CVN: A Collision Value Network for Crowd Navigation

Dense crowd scenes are highly dynamic and un-certain, posing significant challenges to robot navigation. Recent studies have begun to utilize deep neural networks to encode interaction behaviors among agents, followed by reinforcement learning to develop obstacle avoidance strategies. However, accurately encoding the status of complex crowds remains a significant challenge, often leading to frequent interference with agents during navigation, and greatly reduces the experience of traffic participants. To this end, this paper proposes the collision value network (CVN). CVN leverages the multi-head attention mechanism in the transformer to encode the interaction behaviors among different agents, and then addresses the crowd navigation problem with reinforcement learning. Based on the designed reward function, the value iteration method is used to obtain the approximate value function and obstacle avoidance policy. Experiments show that the proposed method can enable the robot to make reasonable obstacle avoidance decisions quickly and accurately in complex scenarios. Across 500 rounds of random navigation tests, the navigation success rate reaches 98 %, the collision probability is 2 %, and the navigation time is 10.21s, showcasing superior performance compared to existing methods.