Core Concepts
The author presents NeuPAN as a real-time, accurate, and environment-invariant robot navigation solution using end-to-end model-based learning. The approach directly maps raw points to distances for collision avoidance.
Abstract
NeuPAN introduces a novel approach to robot navigation by mapping raw points to distances in cluttered environments. The system integrates perception and locomotion through neural networks and mathematical models, demonstrating superior performance in various scenarios.
Key Points:
NeuPAN directly processes high-dimensional raw points for accurate distance calculations.
The system incorporates end-to-end model-based learning for collision avoidance.
DUNE is used to map LiDAR points to multi-frame distance features.
NRMP utilizes learned features for motion planning with physical constraints.
Training of DUNE is conducted using simulated datasets for real-world applications.
Stats
Leveraging a tightly-coupled perception-locomotion framework, NeuPAN outperforms benchmarks in accuracy, efficiency, robustness, and generalization capability.
Experiments demonstrate that NeuPAN operates well in unstructured environments with arbitrary-shape undetectable objects.