Research On Improvement Of Multi-Object Motion Coordination Reinforcement Learning Algorithm In Specific Road Network Environment

With the development of artificial intelligence and robotics,the problem of multiobject motion coordination has become an important issue for multi-mobile robotic systems and has received more and more attention from the academic community.Multiobject motion coordination plays a key role in application scenarios such as multi-robot collaborative handling,cooperative assembly,and warehousing and logistics.In a specific road network environment,mobile robots are modeled as moving objects with moving ability,and the reinforcement learning algorithm is used to train an agent to complete the multi-object motion coordination task.This paper proposes an improved multi-object motion coordination algorithm based on the Double Deep Q-Network(DDQN).First,in order to solve the problem of sparse rewards and unbalanced sample data caused by frequent collisions during training,this method proposes the Partially Tolerant Collision(PTC)processing mode,which allows collisions between objects and punishes collisions.In this way,the agent can learn to avoid the collision from collision experiences,so that the ability of the agent to avoid collision and the round success rate can be improved.Secondly,this method proposes the Dynamic Priority Strategy(DPS)which dynamically sets the scheduling priority for each moving object according to their remaining path length,and constructs the reward function of reinforcement learning based on DPS to guide the agent to consume less time complete motion coordination.The improved multi-object motion coordination algorithm based on DDQN proposed in this paper shows higher round success rates and lower completion time in experiments.Meanwhile,ablation experiments on the PTC and DPS further demonstrate their effectiveness.In order to further solve the problem of motion coordination with more complex collision constraints,this paper proposes a Double Experience Buffer Prioritized Experience Replay multi-object motion coordination algorithm(DEBPER)based on the above-mentioned algorithm.This method uses two experience buffers to store and replay successful and failed experiences respectively,which solves the problem of unbalanced experiences in a single experience buffer,and improves the utilization of experiences so that the algorithm can further improve the round success rate.The comparative experiments in multiple motion coordination tasks show that the DEBPER algorithm proposed in this paper is more capable of handling motion coordination tasks with complex collision constraints.