Reproduce “LangWBC: Language-directed Humanoid Whole-Body Control via End-to-end 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
Yiyang Shao, Xiaoyu Huang, Bike Zhang, Qiayuan Liao, Yuman Gao, Yufeng Chi, Zhongyu Li, Sophia Shao and 1 more
Researchers at the University of California, Berkeley developed LangWBC, an end-to-end learning framework that translates natural language commands directly into agile and robust whole-body motions for humanoid robots. It demonstrates zero-shot sim-to-real transfer and generalizes to unseen commands, enabling smooth transitions and novel behavior generation via latent space interpolation on a Unitree G1 humanoid.
Key insights
- An end-to-end language-conditioned generative controller can achieve robust, agile, and diverse whole-body humanoid motions in the real world.
- A Conditional Variational Autoencoder (CVAE) architecture can form a structured and interpretable latent space for motions, enabling generalization to unseen commands and interpolation for novel behaviors.
- Policy distillation from a strong motion-tracking teacher to a language-conditioned student is effective for learning complex, physically plausible actions from high-dimensional motion capture data.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.