Reproduce “Terrain Consistent Reference-Guided RL for Humanoid Navigation Autonomy” 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
William D. Compton, Zachary Olkin, Aaron D. Ames
Researchers at Caltech developed a method for humanoid robots to perform robust navigation in complex environments by training reinforcement learning policies with dynamically adjusted, terrain-consistent reference trajectories. The approach enables a Unitree G1 robot to autonomously traverse over 70 meters of varied outdoor terrain, including multiple flights of stairs, using entirely onboard sensing and computation.
Key insights
- Dynamically modulating reference trajectories to be terrain-consistent during RL training significantly enhances locomotion robustness and tracking accuracy on complex terrains, such as stairs and slopes.
- A simplified MLP-based policy architecture can achieve performance comparable to more complex CNN/attention-based networks when guided by effective terrain-aware reference synthesis and dense CLF-RL reward signals, reducing computational overhead.
- A perception-conditioned, reference-guided RL policy with a standard SE(2) velocity interface can effectively serve as a deployable locomotion layer for long-horizon autonomous humanoid navigation, bridging the gap between low-level control and high-level planning.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.