Reproduce “Gait-Conditioned Reinforcement Learning with Multi-Phase Curriculum for Humanoid Locomotion” 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
Tianhu Peng, Lingfan Bao, Chengxu Zhou
University College London researchers present a unified, reference-free reinforcement learning framework that enables humanoid robots to perform diverse locomotion behaviors, including standing, walking, running, and smooth transitions, using a single recurrent policy. This approach successfully transferred to a Unitree G1 robot without real-world fine-tuning, demonstrating stable and natural movements.
Key insights
- Dynamically routing rewards based on the current gait mode is critical for enabling a single policy to effectively co-learn and coordinate diverse and potentially conflicting locomotion behaviors.
- A biologically inspired, multi-phase curriculum is essential for the stable acquisition of complex, multi-gait skills, preventing convergence failures seen in simultaneous multi-gait training.
- Integrating biomechanically grounded reward terms, such as those encouraging straight-knee support and anti-phase arm-leg coordination, can effectively promote natural, dynamically stable, and efficient motions without requiring external motion capture data.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.