A hierarchical motion planning framework optimizing probabilistic roadmap, pure pursuit, and deep reinforcement learning for non-holonomic automated guided vehicles

The motion planning is a critical component of autonomous navigation, requiring the vehicle to reach a target location safely. Traditional navigation approaches for four-wheel differential drive automated guided vehicles (AGVs) often suffer from limitations such as inadequate handling of non-holonomic constraints, sharp turning, or susceptibility to becoming stuck in complex environments. This study presents a hierarchical motion planning framework that optimizing Probabilistic Roadmap (PRM), Pure Pursuit (PP), and Reinforcement Learning (RL) based Deep Deterministic Policy Gradient (DDPG) to address the limitations of traditional approaches in non-holonomic vehicle constraints. The proposed PRM–PP–RLDDPG framework was evaluated in three simulated environments: a Gazebo map, a corridor–laboratory map, and an office layout. The experimental results demonstrate that the proposed framework consistently outperforms the traditional PRM–PP–VFH approach across all evaluated simulated environments. The proposed framework reduces the final position percentage error by approximately 41%, indicating improved terminal accuracy indicates higher mean final position accuracy of 99.12%, compared to 98.51% for the traditional method across all environments. These results highlight the effectiveness of integrating learning-based execution-level optimization within a global–local planning architecture for reliable and stable navigation of non-holonomic AGVs in complex indoor environments.