and Xi Chen and Hao-Tien Lewis Chiang and Krzysztof Choromanski and David D'Ambrosio and Sudeep Dasari and Todor Davchev and Coline Devin and Norman Di Palo and Tianli Ding and Adil Dostmohamed and Danny Driess and Yilun Du and Debidatta Dwibedi and Michael Elabd and Claudio Fantacci and Cody Fong and Erik Frey and Chuyuan Fu and Marissa Giustina and Keerthana Gopalakrishnan and Laura Graesser and Leonard Hasenclever and Nicolas Heess and Brandon Hernaez and Alexander Herzog and R. Alex Hofer and Jan Humplik and Atil Iscen and Mithun George Jacob and Deepali Jain and others},
year = {2025},
eprint = {2503.20020},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2503.20020},
}
Open embodied VLMs (7B and 32B) trained for spatial grounding, affordances, trajectories and multi-step task planning — a "brain" that sits above low-level policies.开源具身视觉语言模型（7B 与 32B），面向空间定位、可供性、轨迹预测与多步任务规划，作为底层策略之上的“大脑”。
@misc{team2025robobrain,
title = {{RoboBrain 2.0 Technical Report}},
author = {BAAI RoboBrain Team and Mingyu Cao and Huajie Tan and Yuheng Ji and Xiansheng Chen and Minglan Lin and Zhiyu Li and Zhou Cao and Pengwei Wang and Enshen Zhou and Yi Han and Yingbo Tang and Xiangqi Xu and Wei Guo and Yaoxu Lyu and Yijie Xu and Jiayu Shi and Mengfei Du and Cheng Chi and Mengdi Zhao and Xiaoshuai Hao and Junkai Zhao and Xiaojie Zhang and Shanyu Rong and Huaihai Lyu and Zhengliang Cai and Yankai Fu and Ning Chen and Bolun Zhang and Lingfeng Zhang and Shuyi Zhang and Dong Liu and Xi Feng and Songjing Wang and Xiaodan Liu and Yance Jiao and Mengsi Lyu and Zhuo Chen and Chenrui He and Yulong Ao and Xue Sun and Zheqi He and Jingshu Zheng and Xi Yang and Donghai Shi and Kunchang Xie and Bochao Zhang and Shaokai Nie and Chunlei Men and Yonghua Lin and others},
year = {2025},
eprint = {2507.02029},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2507.02029},
}
A high-level VLM reads open-ended prompts and mid-task corrections and turns them into short language commands for a low-level π0 policy.高层 VLM 理解开放式指令与执行中的纠正，并将其转化为简短语言指令交给底层 π0 策略执行。
@misc{shi2025hi,
title = {{Hi Robot: Open-Ended Instruction Following with Hierarchical Vision-Language-Action Models}},
author = {Lucy Xiaoyang Shi and Brian Ichter and Michael Equi and Liyiming Ke and Karl Pertsch and Quan Vuong and James Tanner and Anna Walling and Haohuan Wang and Niccolo Fusai and Adrian Li-Bell and Danny Driess and Lachy Groom and Sergey Levine and Chelsea Finn},
year = {2025},
eprint = {2502.19417},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2502.19417},
}
ByteDance Seed's generalist VLA: robot imitation data co-trained with web vision-language data, adapted quickly with VR-collected human trajectories; shown on long-horizon bimanual mobile tasks.字节跳动 Seed 的通用 VLA——机器人数据模仿学习并与网络视觉-语言数据联合训练，再用 VR 采集的人类轨迹快速适配；在 ByteMini 双臂移动机器人上展示长时程与衣物操作任务。
@misc{cheang2025gr,
title = {{GR-3 Technical Report}},
author = {Chilam Cheang and Sijin Chen and Zhongren Cui and Yingdong Hu and Liqun Huang and Tao Kong and Hang Li and Yifeng Li and Yuxiao Liu and Xiao Ma and Hao Niu and Wenxuan Ou and Wanli Peng and Zeyu Ren and Haixin Shi and Jiawen Tian and Hongtao Wu and Xin Xiao and Yuyang Xiao and Jiafeng Xu and Yichu Yang},
year = {2025},
eprint = {2507.15493},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2507.15493},
}
Learns task-centric latent actions from videos (in DINO feature space, filtered by language) so one policy can be pretrained across robots and human video; reports beating OpenVLA with far less pretraining compute.在 DINO 特征空间中、借助语言指令过滤，从视频里学习“任务相关”的隐动作，使策略可跨机器人与人类视频预训练；报告以少得多的预训练算力超过 OpenVLA。
@misc{bu2025univla,
title = {{UniVLA: Learning to Act Anywhere with Task-centric Latent Actions}},
author = {Qingwen Bu and Yanting Yang and Jisong Cai and Shenyuan Gao and Guanghui Ren and Maoqing Yao and Ping Luo and Hongyang Li},
year = {2025},
eprint = {2505.06111},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2505.06111},
}
Adds 3D-aware position encoding to visual tokens and adaptive action grids, pretrained on 1.1M real robot episodes; new robot setups are handled by re-discretizing the grids.为视觉 token 加入 3D 位置编码，并用自适应动作网格表示动作，在 110 万条真机轨迹上预训练；迁移到新机器人时只需重新划分动作网格。
@misc{qu2025spatialvla,
title = {{SpatialVLA: Exploring Spatial Representations for Visual-Language-Action Model}},
author = {Delin Qu and Haoming Song and Qizhi Chen and Yuanqi Yao and Xinyi Ye and Yan Ding and Zhigang Wang and JiaYuan Gu and Bin Zhao and Dong Wang and Xuelong Li},
year = {2025},
eprint = {2501.15830},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2501.15830},
}
An open 7B VLA that first outputs depth-aware perception tokens, then an editable 2D trajectory, then actions; reports 70.5% zero-shot on SimplerEnv and 86.6% on LIBERO, with weights, code and data released.开源 7B VLA，先输出深度感知 token，再生成可编辑的二维轨迹，最后输出动作；论文报告 SimplerEnv 零样本 70.5%、LIBERO 86.6%，并开放权重、代码和数据。
@misc{lee2025molmoact,
title = {{MolmoAct: Action Reasoning Models that can Reason in Space}},
author = {Jason Lee and Jiafei Duan and Haoquan Fang and Yuquan Deng and Shuo Liu and Boyang Li and Bohan Fang and Jieyu Zhang and Yi Ru Wang and Sangho Lee and Winson Han and Wilbert Pumacay and Angelica Wu and Rose Hendrix and Karen Farley and Eli VanderBilt and Ali Farhadi and Dieter Fox and Ranjay Krishna},
year = {2025},
eprint = {2508.07917},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2508.07917},
}
Pretrains a grasping VLA almost only on SynGrasp-1B, a billion-frame photorealistic sim dataset, co-trained with web data; reports zero-shot sim-to-real open-vocabulary grasping and few-shot adaptation.几乎只用十亿帧级的照片级仿真抓取数据集 SynGrasp-1B（并联合网络语义数据）预训练抓取 VLA；论文报告可零样本从仿真迁移到真实开放词汇抓取，并支持少样本适配。
@misc{deng2025graspvla,
title = {{GraspVLA: a Grasping Foundation Model Pre-trained on Billion-scale Synthetic Action Data}},
author = {Shengliang Deng and Mi Yan and Songlin Wei and Haixin Ma and Yuxin Yang and Jiayi Chen and Zhiqi Zhang and Taoyu Yang and Xuheng Zhang and Wenhao Zhang and Heming Cui and Zhizheng Zhang and He Wang},
year = {2025},
eprint = {2505.03233},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2505.03233},
}
Gives each data source its own small set of learnable soft-prompt embeddings, so one plain-Transformer flow-matching policy (0.9B) trains across embodiments; evaluated on 6 sim benchmarks and 3 real robots.为每个数据来源配一组可学习的软提示嵌入，使一个基于标准 Transformer 的流匹配策略（0.9B）能跨本体联合训练；在 6 个仿真基准和 3 台真实机器人上评测。
@misc{zheng2025x,
title = {{X-VLA: Soft-Prompted Transformer as Scalable Cross-Embodiment Vision-Language-Action Model}},
author = {Jinliang Zheng and Jianxiong Li and Zhihao Wang and Dongxiu Liu and Xirui Kang and Yuchun Feng and Yinan Zheng and Jiayin Zou and Yilun Chen and Jia Zeng and Ya-Qin Zhang and Jiangmiao Pang and Jingjing Liu and Tai Wang and Xianyuan Zhan},
year = {2025},
eprint = {2510.10274},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2510.10274},
}
Pretrains spatial grounding on 2.3M samples ('where to act'), then post-trains actions with spatial prompts ('how to act'); reports +14.6% on SimplerEnv Google Robot over the same model without spatial guidance.先在 230 万条空间推理样本上做空间定位预训练（"在哪里做"），再用空间提示做动作后训练（"怎么做"）；论文报告在 SimplerEnv Google Robot 上比无空间引导的同一模型高 14.6%。
@misc{chen2025internvla,
title = {{InternVLA-M1: A Spatially Guided Vision-Language-Action Framework for Generalist Robot Policy}},
author = {Xinyi Chen and Yilun Chen and Yanwei Fu and Ning Gao and Jiaya Jia and Weiyang Jin and Hao Li and Yao Mu and Jiangmiao Pang and Yu Qiao and Yang Tian and Bin Wang and Bolun Wang and Fangjing Wang and Hanqing Wang and Tai Wang and Ziqin Wang and Xueyuan Wei and Chao Wu and Shuai Yang and Jinhui Ye and Junqiu Yu and Jia Zeng and Jingjing Zhang and Jinyu Zhang and Shi Zhang and Feng Zheng and Bowen Zhou and Yangkun Zhu},
year = {2025},
eprint = {2510.13778},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2510.13778},
}
One model for UI navigation and robot manipulation, pretrained on images, videos and robot data labelled with Set-of-Mark (what to act on) and Trace-of-Mark (how things move); model and code are public.同一个模型同时做界面操作与机器人操作：在图像、视频和机器人数据上预训练，用 Set-of-Mark 标注可操作对象、用 Trace-of-Mark 标注运动轨迹；模型与代码公开。
@misc{yang2025magma,
title = {{Magma: A Foundation Model for Multimodal AI Agents}},
author = {Jianwei Yang and Reuben Tan and Qianhui Wu and Ruijie Zheng and Baolin Peng and Yongyuan Liang and Yu Gu and Mu Cai and Seonghyeon Ye and Joel Jang and Yuquan Deng and Lars Liden and Jianfeng Gao},
year = {2025},
eprint = {2502.13130},
archivePrefix = {arXiv},
primaryClass = {cs.CV},
url = {https://arxiv.org/abs/2502.13130},
}
A dexterous VLA pretrained on human hand motion from mocap, VR and RGB video ('physical instruction tuning'), with part-level motion tokens, then aligned to 3D space and post-trained for real robot hands.灵巧手 VLA：先在动捕、VR 与普通视频中的人手动作上预训练（"物理指令微调"），使用部件级动作 token，再做三维空间对齐，并针对真实机器人灵巧手后训练。
@misc{luo2025being,
title = {{Being-H0: Vision-Language-Action Pretraining from Large-Scale Human Videos}},
author = {Hao Luo and Yicheng Feng and Wanpeng Zhang and Sipeng Zheng and Ye Wang and Haoqi Yuan and Jiazheng Liu and Chaoyi Xu and Qin Jin and Zongqing Lu},
year = {2025},
eprint = {2507.15597},
archivePrefix = {arXiv},
primaryClass = {cs.CV},
url = {https://arxiv.org/abs/2507.15597},
}
WALL-OSS: an open embodied foundation model that links instruction reasoning, subgoal decomposition and action generation in one differentiable model ('unified cross-level CoT'); training code and checkpoints are public.WALL-OSS：开源具身基础模型，在同一可微模型中串联指令推理、子目标分解与细粒度动作生成（"统一跨层级思维链"）；训练代码与模型权重已开源。
@misc{zhai2025igniting,
title = {{Igniting VLMs toward the Embodied Space}},
author = {Andy Zhai and Brae Liu and Bruno Fang and Chalse Cai and Ellie Ma and Ethan Yin and Hao Wang and Hugo Zhou and James Wang and Lights Shi and Lucy Liang and Make Wang and Qian Wang and Roy Gan and Ryan Yu and Shalfun Li and Starrick Liu and Sylas Chen and Vincent Chen and Zach Xu},
year = {2025},
eprint = {2509.11766},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2509.11766},
}
A hierarchical VLA in which a fine-tuned VLM draws a coarse 2D path for the end effector on the image and a low-level 3D policy follows it, so the VLM can be trained on cheap off-domain data such as sim or videos.分层 VLA：微调后的 VLM 在图像上画出末端执行器的粗略二维路径，再由底层三维策略跟随执行；因此 VLM 可以用仿真、视频等低成本的非本域数据训练。
@misc{li2025hamster,
title = {{HAMSTER: Hierarchical Action Models for Open-World Robot Manipulation}},
author = {Yi Li and Yuquan Deng and Jesse Zhang and Joel Jang and Marius Memmel and Raymond Yu and Caelan Reed Garrett and Fabio Ramos and Dieter Fox and Anqi Li and Abhishek Gupta and Ankit Goyal},
year = {2025},
eprint = {2502.05485},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2502.05485},
}
One model that decides when to think and when to act: it reasons at key moments (subtask done, error, human input) and otherwise acts on its latest reasoning; co-trained with synthetic reasoning-rich VL data.单一模型自主决定何时“想”、何时“做”：在子任务完成、出错或人类插话等关键时刻进行推理，其余时间依据最新推理结果生成动作；并与合成的推理型视觉-语言数据共同训练。
@misc{lin2025onetwovla,
title = {{OneTwoVLA: A Unified Vision-Language-Action Model with Adaptive Reasoning}},
author = {Fanqi Lin and Ruiqian Nai and Yingdong Hu and Jiacheng You and Junming Zhao and Yang Gao},
year = {2025},
eprint = {2505.11917},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2505.11917},
}
Pairs a 2B Qwen2-VL with a ~1B plug-in diffusion action expert with one head per robot, trained in three stages (cross-embodiment, alignment, task); reports shirt folding without task-specific tuning and 97.3 on LIBERO.将 2B 的 Qwen2-VL 与约 1B 参数、按机器人分头输出的可插拔扩散动作专家结合，分三阶段训练（跨本体预训练、本体对齐、任务适配）；报告无需任务专门微调即可叠衬衫，LIBERO 平均 97.3。
@misc{wen2025dexvla,
title = {{DexVLA: Vision-Language Model with Plug-In Diffusion Expert for General Robot Control}},
author = {Junjie Wen and Yichen Zhu and Jinming Li and Zhibin Tang and Chaomin Shen and Feifei Feng},
year = {2025},
eprint = {2502.05855},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2502.05855},
}
Keeps a VLA able to chat: phased alignment training plus a mixture-of-experts split to fight forgetting and task interference; reports strong VQA/MMMU scores and beats OpenVLA on 25 real tasks.让 VLA 保留“对话”能力：通过分阶段对齐训练和混合专家结构缓解遗忘与任务干扰；论文报告在保持 VQA/MMMU 等多模态理解成绩的同时，在 25 个真实操作任务上优于 OpenVLA。
@misc{zhou2025chatvla,
title = {{ChatVLA: Unified Multimodal Understanding and Robot Control with Vision-Language-Action Model}},
author = {Zhongyi Zhou and Yichen Zhu and Minjie Zhu and Junjie Wen and Ning Liu and Zhiyuan Xu and Weibin Meng and Ran Cheng and Yaxin Peng and Chaomin Shen and Feifei Feng},
year = {2025},
eprint = {2502.14420},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2502.14420},
}
Dual-system VLA: a value-query head scores sampled action-chunk candidates (System 2, ~4 Hz) and a light cascaded denoiser refines it (System 1, ~90 Hz); reports gains over π0 on LIBERO, SIMPLER and real robots.双系统 VLA：价值查询头对多组候选动作块打分并择优（System 2，约 4 Hz），轻量级级联去噪策略再将其细化为流畅动作（System 1，约 90 Hz）；论文报告在 LIBERO、SIMPLER 和真机上均优于 π0。
@misc{song2025hume,
title = {{Hume: Introducing System-2 Thinking in Visual-Language-Action Model}},
author = {Haoming Song and Delin Qu and Yuanqi Yao and Qizhi Chen and Qi Lv and Yiwen Tang and Modi Shi and Guanghui Ren and Maoqing Yao and Bin Zhao and Dong Wang and Xuelong Li},
year = {2025},
eprint = {2505.21432},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2505.21432},
}
Builds a bimanual VLA by composing two copies of a pretrained single-arm VLA instead of pretraining on bimanual data; reports beating a similar-size RDT-1B and narrowing the gap to π0 on real and sim tasks.不做双臂预训练，而是把两份预训练单臂 VLA 组合成协同的双臂 VLA；论文报告在真实与仿真双臂任务上优于同规模 RDT-1B，并缩小了与依赖大量双臂数据的 π0 的差距。
@misc{im2025twinvla,
title = {{TwinVLA: Data-Efficient Bimanual Manipulation with Twin Single-Arm Vision-Language-Action Models}},
author = {Hokyun Im and Euijin Jeong and Andrey Kolobov and Jianlong Fu and Youngwoon Lee},
year = {2025},
eprint = {2511.05275},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2511.05275},
}
@misc{wang2025vq,
title = {{VQ-VLA: Improving Vision-Language-Action Models via Scaling Vector-Quantized Action Tokenizers}},
author = {Yating Wang and Haoyi Zhu and Mingyu Liu and Jiange Yang and Hao-Shu Fang and Tong He},
year = {2025},
eprint = {2507.01016},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2507.01016},
}
Trains a VLA largely on world-model-generated data (re-rendered real videos, new viewpoints, sim2real, human-to-robot video); reports gains over π0 on dexterous and mobile tasks, plus a 402M variant for Jetson Orin.VLA 训练大量依赖世界模型生成的数据（真实视频重渲染、新视角、仿真转真实、人类视频转机器人）；论文报告在灵巧与移动操作任务上优于 π0，并提供可在 Jetson Orin 上运行的 4.02 亿参数小模型。
@misc{team2025gigabrain,
title = {{GigaBrain-0: A World Model-Powered Vision-Language-Action Model}},
author = {GigaBrain Team and Angen Ye and Boyuan Wang and Chaojun Ni and Guan Huang and Guosheng Zhao and Haoyun Li and Jie Li and Jiagang Zhu and Lv Feng and Peng Li and Qiuping Deng and Runqi Ouyang and Wenkang Qin and Xinze Chen and Xiaofeng Wang and Yang Wang and Yifan Li and Yilong Li and Yiran Ding and Yuan Xu and Yun Ye and Yukun Zhou and Zhehao Dong and Zhenan Wang and Zhichao Liu and Zheng Zhu},
year = {2025},
eprint = {2510.19430},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2510.19430},
}
~500 h / 100K trajectories on one mobile bimanual robot (R1 Lite) across 50 real scenes, plus G0: a VLM planner over a flow-matching VLA; reports that single-embodiment pretraining matters most for few-shot transfer.在单一移动双臂机器人（R1 Lite）上采集约 500 小时、10 万条轨迹、覆盖 50 个真实场景，并提出 G0：VLM 规划器 + 流匹配 VLA 执行器的双系统模型；论文报告单本体预训练对少样本迁移最关键。
@misc{jiang2025galaxea,
title = {{Galaxea Open-World Dataset and G0 Dual-System VLA Model}},
author = {Tao Jiang and Tianyuan Yuan and Yicheng Liu and Chenhao Lu and Jianning Cui and Xiao Liu and Shuiqi Cheng and Jiyang Gao and Huazhe Xu and Hang Zhao},
year = {2025},
eprint = {2509.00576},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2509.00576},
}
Open-source 'brain–cerebellum' stack: a RoboBrain MLLM plans and assigns tasks across heterogeneous robots, per-robot skill libraries execute, and a shared spatial/temporal/robot memory keeps them in sync.开源的“大脑–小脑”分层系统：RoboBrain 多模态大模型负责跨异构机器人的任务规划与分配，各机器人上的技能库负责执行，共享的空间/时间/本体记忆保持多机协同。
@misc{tan2025roboos,
title = {{RoboOS: A Hierarchical Embodied Framework for Cross-Embodiment and Multi-Agent Collaboration}},
author = {Huajie Tan and Xiaoshuai Hao and Cheng Chi and Minglan Lin and Yaoxu Lyu and Mingyu Cao and Dong Liang and Zhuo Chen and Mengsi Lyu and Cheng Peng and Chenrui He and Yulong Ao and Yonghua Lin and Pengwei Wang and Zhongyuan Wang and Shanghang Zhang},
year = {2025},
eprint = {2505.03673},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2505.03673},
}
A pretrained VLM plans and a diffusion controller grasps with a dexterous hand, working on domain-invariant features from foundation models; reports 90%+ grasp success across thousands of unseen cluttered scenes.由预训练 VLM 做高层规划、扩散控制器驱动灵巧手抓取，底层输入为基础模型提取的域不变特征；报告在数千个未见过的杂乱场景中抓取成功率超过 90%。
@misc{zhong2025dexgraspvla,
title = {{DexGraspVLA: A Vision-Language-Action Framework Towards General Dexterous Grasping}},
author = {Yifan Zhong and Xuchuan Huang and Ruochong Li and Ceyao Zhang and Zhang Chen and Tianrui Guan and Fanlian Zeng and Ka Num Lui and Yuyao Ye and Yitao Liang and Yaodong Yang and Yuanpei Chen},
year = {2025},
eprint = {2502.20900},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2502.20900},
}
One autoregressive model that is both a VLA policy and an image-predicting world model; the two help each other on LIBERO, and an attention mask hiding earlier actions curbs error build-up in action chunks.单个自回归模型同时充当 VLA 策略与预测未来图像的世界模型，二者在 LIBERO 上相互促进；论文还发现自回归生成动作块会累积误差，用屏蔽先前动作的注意力掩码可以缓解。
@misc{cen2025worldvla,
title = {{WorldVLA: Towards Autoregressive Action World Model}},
author = {Jun Cen and Chaohui Yu and Hangjie Yuan and Yuming Jiang and Siteng Huang and Jiayan Guo and Xin Li and Yibing Song and Hao Luo and Fan Wang and Deli Zhao and Hao Chen},
year = {2025},
eprint = {2506.21539},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2506.21539},
}
Online RL for VLAs with GRPO and only binary success rewards on OpenVLA-OFT, plus exploration tweaks; reports 99.1% on LIBERO and 70.4% on RoboTwin 1.0 (vs 39.8% for the SFT baseline).在 OpenVLA-OFT 上用 GRPO 与仅"成功/失败"二值奖励做在线强化学习，并加入若干促进探索的改动；论文报告 LIBERO 成功率 99.1%、RoboTwin 1.0 为 70.4%（SFT 基线 39.8%）。
@misc{li2025simplevla,
title = {{SimpleVLA-RL: Scaling VLA Training via Reinforcement Learning}},
author = {Haozhan Li and Yuxin Zuo and Jiale Yu and Yuhao Zhang and Zhaohui Yang and Kaiyan Zhang and Xuekai Zhu and Yuchen Zhang and Tianxing Chen and Ganqu Cui and Dehui Wang and Dingxiang Luo and Yuchen Fan and Youbang Sun and Jia Zeng and Jiangmiao Pang and Shanghang Zhang and Yu Wang and Yao Mu and Bowen Zhou and Ning Ding},
year = {2025},
eprint = {2509.09674},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2509.09674},
}
Open RL training framework for VLAs (PPO/GRPO across GPU-parallel simulators); reports 98.11% on 130 LIBERO tasks and 97.66% on 25 ManiSkill tasks after RL — a common base for VLA+RL work.面向 VLA 的开源强化学习训练框架（PPO/GRPO，支持 GPU 并行仿真器）；论文报告经 RL 后在 130 个 LIBERO 任务上成功率 98.11%、25 个 ManiSkill 任务上 97.66%，是 VLA+RL 研究常用的底座。
@misc{zang2025rlinf,
title = {{RLinf-VLA: A Unified and Efficient Framework for Reinforcement Learning of Vision-Language-Action Models}},
author = {Hongzhi Zang and Mingjie Wei and Si Xu and Yongji Wu and Zhen Guo and Yuanqing Wang and Hao Lin and Peihong Wang and Liangzhi Shi and Yuqing Xie and Zhexuan Xu and Zhihao Liu and Kang Chen and Wenhao Tang and Quanlu Zhang and Weinan Zhang and Chao Yu and Yu Wang},
year = {2025},
eprint = {2510.06710},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2510.06710},
}
A small open diffusion-head Transformer pretrained on Open X-Embodiment that fine-tunes quickly to new robots and sensors.在 Open X-Embodiment 上预训练的开源扩散头 Transformer 策略，可快速微调到新机器人和新传感器。
@misc{team2024octo,
title = {{Octo: An Open-Source Generalist Robot Policy}},
author = {Octo Model Team and Dibya Ghosh and Homer Walke and Karl Pertsch and Kevin Black and Oier Mees and Sudeep Dasari and Joey Hejna and Tobias Kreiman and Charles Xu and Jianlan Luo and You Liang Tan and Lawrence Yunliang Chen and Pannag Sanketi and Quan Vuong and Ted Xiao and Dorsa Sadigh and Chelsea Finn and Sergey Levine},
year = {2024},
eprint = {2405.12213},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2405.12213},
}
@misc{liu2024rdt,
title = {{RDT-1B: a Diffusion Foundation Model for Bimanual Manipulation}},
author = {Songming Liu and Lingxuan Wu and Bangguo Li and Hengkai Tan and Huayu Chen and Zhengyi Wang and Ke Xu and Hang Su and Jun Zhu},
year = {2024},
eprint = {2410.07864},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2410.07864},
}
Instead of emitting action tokens from the VLM, a separate diffusion action transformer is conditioned on the VLM's "cognition" feature; reports clear gains over OpenVLA in sim and on real robots.不再让 VLM 直接输出动作 token，而是用独立的扩散动作 Transformer 以 VLM 的“认知”特征为条件生成动作；报告在仿真和真机上明显优于 OpenVLA。
@misc{li2024cogact,
title = {{CogACT: A Foundational Vision-Language-Action Model for Synergizing Cognition and Action in Robotic Manipulation}},
author = {Qixiu Li and Yaobo Liang and Zeyu Wang and Lin Luo and Xi Chen and Mozheng Liao and Fangyun Wei and Yu Deng and Sicheng Xu and Yizhong Zhang and Xiaofan Wang and Bei Liu and Jianlong Fu and Jianmin Bao and Dong Chen and Yuanchun Shi and Jiaolong Yang and Baining Guo},
year = {2024},
eprint = {2411.19650},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2411.19650},
}
An early survey of VLA models, sorting them into component work (vision, language, world models), low-level control policies and high-level task planners, plus datasets, simulators and benchmarks; companion paper list.较早的 VLA 综述：把相关工作分为组件研究（视觉、语言、世界模型等）、底层控制策略与高层任务规划器，并汇总数据集、仿真器与基准；附配套论文列表。
@misc{ma2024survey,
title = {{A Survey on Vision-Language-Action Models for Embodied AI}},
author = {Yueen Ma and Zixing Song and Yuzheng Zhuang and Jianye Hao and Irwin King},
year = {2024},
eprint = {2405.14093},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2405.14093},
}
Broad embodied-AI survey — robots, simulators, embodied perception, interaction, agents and sim-to-real — with the role of multimodal and world models; backed by the widely used HCPLab paper list.覆盖面很广的具身智能综述：机器人本体、仿真器、具身感知、交互、具身智能体与 sim-to-real，并讨论多模态大模型与世界模型的作用；配套中山大学 HCPLab 维护的常用论文列表。
@misc{liu2024aligning,
title = {{Aligning Cyber Space with Physical World: A Comprehensive Survey on Embodied AI}},
author = {Yang Liu and Weixing Chen and Yongjie Bai and Xiaodan Liang and Guanbin Li and Wen Gao and Liang Lin},
year = {2024},
eprint = {2407.06886},
archivePrefix = {arXiv},
primaryClass = {cs.CV},
url = {https://arxiv.org/abs/2407.06886},
}
HPT: per-robot stems turn proprioception and vision into a few tokens for one shared transformer trunk, pretrained on 52 datasets (robot, sim, human video); reports 20%+ gains on unseen tasks.HPT：为每种机器人设计“输入头”，把本体感知与视觉映射成短 token 序列，送入共享的 Transformer 主干；在 52 个数据集（真机、仿真、人类视频）上预训练，报告在未见任务上微调提升 20% 以上。
LAPA pretrains a VLA without action labels: it learns discrete latent actions between video frames, predicts them from image + text, then maps them to robot actions; reports beating OpenVLA on real tasks.LAPA 无需动作标注即可预训练 VLA：先从视频相邻帧间学习离散隐动作，再训练模型由图像和指令预测隐动作，最后用少量机器人数据映射到真实动作；报告在真机任务上优于 OpenVLA。
@misc{ye2024latent,
title = {{Latent Action Pretraining from Videos}},
author = {Seonghyeon Ye and Joel Jang and Byeongguk Jeon and Sejune Joo and Jianwei Yang and Baolin Peng and Ajay Mandlekar and Reuben Tan and Yu-Wei Chao and Bill Yuchen Lin and Lars Liden and Kimin Lee and Jianfeng Gao and Luke Zettlemoyer and Dieter Fox and Minjoon Seo},
year = {2024},
eprint = {2410.11758},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2410.11758},
}
A 3D-aware LLM with scene, object and action tokens, aligned with diffusion models so it can imagine goal images and point clouds before planning actions; trained on ~2M 3D-language-action pairs the authors assembled.具备三维感知的大语言模型，引入场景、物体与动作交互 token，并与扩散模型对齐，可在规划动作前“想象”目标图像和点云；训练数据为作者整理的约 200 万条三维-语言-动作样本。
@misc{zhen2024d,
title = {{3D-VLA: A 3D Vision-Language-Action Generative World Model}},
author = {Haoyu Zhen and Xiaowen Qiu and Peihao Chen and Jincheng Yang and Xin Yan and Yilun Du and Yining Hong and Chuang Gan},
year = {2024},
eprint = {2403.09631},
archivePrefix = {arXiv},
primaryClass = {cs.CV},
url = {https://arxiv.org/abs/2403.09631},
}
Pretrained to generate future video on 38M clips, then fine-tuned to predict frames and actions together; reports 97.7% average success over 100+ real tabletop tasks and roughly double GR-1's success in unseen scenes.先在 3800 万段视频上预训练未来帧生成，再在机器人数据上微调，同时预测视频帧与动作轨迹；报告在 100 多个真机桌面任务上平均成功率 97.7%，在未见场景中成功率约为 GR-1 的两倍。
@misc{cheang2024gr,
title = {{GR-2: A Generative Video-Language-Action Model with Web-Scale Knowledge for Robot Manipulation}},
author = {Chi-Lam Cheang and Guangzeng Chen and Ya Jing and Tao Kong and Hang Li and Yifeng Li and Yuxiao Liu and Hongtao Wu and Jiafeng Xu and Yichu Yang and Hanbo Zhang and Minzhao Zhu},
year = {2024},
eprint = {2410.06158},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2410.06158},
}
Draws the robot's recent motion traces on the input image ('visual trace prompting') and fine-tunes OpenVLA on 150K trajectories; reports ~10% gains on SimplerEnv and 3.5x on real WidowX tasks; 4B variant too.把机器人近期的运动轨迹画在输入图像上（“视觉轨迹提示”），并在 15 万条轨迹上微调 OpenVLA；报告在 SimplerEnv 上提升约 10%、真实 WidowX 任务上提升 3.5 倍，另有基于 Phi-3-Vision 的 4B 精简版。
@misc{zheng2024tracevla,
title = {{TraceVLA: Visual Trace Prompting Enhances Spatial-Temporal Awareness for Generalist Robotic Policies}},
author = {Ruijie Zheng and Yongyuan Liang and Shuaiyi Huang and Jianfeng Gao and Hal Daumé and Andrey Kolobov and Furong Huang and Jianwei Yang},
year = {2024},
eprint = {2412.10345},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2412.10345},
}
A compact VLA built from a small multimodal model (up to 1.4B) with LoRA and a diffusion policy head, without large-scale robot pretraining; reports higher real-robot success than OpenVLA and ~20x faster inference.由小型多模态模型（至多 1.4B）加 LoRA 微调与扩散策略头构成的紧凑 VLA，无需大规模机器人预训练；报告真机成功率高于 OpenVLA，推理速度约快 20 倍。
@misc{wen2024tinyvla,
title = {{TinyVLA: Towards Fast, Data-Efficient Vision-Language-Action Models for Robotic Manipulation}},
author = {Junjie Wen and Yichen Zhu and Jinming Li and Minjie Zhu and Kun Wu and Zhiyuan Xu and Ning Liu and Ran Cheng and Chaomin Shen and Yaxin Peng and Feifei Feng and Jian Tang},
year = {2024},
eprint = {2409.12514},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2409.12514},
}
ReKep — a VLM writes manipulation tasks as Python cost functions over 3D keypoints (relational keypoint constraints), solved by a hierarchical optimizer in real time; zero-shot, no task-specific training.ReKep：由视觉语言模型把操作任务写成定义在三维关键点上的 Python 代价函数（关系关键点约束），再由分层优化器实时求解；零样本，无需针对任务训练。
@misc{huang2024rekep,
title = {{ReKep: Spatio-Temporal Reasoning of Relational Keypoint Constraints for Robotic Manipulation}},
author = {Wenlong Huang and Chen Wang and Yunzhu Li and Ruohan Zhang and Li Fei-Fei},
year = {2024},
eprint = {2409.01652},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2409.01652},
}
Extends NaVid into one video VLA model for several navigation tasks (instruction following, object search, question answering, human following).把 NaVid 扩展为单一视频 VLA 模型，统一处理多种导航任务（指令导航、目标搜索、具身问答、跟随人）。
@misc{zhang2024uni,
title = {{Uni-NaVid: A Video-based Vision-Language-Action Model for Unifying Embodied Navigation Tasks}},
author = {Jiazhao Zhang and Kunyu Wang and Shaoan Wang and Minghan Li and Haoran Liu and Songlin Wei and Zhongyuan Wang and Zhizheng Zhang and He Wang},
year = {2024},
eprint = {2412.06224},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2412.06224},
}
A VLA outputs mid-level navigation commands in language that an RL locomotion policy executes, so legged robots follow natural-language routes; adds Isaac Lab VLN benchmarks.VLA 先用语言输出中层导航指令，再由强化学习运动策略执行，让腿式机器人按自然语言指令导航；并提供 Isaac Lab 上的 VLN 基准。
@misc{cheng2024navila,
title = {{NaVILA: Legged Robot Vision-Language-Action Model for Navigation}},
author = {An-Chieh Cheng and Yandong Ji and Zhaojing Yang and Zaitian Gongye and Xueyan Zou and Jan Kautz and Erdem Bıyık and Hongxu Yin and Sifei Liu and Xiaolong Wang},
year = {2024},
eprint = {2412.04453},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2412.04453},
}
@misc{li2024evaluating,
title = {{Evaluating Real-World Robot Manipulation Policies in Simulation}},
author = {Xuanlin Li and Kyle Hsu and Jiayuan Gu and Karl Pertsch and Oier Mees and Homer Rich Walke and Chuyuan Fu and Ishikaa Lunawat and Isabel Sieh and Sean Kirmani and Sergey Levine and Jiajun Wu and Chelsea Finn and Hao Su and Quan Vuong and Ted Xiao},
year = {2024},
eprint = {2405.05941},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2405.05941},
}
100 task categories over 2,000+ objects that need common sense, world knowledge and multi-step reasoning, not just instruction following; evaluates VLAs, embodied agents and VLMs in one suite.100 类任务、2000 多个物体，任务需要常识、世界知识与多步推理，而不只是照指令执行；在同一套件中评测 VLA、具身智能体与 VLM。
@misc{zhang2024vlabench,
title = {{VLABench: A Large-Scale Benchmark for Language-Conditioned Robotics Manipulation with Long-Horizon Reasoning Tasks}},
author = {Shiduo Zhang and Zhe Xu and Peiju Liu and Xiaopeng Yu and Yuan Li and Qinghui Gao and Zhaoye Fei and Zhangyue Yin and Zuxuan Wu and Yu-Gang Jiang and Xipeng Qiu},
year = {2024},
eprint = {2412.18194},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2412.18194},
}
@misc{driess2023palm,
title = {{PaLM-E: An Embodied Multimodal Language Model}},
author = {Danny Driess and Fei Xia and Mehdi S. M. Sajjadi and Corey Lynch and Aakanksha Chowdhery and Brian Ichter and Ayzaan Wahid and Jonathan Tompson and Quan Vuong and Tianhe Yu and Wenlong Huang and Yevgen Chebotar and Pierre Sermanet and Daniel Duckworth and Sergey Levine and Vincent Vanhoucke and Karol Hausman and Marc Toussaint and Klaus Greff and Andy Zeng and Igor Mordatch and Pete Florence},
year = {2023},
eprint = {2303.03378},
archivePrefix = {arXiv},
primaryClass = {cs.LG},
url = {https://arxiv.org/abs/2303.03378},
}
Survey of how pretrained foundation models are used in robot perception, decision-making and control, and what holds them back — scarce robot data, safety and uncertainty, real-time execution.综述预训练基础模型如何用于机器人的感知、决策与控制，以及主要瓶颈：机器人数据稀缺、安全性与不确定性量化、实时执行。
@misc{firoozi2023foundation,
title = {{Foundation Models in Robotics: Applications, Challenges, and the Future}},
author = {Roya Firoozi and Johnathan Tucker and Stephen Tian and Anirudha Majumdar and Jiankai Sun and Weiyu Liu and Yuke Zhu and Shuran Song and Ashish Kapoor and Karol Hausman and Brian Ichter and Danny Driess and Jiajun Wu and Cewu Lu and Mac Schwager},
year = {2023},
eprint = {2312.07843},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2312.07843},
}
Self-supervised image features that work frozen for depth, segmentation and retrieval; a common visual backbone in VLAs (e.g. OpenVLA) and latent world models.自监督学习的图像特征，冻结即可用于深度估计、分割和检索；是 VLA（如 OpenVLA）和潜空间世界模型常用的视觉骨干。
@misc{oquab2023dinov,
title = {{DINOv2: Learning Robust Visual Features without Supervision}},
author = {Maxime Oquab and Timothée Darcet and Théo Moutakanni and Huy Vo and Marc Szafraniec and Vasil Khalidov and Pierre Fernandez and Daniel Haziza and Francisco Massa and Alaaeldin El-Nouby and Mahmoud Assran and Nicolas Ballas and Wojciech Galuba and Russell Howes and Po-Yao Huang and Shang-Wen Li and Ishan Misra and Michael Rabbat and Vasu Sharma and Gabriel Synnaeve and Hu Xu and Hervé Jegou and Julien Mairal and Patrick Labatut and Armand Joulin and Piotr Bojanowski},
year = {2023},
eprint = {2304.07193},
archivePrefix = {arXiv},
primaryClass = {cs.CV},
url = {https://arxiv.org/abs/2304.07193},
}
Tokenized actions plus a Transformer trained on 130k real demos showed that robot policies scale with data diversity.将动作离散化为 token，用 13 万条真实演示训练 Transformer，证明机器人策略会随数据多样性扩展。
@misc{brohan2022rt,
title = {{RT-1: Robotics Transformer for Real-World Control at Scale}},
author = {Anthony Brohan and Noah Brown and Justice Carbajal and Yevgen Chebotar and Joseph Dabis and Chelsea Finn and Keerthana Gopalakrishnan and Karol Hausman and Alex Herzog and Jasmine Hsu and Julian Ibarz and Brian Ichter and Alex Irpan and Tomas Jackson and Sally Jesmonth and Nikhil J Joshi and Ryan Julian and Dmitry Kalashnikov and Yuheng Kuang and Isabel Leal and Kuang-Huei Lee and Sergey Levine and Yao Lu and Utsav Malla and Deeksha Manjunath and Igor Mordatch and Ofir Nachum and Carolina Parada and Jodilyn Peralta and Emily Perez and Karl Pertsch and Jornell Quiambao and Kanishka Rao and Michael Ryoo and Grecia Salazar and Pannag Sanketi and Kevin Sayed and Jaspiar Singh and Sumedh Sontakke and Austin Stone and Clayton Tan and Huong Tran and Vincent Vanhoucke and Steve Vega and Quan Vuong and Fei Xia and Ted Xiao and Peng Xu and Sichun Xu and Tianhe Yu and others},
year = {2022},
eprint = {2212.06817},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2212.06817},
}
Gato — one Transformer with one set of weights plays games, captions images and stacks blocks with a real arm.Gato：同一套权重的 Transformer 既能玩游戏、写图像描述，也能用真实机械臂堆积木。
@misc{reed2022generalist,
title = {{A Generalist Agent}},
author = {Scott Reed and Konrad Zolna and Emilio Parisotto and Sergio Gomez Colmenarejo and Alexander Novikov and Gabriel Barth-Maron and Mai Gimenez and Yury Sulsky and Jackie Kay and Jost Tobias Springenberg and Tom Eccles and Jake Bruce and Ali Razavi and Ashley Edwards and Nicolas Heess and Yutian Chen and Raia Hadsell and Oriol Vinyals and Mahyar Bordbar and Nando de Freitas},
year = {2022},
eprint = {2205.06175},
archivePrefix = {arXiv},
primaryClass = {cs.AI},
url = {https://arxiv.org/abs/2205.06175},
}
Casts many tabletop tasks as interleaved text-and-image prompts to one Transformer policy, with the VIMA-Bench suite for systematic generalization tests.把多种桌面任务统一成图文交错的提示，交给一个 Transformer 策略执行，并配套 VIMA-Bench 用于系统性泛化评测。
@misc{jiang2022vima,
title = {{VIMA: General Robot Manipulation with Multimodal Prompts}},
author = {Yunfan Jiang and Agrim Gupta and Zichen Zhang and Guanzhi Wang and Yongqiang Dou and Yanjun Chen and Li Fei-Fei and Anima Anandkumar and Yuke Zhu and Linxi Fan},
year = {2022},
eprint = {2210.03094},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2210.03094},
}
PerAct — a Perceiver Transformer over voxelized RGB-D observations that predicts the next keyframe gripper pose as voxel classification; one language-conditioned policy for many RLBench and real tasks.PerAct：在体素化的 RGB-D 观测上运行 Perceiver Transformer，把下一关键帧夹爪位姿当作体素分类来预测；一个语言条件策略即可覆盖众多 RLBench 与真机任务。
Chains an LLM (parses landmarks from the instruction), CLIP (finds them in images) and a visual navigation model to follow route instructions on a real robot without language-annotated robot data.将大语言模型（从指令中解析地标）、CLIP（在图像中找到地标）与视觉导航模型串联，无需带语言标注的机器人数据即可在真实机器人上按路线指令导航。
@misc{shah2022lm,
title = {{LM-Nav: Robotic Navigation with Large Pre-Trained Models of Language, Vision, and Action}},
author = {Dhruv Shah and Blazej Osinski and Brian Ichter and Sergey Levine},
year = {2022},
eprint = {2207.04429},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2207.04429},
}
CLIPort — combines CLIP's semantic "what" stream with Transporter's spatial "where" stream for language-conditioned tabletop manipulation; an early bridge from web vision-language models to robot actions.CLIPort：把 CLIP 提供的语义"是什么"通路与 Transporter 的空间"在哪里"通路结合，实现语言条件的桌面操作；是较早把网络视觉语言模型接入机器人动作的工作。
@misc{shridhar2021cliport,
title = {{CLIPort: What and Where Pathways for Robotic Manipulation}},
author = {Mohit Shridhar and Lucas Manuelli and Dieter Fox},
year = {2021},
eprint = {2109.12098},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2109.12098},
}
Chaehyun Kim, Chi-Pin Huang, Jisoo Kim, Min-Hung Chen et al.
Abstract摘要
Vision-Language-Action (VLA) models have made substantial progress in robotic manipulation, but their generalization remains constrained by the diversity and coverage of action-labeled robot demonstrations. Human videos provide much broader behavioral and environmental experience, while learning from them faces a fundamental trade-off between scalable data and explicit action grounding. We introduce ThinkWorld, a VLA learning framework that grounds action reasoning in world dynamics and strengthens it through reinforcement learning with world imagination. ThinkWorld represents each action …
@misc{kim2026thinkworld,
title = {{ThinkWorld: World Imagination Reinforces Chain-of-Dynamics Reasoning}},
author = {Chaehyun Kim and Chi-Pin Huang and Jisoo Kim and Min-Hung Chen and Yu-Chiang Frank Wang and Seungryong Kim and Fu-En Yang},
year = {2026},
eprint = {2610.37666},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.37666},
}
Keru Zhou, Zhirui Fang, Xinrun Xu, Yaopeng Lou et al.
Abstract摘要
As egocentric video becomes increasingly available for embodied learning, an important research question is how to turn this low-cost, scalable source of data into robot manipulation capabilities. Egocentric videos naturally provide rich context for high-level task understanding, while real-robot demonstrations with wrist observations and action annotations provide supervision for fine-grained action alignment. We present Ego-CEVR, a data-native framework for hierarchical visual representation learning that assigns distinct learning objectives to global and interaction representations. The …
@misc{zhou2026ego,
title = {{Ego-CEVR: Learning Cross-Embodiment Visual Representations from Human Egocentric Videos}},
author = {Keru Zhou and Zhirui Fang and Xinrun Xu and Yaopeng Lou and Zeyuan He and Zihao Zhu and Xinze Niu and Chen Wang and others},
year = {2026},
eprint = {2610.37509},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.37509},
}
Hanqing Wang, Zhaotian Cai, Ruichuan An, Bohan Zeng et al.
Abstract摘要
Vision--Language--Action (VLA) models generate robot actions from visual observations and language instructions, but their understanding of physical dynamics remains limited. Video world models can capture scene dynamics, yet generating future videos introduces high latency into the control loop. We introduce PDRVLA, whose lightweight Physical Dynamics Module extracts physical tokens from vision--language features to condition action generation without online video generation. The module learns through direct prediction of future RGB, depth, and optical-flow features, with auxiliary …
@misc{wang2026pdrvla,
title = {{PDRVLA: A Vision-Language-Action Model with Physical Dynamics Representations}},
author = {Hanqing Wang and Zhaotian Cai and Ruichuan An and Bohan Zeng and Tao Zhang and Yilin Chen and Xu Jiang and Kaiquan Chen and others},
year = {2026},
eprint = {2610.37381},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.37381},
}
Bojun Zou, Wenhao Yu, Jiaohao Li, Jiaxuan Lin et al.
Abstract摘要
Vision-Language-Action (VLA) models integrate high-level reasoning with low-level action execution for end-to-end robotic control. However, most VLA models target static manipulation in unconstrained environments, overlooking spatial constraints and dynamic obstacles. Fine-tuning VLAs on obstacle-avoidance data is a straightforward solution, but collecting such data is costly, especially for dynamic scenarios, and may degrade their original task capabilities. To address these limitations, we propose VLA-Stem, a lightweight, reinforcement-learning-based safety mechanism for VLA systems. …
@misc{zou2026vla,
title = {{VLA-Stem: A Lightweight Safety Mechanism for Vision-Language-Action Models via Reinforcement Learning}},
author = {Bojun Zou and Wenhao Yu and Jiaohao Li and Jiaxuan Lin and Guanting Ye and Haohan Geng and Weihao Yu and Yanyong Zhang and others},
year = {2026},
eprint = {2610.37378},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.37378},
}
Hanwen Ren, Meng Zheng, Wenzhe Cui, Benjamin Planche et al.
Abstract摘要
Vision-language-action (VLA) models provide a promising paradigm for scalable robotic manipulation, yet their reliance on success-only behavioral cloning leaves them brittle; lacking corrective training signals, minor execution errors rapidly compound into unrecoverable, out-of-distribution failures. To address this limitation, we propose Adaptive Failure-Informed Learning (AFIL), an end-to-end framework that leverages failure trajectories as adaptive negative guidance for diffusion- and flow-based VLA policies. AFIL uses a pretrained VLA to generate failure rollouts online, avoiding the need …
@misc{ren2026tokens,
title = {{Tokens of Motion: Primitive Motion Learning for Human-Humanoid Skill Transfer}},
author = {Hanwen Ren and Meng Zheng and Wenzhe Cui and Benjamin Planche and Zhongpai Gao and Anwesa Choudhuri and Terrence Chen and Ziyan Wu},
year = {2026},
eprint = {2610.37330},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.37330},
}
Yonghao Yu, Lang Huang, Runyi Li, Toshihiko Yamasaki
Abstract摘要
The vision-language-action (VLA) and world-action model (WAM) communities debate whether control policies should inherit pretrained knowledge from vision-language models or from visual generators. The debate, however, bundles three separable choices: backbone initialization, the objective applied to future observations, and whether the machinery for that future loss is kept at inference. A gap between a VLA and a WAM therefore cannot simply be attributed to any one of them. We show that decoupling these choices yields a common framework for both families, which we term Just Action Model …
@misc{yu2026vision,
title = {{Vision-Language-Action or World-Action? Just Action Models}},
author = {Yonghao Yu and Lang Huang and Runyi Li and Toshihiko Yamasaki},
year = {2026},
eprint = {2610.37219},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.37219},
}
Pham Tri Quang, Bao-Ngoc Dao, Minh Duc Nguyen, Duc-Duy Nguyen et al.
Abstract摘要
Vision-Language-Action (VLA) policies demonstrate remarkable semantic generalization, yet remain highly vulnerable to out-of-distribution (OOD) visual shifts during execution. While existing test-time methods attempt to suppress spurious visual cues, they fundamentally overlook the rich, uncorrupted kinematic priors embedded in the policy's own training experience. We introduce Retrieve to Act (R2A), a test-time scaling framework that formulates OOD robustness purely as a geometric motion primitive retrieval problem. R2A bypasses fragile visual features by structuring successful training …
@misc{quang2026retrieve,
title = {{Retrieve to Act: Motion Primitive Graph Retrieval for Robust VLA Execution}},
author = {Pham Tri Quang and Bao-Ngoc Dao and Minh Duc Nguyen and Duc-Duy Nguyen and Hung T. Ho and Linh Dang Le and Trong-Bao Ho and Thien-Loc Ha and others},
year = {2026},
eprint = {2610.37215},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.37215},
}
We trained four widely used imitation-learning policies, ACT, Diffusion Policy, SmolVLA and $\pi_{0.5}$, on nested subsets of 12, 25, 50, 100 and 200 teleoperated demonstrations of one pick-and-place task with distractor objects, using one fixed training recipe per policy. Every trained model was then run on a physical SO-101 arm on a fixed set of object placements: 1,340 scored episodes in 126 blocks. ACT, trained from scratch, rose from 1 success in 40 episodes at 12 demonstrations to 37 in 40 at 100. $\pi_{0.5}$ with its action expert fine-tuned was the best policy when data was scarce (16 …
Haoyu Zhao, Yinzhou Tang, Zixuan Wang, Minghao Zhu et al.
Abstract摘要
Learning controllable world dynamics from heterogeneous interaction videos is challenging because robot embodiments use incompatible action spaces, while human videos lack native control labels. We present RynnWorld-Latent, a cross-embodiment video world model that uses the 3D-grounded latent actions introduced by RynnLAM as a shared conditioning interface. This representation enables joint pretraining on human interaction videos, robot demonstrations, and simulation rollouts without requiring a common native action space. Built on a pretrained video transformer, RynnWorld-Latent combines …
@misc{zhao2026rynnworld,
title = {{RynnWorld-Latent: A Cross-Embodiment World Model Conditioned on Universal Latent Actions}},
author = {Haoyu Zhao and Yinzhou Tang and Zixuan Wang and Minghao Zhu and Dongchi Huang and Xingyue Zhao and Qin Zhao and Kehan Li and others},
year = {2026},
eprint = {2610.36791},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.36791},
}
Lifan Zheng, Guotao Chen, Zhenyang Li, Shujie Cui et al.
Abstract摘要
We fine-tuned a separate $\pi_{0.5}$ policy for each of the five tasks in the MetaFine Open Competition, with a task-specific training mixture and number of actions executed per prediction. Training started from the official demonstrations and expert procedures. Local evaluation exposed recurring grasping, pushing, and insertion failures. We addressed them by revising the demonstrated motions and by adding expert continuations from states the policy reached, while keeping nominal demonstrations in training. Checkpoints were compared on every instruction, so that a gain on one instruction …
@misc{zheng2026evaluation,
title = {{Evaluation-Guided Demonstration Refinement for Vision-Language-Action Policies: The Winning Entry in the MetaFine Open Competition}},
author = {Lifan Zheng and Guotao Chen and Zhenyang Li and Shujie Cui and Qiang Huang and Fan Liu},
year = {2026},
eprint = {2610.36765},
archivePrefix = {arXiv},
primaryClass = {cs.CV},
url = {https://arxiv.org/abs/2610.36765},
}
Bozhao Wang, Yinglei Zhu, Zhengqing Li, Zhiyuan Zhang et al.
Abstract摘要
Differences in musculoskeletal morphology challenge policy reuse across individual models. Here we present MyoAcross, a framework that combines individual musculoskeletal model construction, morphology-conditioned policy learning and virtual model augmentation to enable muscle-driven reference-motion tracking without individual fine-tuning. Across 31 held-out individual models, MyoAcross achieves 99.24% completion and reduces tracking error by 22.60% relative to an unconditioned multilayer perceptron. Policy reuse extends to models of 22 held-out individuals constructed from single-image …
@misc{wang2026generalist,
title = {{A generalist policy for muscle-driven control across individual musculoskeletal models}},
author = {Bozhao Wang and Yinglei Zhu and Zhengqing Li and Zhiyuan Zhang and Weiming Kang and Yiheng He and XinJun Liu and Senhao Hou and others},
year = {2026},
eprint = {2610.36535},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.36535},
}
Liang Li, Guoyao Zhang, Xingyu Chen, Sikai Wu et al.
Abstract摘要
Humanoid behavior foundation models (BFMs) provide reusable whole-body motion priors: given a kinematic reference, they generate coordinated joint commands for motion replay, teleoperation, and VLA control. However, trained primarily on flat-ground motion, BFMs lack an explicit mechanism to adapt foot placement, swing clearance, and contact timing to local terrain. Many terrain-aware BFM approaches address this limitation by retraining or fine-tuning the tracker together with perception, incurring substantial data and computation costs while coupling ordinary tracking to the perception …
@misc{li2026graft,
title = {{GRAFT: Geometry-Aware Residual Adaptation for Foundation Trackers}},
author = {Liang Li and Guoyao Zhang and Xingyu Chen and Sikai Wu and Qianyang Wu and Yuhang Lin and Renjing Xu and Mingliang Zhou and others},
year = {2026},
eprint = {2610.36507},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.36507},
}
Navigation vision-language-action (VLA) policies repeatedly invoke costly semantic reasoning even when the next command remains unchanged. Command reuse can reduce computation but risks missing instruction-dependent turns or stops. We introduce ASPEN, an event-triggered scheduler based on command persistence prediction: estimating whether the frozen base policy would reissue its previous command. A lightweight multimodal predictor combines temporal observations, language instructions, execution history, and cached VLA confidence. At eligible command boundaries, sufficient predicted …
Frontier multimodal large language models (MLLMs) can control robots without robot-specific training, but how their capabilities compare and interact with post-trained vision-language-action (VLA) and world action model (WAM) policies across manipulation demands remains to be thoroughly evaluated. In this report, we evaluate GPT-6-Astra on EBench as the core of an agent that controls a robot through an agent harness under the same visual and proprioceptive observation protocol as post-trained policies. We study two agent configurations: direct control, in which the agent generates waypoint …
@misc{yao2026profiling,
title = {{Profiling a Frontier Agent with EBench: From Direct Control to Harness with Policy}},
author = {Yifei Yao and Yude Zou and Guanbo Jiao and Xing Gao and Ning Gao and Yukai Wang and Jiantong Chen and Zihou Zhu and others},
year = {2026},
eprint = {2610.35832},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.35832},
}
Efficient robotic data collection remains a major bottleneck for vision-language-action (VLA) policy learning, especially for tasks requiring accurate end-effector positioning and camera-action alignment. This challenge is particularly severe in electronic-waste disassembly, where robots operate in cluttered assemblies of densely packed, visually similar components. Full digital-twin simulation is difficult because many parts are thin, reflective, or mechanically complex, making accurate object-level meshes and contact models difficult to obtain. In this work, we focus on pre-contact reaching …
Wei Wang, Wenqiao Zhang, Zheqi Lv, Zheng Dong et al.
Abstract摘要
Vision-language-action (VLA) policies can select among discrete candidates for pick-and-place manipulation or generate continuous action chunks for motion control. Improving both forms of decision making through a common mechanism while retaining the capabilities of a pretrained policy is an important adaptation problem. Inspired by Jev's structured, probabilistic decision interface, we propose JevVLA, a lightweight energy-based adaptation framework for VLA policies. JevVLA instantiates this inspiration by modeling the relative quality of action proposals: with the base policy frozen, it …
@misc{wang2026jevvla,
title = {{JevVLA: Jev-Inspired Energy-Based Adaptation for Vision-Language-Action Policies}},
author = {Wei Wang and Wenqiao Zhang and Zheqi Lv and Zheng Dong and Yingjie Niu and Jiaqi Zhu and Jun Xiao and Chao Li and others},
year = {2026},
eprint = {2610.35010},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.35010},
}
Humans exhibit biological sensitivity to safety-critical stimuli that can trigger defensive action. Across three models and two video benchmarks, we investigate whether similar responses emerge in vision-language-action models by examining their internal representations and action tendencies in response to visible hazards and harmful consequences. Our findings suggest that task execution may remain the default response even when danger is internally registered.
Borui Xu, Hongshi Tan, Wenqi Jiang, Bingsheng He et al.
Abstract摘要
Vision--language--action (VLA) models increasingly run on edge devices aboard robots, where inference latency constrains closed-loop robot control. However, it remains unclear how to design VLA models, their inference software and edge hardware so that robots can run VLA inference in real time. Most VLA models report only task success, while prior studies cover limited action-generation strategies, divide inference into at most three stages, and lack comparisons of inference optimizations on the same models.benchmarking 18 open VLA models across released and alternative backends on NVIDIA …
@misc{xu2026vla,
title = {{VLA-HWBench: Where the Time Goes When VLA Models Run on Edge Hardware}},
author = {Borui Xu and Hongshi Tan and Wenqi Jiang and Bingsheng He and Ming Wu and Jason Zeng and Michael Heinrich},
year = {2026},
eprint = {2610.34717},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.34717},
}
Lingjun Mao, Jinglin Cao, Lukun He, Wenpeng Xu et al.
Abstract摘要
Recent advances in vision-language-action (VLA) and world-action models (WAMs) have significantly improved robotic manipulation, yet deploying robots in the open world demands a higher level of autonomy beyond executing predefined tasks. Robot agents offer a promising paradigm for this transition, integrating task discovery, contextual reasoning, physical execution, and adaptive recovery into a closed interaction loop. However, existing benchmarks largely evaluate isolated capabilities, failing to capture the complexity of open-world robot autonomy. We introduce RobotAgentBench (RAB), a …
@misc{mao2026robotagentbench,
title = {{RobotAgentBench: Towards Open-world Robot Autonomy}},
author = {Lingjun Mao and Jinglin Cao and Lukun He and Wenpeng Xu and Yuchen Yan and Hengyu Jin and Zijun Zhang and Jinzhou Tang and others},
year = {2026},
eprint = {2610.34649},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.34649},
}
Vision-language-action (VLA) models are increasingly used for robot manipulation because they map camera images and a language instruction directly to actions. Robots that cannot be tethered to a workstation, such as mobile manipulators and humanoids, must run them on onboard embedded boards, where the prefill latency of the language backbone grows with the hundreds of visual tokens processed at every control step. To meet real-time limits, training-free accelerators recompute only a subset of the visual tokens and skip the rest, either by pruning them or by reusing their cached key-value …
Haoyu Zhao, Xingyue Zhao, Zhian Su, Bohan Hou et al.
Abstract摘要
Scaling vision-language-action (VLA) models is constrained by scarce action-labeled demonstrations and incompatible control spaces across robot embodiments. We present RynnVLA-Latent, a framework that turns heterogeneous interaction videos into scalable supervision for robot policy pretraining. At its core, RynnLAM learns continuous, 3D-grounded latent actions from visual transitions without requiring native robot controls. By combining residual future-feature prediction, camera-compensated 3D-flow supervision, and camera-motion decoupling, RynnLAM encourages representations that capture …
@misc{zhao2026rynnvla,
title = {{RynnVLA-Latent: Scaling Robotic Manipulation Foundation Models with 3D-Grounded Latent Actions}},
author = {Haoyu Zhao and Xingyue Zhao and Zhian Su and Bohan Hou and Yinzhou Tang and Fengyi Wang and Kehan Li and Yuming Jiang and others},
year = {2026},
eprint = {2610.34356},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.34356},
}
Xinyu Liu, Daiying Tian, Chenwei Tang, Ruyang Han et al.
Abstract摘要
Vision-Language-Action (VLA) models have shown remarkable success in robotic manipulation, yet their ability to execute long-horizon tasks remains limited by insufficient long- term memory modeling. Existing methods either rely on implicit temporal representations within policy networks, where historical information can be gradually diluted, or maintain explicit memory buffers that suffer from redundancy, noise accumulation, and inefficient retrieval. In this paper, we propose CogMemory, a novel variational cognitive memory framework that enables VLA systems to selectively preserve and reuse …
@misc{liu2026variational,
title = {{Variational Cognitive Memory Abstraction for Long-Horizon Vision-Language-Action Models}},
author = {Xinyu Liu and Daiying Tian and Chenwei Tang and Ruyang Han and Chengyang Zhang and Qingyu Zeng and Wei Ju and Liangli Zhen and others},
year = {2026},
eprint = {2610.34251},
archivePrefix = {arXiv},
primaryClass = {cs.AI},
url = {https://arxiv.org/abs/2610.34251},
}
Yang Liu, Jiacheng Liu, Hengtao Li, Yuheng Huang et al.
Abstract摘要
Extending Vision-Language-Action models (VLAs) to humanoid robots requires coordinated whole-body control and diverse demonstrations that are costly to collect in real-world scenarios. Although egocentric human demonstrations provide a complementary source of task experience, differences in visual embodiment and missing robot proprioception complicate transfer to humanoid policies. To address these challenges, we introduce HuGe, a family of Humanoid Generalist VLAs pre-trained on over 2,000 hours of egocentric human and robot data. Specifically, HuGe employs three existing representative VLA …
@misc{liu2026huge,
title = {{HuGe: Scaling Humanoid Generalist Vision-Language-Action Models with Egocentric Experience}},
author = {Yang Liu and Jiacheng Liu and Hengtao Li and Yuheng Huang and Zijian Yang and Ji Dai and Maoxin Wang and Yi Gao and others},
year = {2026},
eprint = {2610.34186},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.34186},
}
Vision-Language-Action (VLA) models map a language instruction and the current observation directly to actions, without checking whether the instruction has been completed. For long-horizon tasks, a single failed step therefore goes unnoticed and leaves the whole task unfinished. Such failed episodes, however, often end with the task partially complete. Moreover, the unfinished steps are reusable atomic skills that recur across demonstrations. A failed attempt can thus be recovered by composing with atomic skills. We propose RoboDuo, a closed-loop dual-system framework that uses code as the …
@misc{wang2026roboduo,
title = {{RoboDuo: Attempt, Check and Atomic Recovery through Code for Vision-Language-Action Models}},
author = {Jie Wang and Bo Zhang and Renzhi Lu and Jianyu Chen},
year = {2026},
eprint = {2610.34108},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.34108},
}
Sirui Chen, Huang Huang, Qineng Wang, Wenlong Huang et al.
Abstract摘要
Robot tasks require grounding spatial language in physical scenes, but task success alone cannot establish whether spatial instructions are followed. We present BEHAVIOR-Spatial, a benchmark built on BEHAVIOR-1K that measures spatial question answering, point-based target grounding, and robot action to examine how VLM spatial understanding relates to VLA behavior. It comprises 3,256 spatial questions across 49 activities, 1,240 grounding queries across 14 manipulation tasks, and action evaluations covering those 14 tasks and 10 navigation tasks. Our manipulation protocol evaluates held-out …
@misc{chen2026behavior,
title = {{Behavior Spatial Benchmark: Evaluating Spatial Understanding for Robotics}},
author = {Sirui Chen and Huang Huang and Qineng Wang and Wenlong Huang and Manling Li and Ruohan Zhang and Li Fei-Fei and Jiajun Wu},
year = {2026},
eprint = {2610.34051},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.34051},
}
Vision-language-action (VLA) models enable robots to follow instructions in typed text, limiting broader deployment in natural speech. The traditional solution by prepending a speech recognition before VLA to convert speech into texts would incur additional latency and propagate errors. To address this, we present Speech-$\pi_{0.5}$, which maps raw speech directly to motor commands without explicit textual transcriptions. Specifically, built on $\pi_{0.5}$, we use a cross-modal projection to convert voice frames directly into audio tokens with significant response latency reduction. To …
@misc{nan2026late,
title = {{Late Breaking Results: Audio-Language-Action Models with Direct Speech Conditioning for Robotic Manipulation}},
author = {Enfu Nan and Pu Zhao and Yixiao Chen and Lin Zhao and Juyi Lin and Chen Wang and Weiwei Chen and Yanzhi Wang},
year = {2026},
eprint = {2610.33866},
archivePrefix = {arXiv},
primaryClass = {cs.CV},
url = {https://arxiv.org/abs/2610.33866},
}
Zhuoyuan Yu, Yuxing Long, Du Wang, Junyang Wang et al.
Abstract摘要
Deploying a language-guided navigation policy in a new building usually means watching it fail. Public-benchmark VLAs degrade scene by scene as layout, appearance, and language drift, and the standard fix of teleoperating fresh trajectories on site does not scale. We replace human demonstration with a single brief iPhone scan. Scan2Instruction compiles that scan into a complete training environment, automatically synthesizing executable navigation episodes with grounded instructions. PRPT-KS then adapts the policy through a closed-form Pareto utility read directly off the reconstructed scene, …
@misc{yu2026one,
title = {{One Scan is Enough: Demonstration-Free Adaptation for Language-Guided Navigation}},
author = {Zhuoyuan Yu and Yuxing Long and Du Wang and Junyang Wang and Hongwei Fan and Hao Dong},
year = {2026},
eprint = {2610.33794},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.33794},
}
Qingyuan Cai, Bo Zhang, Yongzhen Huang, Jianyu Chen
Abstract摘要
Vision-Language-Action (VLA) models are bottlenecked by the scarcity of robot data. Egocentric human manipulation videos offer an abundant, diverse, and semantically-rich alternative that inherently captures manipulation intent. However, the embodiment gap makes it difficult to extract action-relevant representations from these videos. From the perspective of action execution, human hands and robots share analogous manipulation primitives---reaching, grasping, placing---yet exhibit distinct kinematics tied to their embodiments. This might motivate our key approach: egocentric video …
@misc{cai2026exe,
title = {{EXE: Egocentric Cross-Embodiment Pretraining via Factorized Latent Actions for VLA Models}},
author = {Qingyuan Cai and Bo Zhang and Yongzhen Huang and Jianyu Chen},
year = {2026},
eprint = {2610.33685},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.33685},
}
Vision-language-action (VLA) models have demonstrated strong capabilities in robotic manipulation, yet their reliance on dense, continuous-valued computation presents a fundamental challenge for energy-constrained deployment. Inspired by the sparse, event-driven signaling and temporal integration mechanisms of biological neurons, a promising alternative is to introduce spiking computation into VLA models. However, enabling such computation without restructuring pretrained policies or compromising closed-loop control remains an open challenge. To address this challenge, we introduce the …
@misc{liu2026power,
title = {{Power-of-Two Leaky Integrate-and-Fire Neurons for Event-Driven Vision-Language-Action Inference}},
author = {Rui Liu and He Lv and Ziyao Zhang and Rui Ma and Rongxiang Nie},
year = {2026},
eprint = {2610.33658},
archivePrefix = {arXiv},
primaryClass = {cs.AI},
url = {https://arxiv.org/abs/2610.33658},
}
Online distillation of vision-language-action (VLA) policies commonly relies on teacher action scores that an inference-only interface may not expose. Trajectory-only Adversarial Distillation (\gad{}) instead turns a fixed bank of teacher joint-position features into adaptive feedback: a discriminator compares feature marginals from teacher and student rollouts and supplies rewards for on-policy learning. The student interacts with the environment without online teacher queries or environment task rewards. On ManiSkill \texttt{PutOnPlateInScene25Main}, \gad{} improves OpenVLA-OFT (7B) from …
Xuhong Qiu, Zhuoyu Wu, Qiawei Zheng, Wenhui Ou et al.
Abstract摘要
Vision-language-action (VLA) models enable robots to perform tasks specified through natural-language instructions, but their large weight footprint challenges deployment on memory-constrained onboard platforms. Ultra-low-bit quantization reduces this footprint, yet preserving task performance and controlling weight recovery costs remain challenging. We propose Action-Guided Quantization, which combines action sensitivity with a measured proxy for weight recovery cost to allocate matrix bitwidths under storage budgets. Per-channel calibration compensates for quantization error, while a …
@misc{qiu2026agq,
title = {{AGQ: Action-Guided Ultra-Low-Bit Quantization for Vision-Language-Action Models}},
author = {Xuhong Qiu and Zhuoyu Wu and Qiawei Zheng and Wenhui Ou and Wenqi Fang and Rapha\"{e}l C. -W. Phan},
year = {2026},
eprint = {2610.33074},
archivePrefix = {arXiv},
primaryClass = {cs.CV},
url = {https://arxiv.org/abs/2610.33074},
}
Vision-Language-Action (VLA) models perform well on robotic manipulation, yet remain brittle once their own actions carry them into states their training data barely covers. Recovering from such failures usually requires recovery demonstrations, policy retraining, or a separate model at inference time. We introduce Backtrack, a lightweight failure detection and recovery mechanism for frozen VLAs that needs none of these. A small failure prediction head, trained only on outcome-labeled rollouts of the frozen policy, reads the VLA's existing internal representations to estimate whether the …
Vision-language-action (VLA) policies can adapt to local human intervention, while federated learning can combine distributed robot experience without pooling raw records. We examine their prospective intersection: sharing adaptations derived from diverse human corrections across tasks, environments, people, and robot embodiments. Corrections differ in modality, safety semantics(urgency, behavioral scope, etc.), and privacy sensitivity. A shared update that improves mean task success may suppress a rare safety intervention or produce an unsafe response on another robot. We argue for three …
Zongyao Zhao, Cheng-Jie Zhao, Ruihang Jiang, Bing Zheng et al.
Abstract摘要
A failed vision-language-action (VLA) rollout does not imply that every action contributed to failure. Yet in group-relative policy optimization (GRPO), a single trajectory-level advantage is assigned to all actions in a rollout. Potentially useful recovery attempts can thus be penalized alongside the errors they aim to correct. Ideal stopping conditions for failure feedback are characterized using success-probability and policy-score martingales. At these stops, the removed gradient contribution has zero conditional expectation given the retained information. Truncation therefore preserves …
@misc{zhao2026when,
title = {{When to Stop Penalizing Actions? Martingale GRPO for VLA Reinforcement Learning}},
author = {Zongyao Zhao and Cheng-Jie Zhao and Ruihang Jiang and Bing Zheng and Zhaolin Wang and Yuanwei Liu},
year = {2026},
eprint = {2610.32431},
archivePrefix = {arXiv},
primaryClass = {cs.AI},
url = {https://arxiv.org/abs/2610.32431},
}
Jiachi Yang, Zhiyi Xu, Hongquan Zhang, Xin Tan et al.
Abstract摘要
Generalist robot manipulation requires a single vision-language-action (VLA) policy to execute diverse skills across scenes and instructions. However, current multi-skill VLA policies often leave two aspects of an instruction insufficiently modeled: which primitive skill to execute and where/how to contact and move objects. We propose MAST-VLA (Multi-skill Adaptation with Skill-conditioned experts and Trajectory-affordance priors), a dual-conditioning framework for robot manipulation. MAST-VLA introduces skill-conditioned LoRA Mixture-of-Experts (LoRA-MoE) adapters in the LLM backbone, where …
@misc{yang2026mast,
title = {{MAST-VLA: Skill-Conditioned Adaptation with Trajectory-Affordance Priors for Multi-Skill Robot Manipulation}},
author = {Jiachi Yang and Zhiyi Xu and Hongquan Zhang and Xin Tan and Zhizhong Zhang and Jingyu Gong and Yuan Xie},
year = {2026},
eprint = {2610.32312},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.32312},
}
Giovanni Minelli, Antonio Langella, Giovanni B. Dessy, Lorenzo Amatucci et al.
Abstract摘要
Vision-language-action (VLA) models are typically pretrained on broad, diverse datasets and subsequently fine-tuned on a comparatively small dataset tailored to a specific target robot. Such adaptation recipe is now widely established, achieving its most mature results on fixed-base arms due to their static base dynamics and consistent reference frame. Mounting the same arm on a quadruped extends the workspace to the entire floor, but fundamentally changes the embodiment: rather than controlling the arm in isolation, the policy must jointly coordinate base navigation, body posture, and …
@misc{minelli2026from,
title = {{From Tabletop to Walking Bot: Deploying a VLA Model on a Legged Manipulator Robot}},
author = {Giovanni Minelli and Antonio Langella and Giovanni B. Dessy and Lorenzo Amatucci and Giulio Turrisi and Victor Barasuol and Claudio Semini},
year = {2026},
eprint = {2610.32271},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.32271},
}
Vision-Language-Action (VLA) models have emerged as powerful general-purpose robot policies, but their successes and failures remain difficult to diagnose and understand. In this work, we show that sparse autoencoder (SAE) representations of OpenVLA activations contain outcome-predictive information and can be used to predict task failure before the robot reaches an unrecoverable state. Moreover, to provide insights into what common behaviors may contribute to these outcomes, we identify individual SAE features whose strong activations are associated with specific stages of success or …
@misc{rogers2026do,
title = {{Do VLAs Know They Will Fail? Interpretable Success and Failure Patterns in Physical AI Systems}},
author = {Christian A. Rogers and Qian Lin and Daniel S. Brown},
year = {2026},
eprint = {2610.32208},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.32208},
}
Jiabing Yang, Yuan Xu, Yixiang Chen, Qisen Ma et al.
Abstract摘要
Flow-matching Vision-Language-Action (VLA) models generate expressive continuous trajectories for robotic manipulation, yet balancing real-time responsiveness with execution stability remains a persistent challenge. Standard deployment workflows typically treat numerical solver depth and execution horizon as decoupled choices. The horizon amortizes inference latency, while solver depth is kept high to preserve integration fidelity. In this paper, we reveal a fundamental compute-execution coupling in flow-matching VLAs governed by a systematic sign reversal. While coarse numerical integration …
@misc{yang2026coupling,
title = {{Coupling Execution Horizon and Sampling Depth in Flow-Matching VLA Policies}},
author = {Jiabing Yang and Yuan Xu and Yixiang Chen and Qisen Ma and Peiyan Li and Kai Wang and Chenhang Cui and Tao Yu and others},
year = {2026},
eprint = {2610.32102},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.32102},
}
Qiangong Zhou, Jiafeng Liang, Yimin Dong, Yang He et al.
Abstract摘要
Long-horizon robot tasks require retaining and using past information. We integrate Test-Time Training (TTT), which stores context in fast weights, into InternVLA-A1 and Pi0.5 to study action generation conditioned on human demonstrations through training on paired demonstrations and robot trajectories. We implement memory transfer between the two phases and truncated backpropagation, and examine training objectives, gradient coverage, and contrastive regularization across video, short-video, and single-frame cue tasks. In our experiments, adding TTT and training on paired data at the current …
@misc{zhou2026ttt,
title = {{TTT in VLA Models: Exploring Training with Human Demonstrations}},
author = {Qiangong Zhou and Jiafeng Liang and Yimin Dong and Yang He and Hongxiang Li},
year = {2026},
eprint = {2610.32094},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2610.32094},
}
When a vision-language-action (VLA) model is fine-tuned with its trainable weights in bfloat16 (bf16) and no fp32 master copy, AdamW writes each update straight into bf16, silently rounding away any update smaller than 1/512 of a weight. In 9 of 12 open VLA codebases a default or documented recipe trains this way, 4 on full weight matrices, including LeRobot's SmolVLA default and its $\pi_0$/$\pi_{0.5}$ guides. In our LIBERO runs without the copy, a probe every 100 steps finds at least 79% of updates to bf16 weights erased. On the Hugging Face Hub, 99% of sampled full-weight SmolVLA …