NeuPAN: Direct Point Robot Navigation with End-to-End Model-based Learning

NeuPAN is a novel real-time robot motion planner that employs an end-to-end model-based learning approach, directly mapping raw point cloud data to a latent distance feature space for collision-free motion generation. Key innovations include a tightly coupled perception-to-control framework and a plug-and-play proximal alternating-minimization network that incorporates point-level constraints, allowing for interpretable and efficient motion generation. NeuPAN has been evaluated on various robotic platforms in simulated and real-world scenarios, demonstrating superior accuracy, efficiency, and robustness compared to existing methods, making it highly relevant for practitioners working on navigation in complex environments.