Reproduce “ZeroWBC: Learning Natural Whole-Body Humanoid Interaction from Human Egocentric Data” 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
Haoran Yang, Jiacheng Bao, Yucheng Xin, Haoming Song, Yuyang Tian, Bin Zhao, Dong Wang, Xuelong Li
A framework called ZeroWBC enables humanoid robots to learn versatile and natural whole-body interactions directly from human egocentric data, bypassing the need for expensive robot teleoperation. The system achieved high success rates on tasks such as 96% for obstacle avoidance and 76% for zero-shot chair sitting on a Unitree G1 robot.
Key insights
- Human egocentric data provides a scalable and cost-effective supervision source, enabling the learning of diverse whole-body humanoid interactions without costly robot teleoperation.
- Prioritizing the tracking of global root and key body-part trajectories is crucial for preserving task-critical interaction intent during human-to-humanoid transfer, despite morphological differences.
- A two-stage fine-tuning approach for Vision-Language Models, combining large-scale general data with domain-specific egocentric data, is effective for achieving both broad semantic understanding and spatially grounded motion generation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.