Reinforcement Learning Based Collision Aware AI-Driven Intelligent Robotics Navigation System

Learning-based robot navigation methods suffer from high data dependency and imperfect performance in specific environments, such as inability to walk straight in open spaces and high collision rates in densely cluttered environments. To enhance artificial intelligence (AI) driven robot navigation performance, a reinforcement imitation learning navigation method based on collision prediction is proposed. In the absence of a model, the state space, action space, and reward function required for the Markov decision process (MDP) are established based on the robot's performance. Deep reinforcement learning (DRL) is utilized for training in a simulation environment, enabling the robot to navigate and avoid obstacles in multi-obstacle environments. Expert data is collected and used to continue training the policy using imitation learning methods to address the imperfections of reinforcement learning in both sparse and dense obstacle scenarios. A collision prediction model is designed, combining traditional control with deep learning. Based on the prediction results, the AI driven robot adaptively selects appropriate control strategies in different environments, greatly improving navigation safety. Through experiments, the proposed method's navigation performance and generalization ability are validated in numerous previously unseen scenarios.