Hybrid RL-DWA boosts 3D robot navigation with near-perfect path completion

A new paper accepted to IEEE/ASME AIM 2026 introduces 3D RL-DWA, a hybrid framework that fuses Reinforcement Learning (RL) with the Dynamic Window Approach (DWA) for goal-directed local navigation of high-degree-of-freedom robotic systems. The key innovation is the use of sparse point cloud data to simultaneously control both the motion and the deformable shape of a microrobot, allowing it to maximize occupied volume while navigating through tight, constrained environments. The researchers evaluated their method in a simulated vascular network, conducting 1080 trials across training and unseen test scenarios.

The results show that 3D RL-DWA consistently achieves high deformation and near-perfect path completion during training, and maintains robust performance when faced with novel environments. Compared to pure RL and model-based planners, the hybrid approach significantly enhances both deformation and navigation capabilities. This work highlights the potential of combining RL's adaptive decision-making with DWA's real-time path optimization for efficient 3D navigation under sparse sensory conditions, with direct applications to medical microrobots operating in the human vasculature.