A Local Perception-Based Navigation Algorithm for Semi-Omnidirectional Mecanum Wheel Ground Robots Using Deep Reinforcement Learning

The integration of autonomous navigation capabilities in mobile robots operating with limited environmental information remains a significant challenge in robotics research. Traditional navigation methods often struggle with limited local information and the unique kinematics of mecanum wheel ground robots, leading to suboptimal path planning and potential collisions. An improved twin delayed deep deterministic policy gradient algorithm is proposed, which incorporates adaptive exploration noise based on LiDAR distance information and introduces priority-based learning rates for more efficient navigation. The proposed method enhances robotic adaptability in complex environments through dynamic exploration strategies guided by real-time obstacle proximity and prioritized learning mechanisms. Comparative simulations demonstrate markedly improved navigation efficiency and obstacle avoidance capabilities relative to conventional reinforcement learning approaches, with consistently faster reward convergence observed across varied scenarios.