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

• DOI: 10.48550/arxiv.2403.06828
• Publication Date: 2024-03-11
• Authors: Ruihua Han; Shuai Wang; Shuaijun Wang; Zeqing Zhang; Jianjun Chen; Shijie Lin; Chengyang Li; Cheng‐Zhong Xu; Yonina C. Eldar; Hao Qi; Jia Pan
• Keywords: End-to-end principle

Navigating a nonholonomic robot in cluttered environment requires extremely accurate perception and locomotion for collision avoidance. This paper presents NeuPAN: real-time, highly-accurate, map-free, robot-agnostic, environment-invariant navigation solution. Leveraging tightly-coupled perception-locomotion framework, NeuPAN has two key innovations compared to existing approaches: 1) it directly maps raw points learned multi-frame distance space, avoiding error propagation from control; 2) is interpretable an end-to-end model-based learning perspective, enabling provable convergence. The crux of solve high-dimensional mathematical model with various point-level constraints using the plug-and-play (PnP) proximal alternating-minimization network (PAN) neurons loop. allows generate end-to-end, physically-interpretable motions point clouds, which seamlessly integrates data- knowledge-engines, where its parameters are adjusted via back propagation. We evaluate on car-like robot, wheel-legged passenger autonomous vehicle, both simulated real-world environments. Experiments demonstrate that outperforms benchmarks, terms accuracy, efficiency, robustness, generalization capability across environments, including sandbox, office, corridor, parking lot. show works well unstructured environments arbitrary-shape undetectable objects, making impassable ways passable.