Reproduce “Towards Torque-Driven Reinforcement Learning for Quadruped Locomotion” on OpenResearch

About this paper

Jordan Dowdy, Jean Chagas Vaz

Researchers at the University of Louisville developed a torque-driven reinforcement learning framework for the heavyweight Unitree B1 quadruped robot. The policy demonstrated stable locomotion and accurate tracking of desired linear speeds up to 3.5 m/s and angular speeds up to 1.5 rad/s in simulation, while exhibiting robustness against strong external disturbances and rough terrain without explicit linear velocity observations.

Key insights

• Torque-driven control policies can provide enhanced dynamic stability and adaptability for high-inertia quadrupeds, surpassing the capabilities of many position-based control methods.
• Robust locomotion and accurate linear velocity tracking can be achieved without direct linear velocity input to the policy, indicating an implicit learning of necessary state relationships from other observational cues.
• A relatively simple multilayer perceptron network architecture, when combined with careful reward and observation shaping, is sufficient for robust locomotion in complex, heavyweight quadruped robots, promoting modularity and adaptability.

Summary by alphaXiv.