Preprint demonstrates differentiable learning for quadruped obstacle avoidance

Researchers report deploying an end-to-end differentiable reinforcement learning system for three-dimensional obstacle avoidance on a Unitree Go2 robot in static and dynamic environments. Their arXiv preprint presents this as the first transfer of that approach from simulation to a physical robot for this task. The system learns obstacle interactions using soft penetration penalties, then transfers a simulated teacher's behavior to a controller using depth observations. A terrain-aware module blends it with a base walking policy. The paper reports better sample efficiency and obstacle avoidance than its comparison methods in simulation; the physical demonstrations establish a narrower deployment result.