Event

RangeNav: A LiDAR Navigation Encoder Pretrained Purely on Synthetic Data

1 reports · 1 sources

Samuel Foo Enze and the author team proposed and released RangeNav, a LiDAR range-image encoder with only 183,000 parameters, pretrained entirely on synthetic data, using anisotropic convolutions and circular padding to handle the extreme aspect ratio and azimuthal seam of range images. The authors report that adding semantic and BEV supervision in simulation improves success rates by an average of 32.03 and 23.73 percentage points, respectively, over a reconstruction-only baseline in procedurally generated training environments and out-of-distribution photorealistic test environments; in zero-shot real-world deployment on a Unitree B2-W quadruped robot, the method achieved robust navigation in dense crowds, low light, and construction areas. The code and models will be open-sourced, and the paper has been submitted to ICRA 2027 and is under review.