RangeNav: Task-Aligned LiDAR Range-Image Pretraining for Robot Navigation

Key points

• Autonomous mobile robots operating in dense crowds and cluttered sites require $360^\circ$ perception to react to surrounding hazards.1
• While panoramic LiDAR offers full-surround coverage, learning representations within reinforcement learning (RL) is sample-inefficient under sparse rewards, and existing 3D pretrained point-cloud encoders are too memory-heavy for massively parallel RL simulation.1
• We pretrain the network using a navigation-aligned objective combining range reconstruction, semantic segmentation, and bird's-eye-view (BEV) prediction, then freeze it for policy learning.1
• Zero-shot real-world deployments on a Unitree B2-W quadruped demonstrate robust navigation through dense crowds, low-light settings, and construction zones where limited-field-of-view depth policies struggle.1

Sources (1)

• [1]RangeNav: Task-Aligned LiDAR Range-Image Pretraining for Robot Navigation
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 21, 03:06 PM

  We present RangeNav, a compact (183k parameters) LiDAR range-image encoder pretrained entirely on synthetic data.

  Autonomous mobile robots operating in dense crowds and cluttered sites require $360^\circ$ perception to react to surrounding hazards.

Extractive summary: sentences quoted from the sources.