Legged Robots that Keep on Learning: Fine-Tuning Locomotion Policies in the Real World

TL;DR

This paper presents a practical reinforcement learning system that allows legged robots to autonomously fine-tune their locomotion policies in real-world environments, improving robustness and adaptability through modest online training.

Contribution

It introduces a real-world reinforcement learning approach enabling robots to continually adapt their locomotion policies during deployment, addressing efficiency, safety, and autonomy challenges.

Findings

Real-world fine-tuning significantly improves robot performance.

Robots can autonomously adapt to diverse environments.

Modest training suffices for substantial performance gains.

Abstract

Legged robots are physically capable of traversing a wide range of challenging environments, but designing controllers that are sufficiently robust to handle this diversity has been a long-standing challenge in robotics. Reinforcement learning presents an appealing approach for automating the controller design process and has been able to produce remarkably robust controllers when trained in a suitable range of environments. However, it is difficult to predict all likely conditions the robot will encounter during deployment and enumerate them at training-time. What if instead of training controllers that are robust enough to handle any eventuality, we enable the robot to continually learn in any setting it finds itself in? This kind of real-world reinforcement learning poses a number of challenges, including efficiency, safety, and autonomy. To address these challenges, we propose a practical robot reinforcement learning system for fine-tuning locomotion policies in the real world. We demonstrate that a modest amount of real-world training can substantially improve performance during deployment, and this enables a real A1 quadrupedal robot to autonomously fine-tune multiple locomotion skills in a range of environments, including an outdoor lawn and a variety of indoor terrains.