MotionDisco: Motion Discovery for Extreme Humanoid Loco-Manipulation
Discovers long-horizon, contact-rich humanoid loco-manipulation from scratch: LLM-guided evolutionary search over interaction sequences plus kinodynamic trajectory optimization, then RL tracking on a real humanoid.从零发现长时域、多接触的人形机器人移动操作动作：由大模型引导的进化搜索交互序列，配合运动动力学轨迹优化，再训练强化学习跟踪策略部署到真实人形机器人。
BibTeX
@misc{taouil2026motiondisco,
title = {{MotionDisco: Motion Discovery for Extreme Humanoid Loco-Manipulation}},
author = {Ilyass Taouil and Michal Ciebelski and Shafeef Omar and Haizhou Zhao and Angela Dai and Aaron M. Johnson and Majid Khadiv},
year = {2026},
eprint = {2606.06139},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2606.06139},
}Locomotion & Humanoids运动控制与人形机器人Humanoid人形humanoidloco-manipulationllmevolutionary-searchtrajectory-optimizationreal