Design a path - planning strategy for mobile robot in multi-structured environment based on distributional reinforcement learning.

The article proposes a novel path-planning strategy for mobile robots moving in unknown environments by integrating Lightweight Learned Image Denoising with Instance Adaptation (LIDIA) and Quantile Regression Deep Q-Network (QR-DQN) within a unified framework. This combination approach aims to allow robots to navigate in multi-object areas without collision. Initially, depth images of the environment are captured through the robot's onboard camera. Since these raw images often contain noise, the LIDIA technique is applied to calibrate this dataset and enhance the image quality. The refined depth data is then employed to train a distributional reinforcement learning model, named QR-DQN. This process enhances the robot ability to make informed and flexible decisions under uncertainty. Moreover, a sub-goal mechanism also is integrated into the model to guide the robot through complex environments by breaking down its tasks into manageable steps. The image data before and after denoising processes, as well as the values of the reward function, are analyzed across various environmental scenarios to evaluate the effectiveness of this framework. The results in different map layouts show that the proposed method can achieve shorter distances with a smooth curve path-planning compared to other methods, even in sophisticated environmental conditions.•The depth image dataset collected via the camera is denoised and calibrated based on the Lightweight Learned Image Denoising with Instance Adaptation. The depth image dataset then serves as the data for training robot's model.•A distributional reinforcement learning model, named The Quantile Regression Deep Q-Network (QR-DQN), is applied to automatically generate the path-planning for mobile robots.•The path-planning results of QR-DQN method are compared well with other methods in different case studies.