Autonomous Steering System Using Fuzzy TD3 and GRU-Attention for Omnidirectional AMRs in Dynamic Environments

Abstract This paper addresses the critical challenge of enabling omnidirectional autonomous mobile robots (OAMRs) to navigate reliably in dynamic and uncertain environments. A novel autonomous steering system is proposed, which combines a fuzzy twin delayed deep deterministic policy gradient (Fuzzy TD3) algorithm and a gated recurrent unit (GRU)-attention mechanism for autonomous steering OAMR in dynamic environments. The fuzzy TD3 algorithm is first proposed to improve the performance of its original TD3 by fuzzifying the inputs related to robot navigation, and the GRU-attention mechanism is then applied to let the mobile robot pay attention to any static and dynamic obstacles. By integrating the fuzzy TD3 and GRU-attention mechanism, a novel autonomous steering method, dubbed FGA-TD3, is proposed to achieve point-to-point navigation by driving the OAMR. Unlike conventional Deep Reinforcement Learning (DRL) methods, FGA-TD3 leverages fuzzy logic to smooth policy gradients and incorporates an attention mechanism for robust decision-making in cluttered environments. Extensive testing demonstrates that FGA-TD3 achieves a 96% success rate in static environments and 93% in dynamic environments, significantly outperforming both TD3 (73% static, 52% dynamic) and Fuzzy TD3 (92% static, 89% dynamic). Additionally, the integration of the A* global path planning algorithm with FGA-TD3 yields a 100% success rate across various challenging environments. Theoretical analysis establishes both the convergence of the FGA-TD3 policy to a local optimum and the practical stability of the closed-loop system. The effectiveness and superiority of the proposed steering system are well exemplified by conducting comparative simulations and experiments, showing that the proposed steering system is capable of autonomously driving OAMRs to achieve desired navigation tasks.