Reproduce “Learning Holistic Whole-Body Loco-Manipulation with a Bipedal Mobile Manipulator” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Zhongyu Chen, Yuxuan Nai, Qian Chen, Yidong Zhu, Chen Jing, Qihan Wang, Xudong Li, Zhizhan Li and 3 more

A whole-body reinforcement-learning controller enables a 14-joint bipedal mobile manipulator to reach six-degree-of-freedom end-effector targets using only task-space commands, automatically combining arm motion with squatting, stepping, and base repositioning. With a Transformer-GRU temporal estimator and phase-aware reward gating, it achieves 88.30% simulated success, tracks targets from roughly 3–191 cm in hardware demonstrations, and outperforms matched baselines in coordinated reachability and motion smoothness.

Key insights

  • Whole-body behavior can emerge from a single end-effector interface: the policy learns to choose between arm-dominant reaching, postural adaptation, stepping, and stabilization without explicit locomotion commands.
  • Combining short-term Transformer attention with longer-term GRU memory is more effective than either estimator alone, raising simulated success from 73.23% with GRU-only context and 84.43% with Transformer-only context to 88.30%.
  • Reward gating improves coordinated task completion and reduces large position failures, although it introduces a trade-off in which the additive-reward baseline attains slightly better orientation errors.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.