Robotic Guide Dog Using Parallel Deep Reinforcement Learning

This abstract presents a massively parallel deep reinforcement learning approach designed to interpret human behavior within a quadruped robot system functioning as a guide dog for visually impaired individuals. The system includes a supervisory module integrated with a control scheme that ensures safe interaction between the human and the quadruped robot. It effectively distinguishes between different human actions, such as stopping, turning, and walking. The trained network builds upon another network capable of robustly managing quadruped locomotion over challenging terrains. The efficacy of this approach is evaluated within a simulation environment based on a physics engine.