Visual Trajectory Planning for Mobile Robots Based on Improved DDPG Algorithm

Mobile robots are widely used across various fields, with trajectory planning serving as a crucial technology for their functionality. In partially known environments, mobile robots must determine feasible trajectories to reach their destinations. The Deep Deterministic Policy Gradient (DDPG) algorithm, a reinforcement learning technique, is commonly employed to address continuous motion control problems. However, the DDPG algorithm suffers from high training costs and slow convergence. This paper proposes an enhanced trajectory planning method for mobile robots based on an improved DDPG algorithm. The method incorporates a semantic segmentation network to binarize the visual feasible region, integrates a Gated Recurrent Unit (GRU) function into the original DDPG network, and introduces a reward function based on the Artificial Potential Field (APF) algorithm. The combination of semantic segmentation binarization and the APF-based reward function reduces training complexity, while the inclusion of the GRU function improves historical memory capabilities, thus accelerating training. Experimental comparisons between Deep Q-Network (DQN), Proximal Policy Optimization (PPO), basic DDPG, and the proposed algorithm demonstrate that the latter outperforms the others in terms of trajectory planning, training efficiency, and convergence speed.