Shape-Aware End-to-End Robot Navigation with 3D Point Clouds via Sparse Convolution and Batch Normalization for Enhanced Training Efficiency

초록

End-to-end navigation methods have emerged as promising alternatives to traditional approaches in mobile robot navigation, thanks to their ability to handle obstacles of arbitrary shapes and unbounded categories. However, these methods often require a large number of training steps due to the complexity of interpreting high-dimensional sensor inputs. To address this limitation, we propose a shape-aware end-to-end navigation system equipped with a neural encoder specifically designed for 3D point clouds. By incorporating batch normalization and sparse convolution, the encoder enhances both sample efficiency and training speed. Our system achieves superior navigation performance compared to existing methods, while also significantly accelerating the training process. © 2025 ICROS.

키워드