This article discusses the use of reinforcement learning to improve the mobility of quadruped robots on risky terrains. The researchers trained a generalist policy for agile locomotion on disorderly and sparse stepping stones before transferring its reusable knowledge to various more challenging terrains by fine-tuning specialist policies. The researchers formulated the task as a navigation task and proposed an exploration strategy to overcome sparse rewards and achieve high robustness. The method was validated through simulation and real-world experiments on an ANYmal-D robot, achieving a peak forward velocity of 2.5m/s on sparse stepping stones and narrow balance beams.
Publication date: 20 Nov 2023
Project Page: https://youtu.be/Z5X0J8OH6z4
Paper: https://arxiv.org/pdf/2311.10484
