Reproduce “Human2LocoMan: Learning Versatile Quadrupedal Manipulation with Human Pretraining” 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
Yaru Niu, Yunzhe Zhang, Mingyang Yu, Changyi Lin, Chenhao Li, Yikai Wang, Yuxiang Yang, Wenhao Yu and 6 more
Human2LocoMan introduces a system and Modularized Cross-embodiment Transformer (MXT) to enable versatile quadrupedal manipulation by effectively transferring motor policies from human demonstrations. The approach significantly improves robot task success rates and data efficiency, achieving up to an 82.7% increase in out-of-distribution success rate with human pretraining.
Key insights
- Human demonstrations can positively transfer complex motor skills to non-humanoid quadrupedal robots despite significant embodiment differences, substantially reducing robot data requirements.
- A modularized deep learning architecture (MXT) with modality-specific tokenizers and detokenizers is crucial for preserving unique data distributions and enabling effective cross-embodiment knowledge transfer.
- A unified data collection pipeline that aligns observation and action spaces across human and robot embodiments is essential for scalable imitation learning from heterogeneous sources.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.