Abstract: To address the collision avoidance problem in multi-robot systems, each robot is required to dynamically select an optimal velocity based on current environment information. However, most existing multi-robot collision avoidance algorithms either tend to deadlock in dense scenarios or sacrifice travel efficiency for safety. To this end, this paper proposes a distributed multi-robot navigation method that integrates buffered Voronoi cells (BVCs) with reinforcement learning. The method constructs BVC using only position information and uses them as observation inputs. A neural network based on a bidirectional gated recurrent module then maps the current state to a navigation action to determine an appropriate velocity. The reward function is also designed based on the distance between the BVC region and the target point, aiming to anticipate and minimize collision risk. Simulation experiments compared the proposed method with mainstream baseline methods in terms of metrics such as success rate, average speed, and travel time. The results indicate that the proposed method outperforms algorithms like RL-RVO across various scenarios, demonstrating superior adaptability and effectiveness.