Reinforcement Learning-Driven Autonomous Navigation System for Mobile Robots in Unstructured and Dynamic Terrains

Abstract

The control of the autonomous movement of mobile robots in unstructured and dynamic environments has been a major challenge because of the changed environmental factors, which are unpredictable, moving objects, partial sensory feedback and constraints of the conventional rule-based path planning system. Traditional navigation algorithms can be slow to make effective real-time decisions in dynamic non-linear environments where terrain features and obstacle patterns continually evolve. In order to overcome these difficulties, this study suggests a Deep Q-Network (DQN)-based autonomous navigation system that can facilitate intelligent decision-making and adaptive path planning to mobile robots that have to work in unpredictable surroundings. The given framework combines the concept of reinforcement learning with the state representation based on the environment perception to enable the robot to acquire the best policies governing navigation by interacting with the environment continuously. The navigation model takes into consideration the proximity of obstacles, the direction of the target, the location of the robot and the movement actions to maximize the safe movements and efficient movements. The active terrain control is carried out with the help of adaptive optimization of rewards and experience replay mechanisms enhancing stability in navigation and convergence in learning. The system is tested and run in a simulated robotics setting developed around ROS and Gazebo, where several environments related to dynamic obstacles are tested, as well as irregular terrain conditions. Key navigation measures such as Success weighted by Path Length (SPL) and Success rate (percentage) are used to assess the performance of the proposed method to quantify reliability in navigation and efficiency in a path. The experimental outcomes show that the suggested DQN-based framework is more successful in navigation, has a stronger obstacle avoidance capability, and has more efficient path traversal than the traditional methods of navigation. The research adds a smart reinforcement learning-based architecture of navigation that improves autonomous movement through uncertain worlds as well as offers a scalable basis to upcoming real-world autonomous robotic navigation devices.