A Method of Path Planning and Intelligent Exploration for Robot Based on Deep Reinforcement Learning

The field of path planning and intelligent exploration for robots based on reinforcement learning consistently garners widespread interest. However, robots often encounter challenges such as sparse rewards and incomplete information during environmental exploration, complicating the task at hand. This paper aims to introduce a novel deep reinforcement learning algorithmic approach that integrates random network distillation with self-imitation learning to enhance intelligent exploration. The proposed algorithm uses a curiosity-driven exploration mechanism to generate experience data throughout the exploration process. Then, it employs prioritized experience replay to select high-quality samples and utilizes self-imitation learning to imitate optimal sequence trajectories. Concurrently, the algorithm updates a new policy network designed to guide exploration behavior. The policy is further updated and refined through the Proximal Policy Optimization (PPO) algorithm. To validate the algorithm’s effectiveness, we have employed the Minigrid environment to simulate the robot’s path exploration endeavors and have conducted both ablation and comparative studies on the algorithm. The findings reveal that the algorithm is capable of accomplishing more complex exploration tasks within partially observable environments, demonstrating a significant advantage in terms of convergence speed. Consequently, the algorithm proves effective in acquiring valuable information within partially observable settings and obtaining superior exploration pathways at a small cost for robot exploration.