Adaptive navigation of mobile robots: synergising attractor dynamics and DDPG reinforcement learning for safe dynamic obstacle avoidance

Abstract

Robot navigation in complex and dynamic environments remains a challenging problem, requiring methods that can efficiently adapt to unforeseen obstacles and goal-oriented tasks. This paper presents a novel approach that combines the biologically-inspired Attractor Dynamics Approach with the Deep Deterministic Policy Gradient (DDPG) algorithm to enable a mobile robot, specifically the e-puck robot, to navigate through cluttered spaces while avoiding collisions with moving obstacles effectively. The Attractor Dynamics Approach utilises attractors as goals and repulsive forces to avoid obstacles, offering robust and goal-oriented navigation even with very low-level sensory information. In parallel, the DDPG-based reinforcement learning component fine-tunes the robot's motion controls based on range sensor readings, ensuring precise and adaptive obstacle avoidance. The integration of these two techniques empowers the robot to autonomously explore its environment, dynamically adjust its trajectory and reach predefined targets successfully and safely.