Reproduce “CWI: Composite Humanoid Whole-Body Imitation System for Loco-manipulation” 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

Wenqi Ge, Junde Guo, Zhen Fu, Shunpeng Yang, Jiayu Chen, Hua Chen

The Composite Whole-Body Imitation (CWI) framework enables robust humanoid loco-manipulation by decoupling motion capture data utilization and employing a multi-critic, teacher-student learning approach. This system allows stable, command-conditioned locomotion and precise manipulation on a physical robot using a portable VR-based teleoperation interface.

Key insights

  • Decoupling MoCap data by body part effectively addresses data imbalance, preserving rich upper-body motions while ensuring stable, command-conditioned lower-body locomotion without extensive data filtering.
  • The multi-critic architecture significantly improves learning stability and convergence in multi-objective reinforcement learning by reducing interference between locomotion, manipulation, and motion style objectives.
  • Teacher-student distillation enables practical robot deployment with a minimalist teleoperation interface (bimanual hand poses), bridging the gap between dense training references and sparse deployment commands.

Summary by alphaXiv.

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