Efficient Dynamic Window Approach With Neural Multi-Trajectory Proposer

We propose a control method for mobile robots that improves navigation in complex environments by integrating the dynamic window approach (DWA) and reinforcement learning technique. The growing presence of mobile robots in many fields highlights the increasing need for effective navigation that can cope with narrow paths and irregularly positioned obstacles. Traditional DWA and its improved methods are sensitive to the parameter values and the length of the prediction horizon. In this study, we propose the Neural Multi-trajectory Proposer (NMP), which generates prediction trajectories on the basis of reinforcement learning. Instead of prediction assuming the control input is constant, which is done in DWA, the proposed method with NMP improves the control efficiency while maintaining safety. The efficiency of the proposed method was validated in various scenarios, demonstrating superior performance compared with conventional DWA by a factor of at most 2.3 in average moving speed. Moreover, it is shown that the degradation in efficiency for longer prediction horizons is suppressed to 3.9% compared with 32% for the conventional DWA. As a feasibility study, the proposed method was evaluated in a real-world experiment, achieving over twice the average moving speed of conventional DWA.