Reproduce “Switch: Learning Agile Skills Switching for Humanoid Robots” 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

Yuen-Fui Lau, Qihan Zhao, Yinhuai Wang, Runyi Yu, Hok Wai Tsui, Qifeng Chen, Ping Tan

A hierarchical control system named "Switch" enables humanoid robots to perform smooth, agile transitions between diverse dynamic skills, achieving a 100% success rate and improved motion fidelity through a novel skill graph and reinforcement learning approach from The Hong Kong University of Science and Technology.

Key insights

  • Augmenting motion datasets with a Skill Graph and buffer nodes effectively creates feasible inter-skill transition paths, overcoming the challenge of combinatorially sparse transition data.
  • Explicitly supervising high-frequency foot-ground contact events with the Foot-Ground Contact Reward significantly improves lower-body coordination and motion fidelity in agile skills.
  • A hierarchical system integrating offline data augmentation, specialized DRL training, and online planning enables both agile skill switching and real-time adaptation to dynamic environments and perturbations.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.