Low-Latency LiDAR Perception Based on Point Transformer-Voxel Model for Underground Robot Autonomous Obstacle Avoidance

Abstract

Underground corridors have repetitive geometry, weak illumination, airborne dust and narrow clearances, and are thus unreliable for image-only obstacle perception. This paper introduces a geometry-first point transformer-voxel network that voxelizes streaming LiDAR data, uses neighborhood attention to recover local boundary details, and combines calibrated collision probabilities with a speed-aware local controller. Tests on 18.6 km of tunnel runs and 42,000 labelled frames show a mean intersection-over-union of 74.8%, a small-obstacle recall of 92.6% and an end-to-end latency of 31.7ms on an embedded GPU. At a speed of 1.2 m/s, the robot has completed 97.8% of the routes without collision and outperformed the strongest compact baseline by 4.9%, reducing perception energy by 18.3%. Based on the above evidence, sparse voxel context and point-level refinement are complementary in the context of latency and memory constraints for underground inspection.