Robot Map-Free Navigation System Based on Improved TD3 Algorithm

To address the critical limitations of traditional robot navigation algorithms that rely on map construction and maintenance-such as high system costs and failure in unknown environments, this paper proposes a map-free navigation system for wheeled robots based on an improved Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm. The proposed solution achieves autonomous navigation and dynamic obstacle avoidance in unknown environments through two-stage optimization: (a) To resolve the local optima trap caused by insufficient exploration efficiency, we develop an adaptive action noise decay mechanism and optimize the network architecture, enhancing path discovery capabilities in complex environments through dynamic exploration strategy adjustment. (b) For high collision risks in dynamic obstacle scenarios, we construct an LSTM-based trajectory prediction reward shaping model that enables intelligent obstacle avoidance through spatiotemporal consistency evaluation of predicted trajectories. Ablation studies demonstrate 7.56^% and 20.31% improvements in average navigation success rates from exploration mechanism enhancement and obstacle avoidance module refinement respectively in maze environments. Real-world office scenario validations show phased navigation success rate improvements of 8% and 18%, confirming the effectiveness of the algorithmic advancements.