Reproduce “Legged Robots that Keep on Learning: Fine-Tuning Locomotion Policies in the Real World” 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
Laura Smith, J. Chase Kew, Xue Bin Peng, Sehoon Ha, Jie Tan, Sergey Levine
Researchers from UC Berkeley and Google Research developed a system that enables legged robots to autonomously fine-tune their locomotion policies in real-world environments. This approach allows an A1 quadruped to adapt agile skills to diverse terrains, such as outdoor grass and memory foam, with under 2.5 hours of real-world training, effectively overcoming the limitations of pre-trained policies in novel conditions.
Key insights
- Continuous real-world fine-tuning offers a practical method for legged robots to adapt to novel, out-of-distribution environments where pre-trained generalization alone is insufficient.
- Combining sample-efficient off-policy RL (REDQ) with simulation pre-training enables significant performance improvements with a modest amount of real-world data collection, making physical robot learning feasible.
- A learned, versatile recovery policy is crucial for achieving high autonomy in real-world robot learning, allowing robots to self-right and reset efficiently without external assistance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.