Reproduce “HumanoidArena: Benchmarking Egocentric Hierarchical Whole-body Learning” 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
Taowen Wang, Zikang Xie, Bin Yang, Yunheng Wang, Zizhao Yuan, Yuetong Fang, Yixiao Feng, Yichi Wang and 8 more
HUMANOIDARENA is a new simulation benchmark that evaluates egocentric hierarchical whole-body learning for humanoid robots, emphasizing complex "leg-critical" interaction tasks. It reveals that learned policies struggle with cross-General Motion Tracker (GMT) transfer and are sensitive to visual perturbations, experiencing significant performance drops in deployment across different low-level execution backends.
Key insights
- The performance of hierarchical whole-body policies is significantly a property of the policy-tracker pair, with certain GMTs enabling higher success rates than others (e.g., SONIC generally outperformed TWIST2).
- Policies exhibit varied robustness; visual perturbations consistently caused the most substantial performance degradation, highlighting sensitivity to egocentric appearance shifts and the need for robust perception.
- There are significant and asymmetric failures in cross-GMT transfer, indicating that high-level policies tend to overfit to the motion priors and dynamics of the GMT used during training, rather than learning truly tracker-invariant intermediate actions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.