DRL-VO: Learning to Navigate Through Crowded Dynamic Scenes Using Velocity Obstacles

This article proposes a novel learning-based control policy with strong generalizability to new environments that enables mobile robot navigate autonomously through spaces filled both static obstacles and dense crowds of pedestrians. The uses unique combination input data generate the desired steering angle forward velocity: short history lidar data, kinematic about nearby pedestrians, subgoal point. is trained in reinforcement learning setting using reward function contains term based on velocity guide actively avoid pedestrians move toward goal. Through series 3-D simulated experiments up 55 this able achieve better balance between collision avoidance speed (i.e., higher success rate faster average speed) than state-of-the-art model-based policies, it also generalizes different crowd sizes unseen environments. An extensive hardware demonstrate ability directly work real-world zero retraining. Furthermore, show works highly constrained platform without any additional training. Lastly, several important lessons can be applied other systems are summarized.