Understanding Human Behaviour in a Robotic Guide Dog Using Parallel Deep Reinforcement Learning

This paper employs a massively parallel deep re-inforcement learning approach to model human behavior in a quadruped robot system designed to function as a guide dog for visually impaired individuals. The supervisory network operates alongside a control scheme that ensures the safe interaction between the human user and the robot. It distinguishes between several simplified human behaviors, including stopping, turning, spinning, and walking straight. The trained network builds upon a low-level control network that effectively manages the quadruped's locomotion across challenging terrains. The effectiveness of the proposed approach is validated in a physics-engine-based simulation environment.