Autonomous Navigation of Mobile Robots in Complex Environments with Global Path Smoothing and Adaptive Local Control

Autonomous navigation of autonomous mobile robots (AMRs) in complex environments requires simultaneously addressing the validity of global paths and the optimization of local trajectories. This paper proposes a hierarchical AMR autonomous navigation framework that integrates a smoothed rapidly-exploring random tree with soft actor-critic (SRRT-SAC), comprising an upper-level global motion planning layer and a lower-level local motion control layer. In the planning stage, we enhance the conventional rapid-exploring random tree (RRT) algorithm by applying B-spline interpolation to the globally planned path, thereby improving global path smoothness. In the control stage, the AMR navigation problem is formulated as a Markov decision process (MDP) and solved using deep reinforcement learning (DRL) with the Soft Actor–Critic (SAC) algorithm. We design a composite reward scheme that integrates heuristic functions to effectively mitigate the sparse-reward problem in DRL, and we apply dynamic normalization of the state space to accelerate and stabilize neural network training. Experimental results show that during training, SRRT-SAC achieves an average reward convergence speed more than 36.8% faster than baseline methods such as SAC and TD3. In generalization tests, SRRT-SAC yields notable improvements over conventional DRL approaches across metrics, including path length, task completion time, number of turns, and minimum obstacle clearance; notably, in a 20 m × 20 m environment, it successfully completed navigation tasks that SAC, TD3 and A2C could not.