Reproduce “ThorArena: Benchmarking Humanoid Physical Interaction with Human Motion-Force Demonstrations” 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
Chenhao Yu, Hongwu Wang, Weitao Zhang, Youhao Hu, Jiachen Zhang, Gangyang Li, Alois Knoll, Shaqi Luo
ThorArena introduces a benchmark for evaluating humanoid robot control policies under physical interaction, integrating human motion-force demonstrations, a force-replay protocol, and force-aware metrics. This framework revealed that existing policies designed for kinematic tasks degrade significantly under external forces, with Thor2 demonstrating superior robust tracking and survival rates in contact-rich scenarios.
Key insights
- Evaluating humanoid policies solely under kinematic conditions is insufficient, as policies struggle with stability and accuracy when subjected to external interaction forces.
- Physical interaction forces significantly impact a humanoid's whole-body balance, motion tracking precision, and required control effort.
- Comprehensive evaluation requires metrics that go beyond kinematic error to include robustness to varying force levels, control effort, and long-horizon stability.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.