Скачать или смотреть End-to-End Social Mobile Robot Navigation with Reinforcement Learning: A Comparative Study

This video presents the real-world validation of an end-to-end social navigation framework developed at the Department of Computer Engineering and Mathematical Sciences (DIISM) of the University of Siena. We demonstrate the capabilities of a Truncated Quantile Critics (TQC) policy deployed on a TurtleBot 4, which navigates a human-populated corridor using only stacked 2D LiDAR scans and odometry. Following a comparative study where TQC outperformed PPO and SAC in terms of reliability and safety, this experiment highlights the effective sim-to-real transfer from a high-fidelity MuJoCo training environment utilizing the Headed Social Force Model. As shown, the robot successfully handles dynamic interactions by pausing, reorienting, and respecting social distances to reach its goal without relying on camera inputs.