Reproduce “FastStair: Learning to Run Up Stairs with 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
Yan Liu, Tao Yu, Haolin Song, Hongbo Zhu, Nianzong Hu, Yuzhi Hao, Xiuyong Yao, Xizhe Zang and 2 more
FastStair introduces a planner-guided, multi-stage reinforcement learning framework that enables humanoid robots to run up and down stairs with both high speed and stability. This method leverages parallel foothold planning and achieves robust real-world performance, including winning the Canton Tower Robot Run Up Competition by climbing a 33-step spiral staircase in 12 seconds with a peak velocity of 1.65 m/s.
Key insights
- Reformulating model-based optimization into a parallel discrete search can overcome computational bottlenecks, making explicit stability guidance tractable for large-scale reinforcement learning.
- A multi-stage learning approach effectively reconciles the agility-stability trade-off by initially imbuing the policy with planner-derived safety priors, then specializing for speed, and finally unifying policies for seamless performance across velocity ranges.
- Integrating Low-Rank Adaptation (LoRA) layers allows for the merging of speed-specialized 'expert' policies into a single network, enabling smooth and robust transitions across diverse velocity commands.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.