Path Planning for Moving Robots in an Unknown Dynamic Area Using RND-Based Deep Reinforcement Learning

Robots can be used for various purposes in environments humans cannot reach. Dynamic environments with static and dynamic obstacles are challenging areas in robotic implementations. In this proposed study, robotic exploration is presented in an unknown dynamic environment using Random Network Distillation(RND) algorithm, which was introduced for Atari games. RND algorithm is used here for robotic navigation depending on obstacle location and environmental features. The proposed route planning robot is implemented using the Robot Operating System(ROS) and the Gazebo simulator. In previous works, Deep Reinforcement Learning based agents have shown promising outcomes. However, the reward system needs to be more robust for this path-planning task. In order to unravel the sparse reward issue, a revised RND approach is suggested. In order to dynamically alter intrinsic and extrinsic rewards, this algorithm determines the variation of the training steps count in each training episode. Laser inputs from Gazebo are fed to the RND algorithm to get the actions. Actions will define by avoiding walls or obstacles. The robotic path and map are visualized on the Gazebo. The results show the robot's actions and combine intrinsic and extrinsic rewards from RND on the application. Further, employing the RND-trained agent learns to cover environments presented as a novel part of the research. Four simulations are conducted based on the environment to validate the algorithm, including without obstacles, static obstacles, dynamic obstacles, and both. The experimental results indicate that RND agents trained in different environments can reach the goal accurately without crashes.