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

RangeNav: from synthetic pretraining to real-world navigation.

360°LiDAR coverage

183kencoder parameters

1,920parallel RL environments

Zero-shotreal-world transfer

Abstract

Autonomous mobile robots operating in dense crowds and cluttered sites require 360° perception to react to surrounding hazards. 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.

We present RangeNav, a compact (183k parameters) LiDAR range-image encoder pretrained entirely on synthetic data. RangeNav uses anisotropic convolutions and circular padding to handle the extreme aspect ratio and azimuthal seam of range images. 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.

In simulation, adding semantic and BEV supervision increases success rates over a reconstruction-only baseline by an average of 32.03 and 23.73 percentage points (pp) in procedurally generated training and out-of-distribution realistic test environments, respectively. 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. Code and models will be open-sourced.

Method

We generate LiDAR observations from TartanGround, pretrain a compact encoder with three complementary objectives, and freeze it for policy learning. The decoder heads are used only during pretraining.

LiDAR projection, encoder and policy, auxiliary decoder heads, and robot deployments with dynamic obstacles, at a construction site, and in low light. — **System overview.** A panoramic LiDAR scan is encoded for a recurrent navigation policy.

Six depth views projected into a panoramic LiDAR range image. — **Synthetic data.** Six depth views are projected into a LiDAR range image. Noise and masking simulate imperfect observations.

RangeNav training pipeline, from synthetic observations and encoder pretraining to reinforcement-learning navigation. — **Pretraining and policy learning.** Reconstruction, semantics, and bird’s-eye-view prediction supervise the encoder. The frozen encoder then feeds the navigation policy.

Key insight: Geometry, semantic labels, and top-down occupancy teach the encoder which scene information matters for navigation.

Real-World Navigation

The system runs onboard a Unitree B2-W with a 32-beam RoboSense Helios LiDAR and Jetson AGX Orin. Demonstrations cover pedestrians, construction sites, narrow passages, and low-light environments.

Real-world navigation. Zero-shot deployment on a Unitree B2-W in crowds, cluttered sites, and low light.

ARCHE deployment. Outdoor navigation at ARCHE.

Dynamic obstacle avoidance. Navigation around moving obstacles.

Real-world scenes with pedestrians, a fence, and vehicles, alongside LiDAR inputs and decoder predictions. — **Real-world perception.** LiDAR inputs and predictions of range, semantics, and bird’s-eye-view occupancy. Decoder outputs are shown for visualization.

Key insight: The same encoder and policy transfer from simulation without real-world fine-tuning.

Simulation

Training uses procedural mazes, pillars, stairs, and pits. Evaluation adds held-out scenes, starts close to walls, and moving pedestrians. Each environment uses 1,024 fixed start–goal scenarios.

Procedural terrain. Navigation across mazes, pillars, stairs, and pits.

Held-out environments. Navigation in scenes excluded from reinforcement-learning training.

Starting beside a wall. Depth on the left; LiDAR on the right. Success increases from 79.98% to 88.87%.

Moving pedestrians. Depth on the left; LiDAR on the right. Success increases from 77.64% to 79.49%.

Training and evaluation environments including maze, stairs, pits, side wall, Gascola, city, cable, and moving obstacles. — **Evaluation environments.** Procedural terrain, realistic scenes, wall-adjacent starts, and moving obstacles.

Blue LiDAR and orange depth trajectories around dead ends and local clutter. — **Navigation trajectories.** LiDAR (blue) avoids a dead-end detour and escapes local clutter in these representative rollouts; depth is shown in orange.

Key insight: Panoramic LiDAR improves success over front-facing depth in side-wall and dynamic-obstacle scenarios. Depth remains stronger on pillars, stairs, and AbandonedCable.

Results

Adding semantic and BEV supervision improves average success over reconstruction alone by 32.03 percentage points on the base mixture and 23.73 percentage points on held-out TartanGround scenes.

Navigation success rate (%) ↑
Input / pretraining	Base mixture	TartanGround	Side-wall	Dynamic
Depth baseline	95.36	72.95	79.98	77.64
LiDAR · Reconstruction	63.43	49.44	61.52	63.18
LiDAR · + Semantics	88.50	67.43	85.94	78.03
RangeNav · + BEV	95.46	73.17	88.87	79.49

Success rate (%). Base mixture and TartanGround each average four environments; policies are trained for 15,000 iterations.

The 183k-parameter encoder supports 1,920 parallel environments and 1,469 training steps/s on one RTX 4090.

Detailed results and ablations

Per-environment success

Success rate by evaluation environment (%) ↑
Environment	Reconstruction only	RangeNav
Maze	78.81	96.88
Pillars	80.08	98.54
Stairs	68.07	94.34
Pits	26.76	92.09
AbandonedCable	51.46	81.45
Gascola	29.98	51.56
ModernCityDowntown	46.39	86.13
Market	69.92	73.54

Route efficiency

Route efficiency on the procedural base mixture
Metric	Depth	LiDAR
Success weighted by path length (%) ↑	84.53	85.11
Time to goal (s) ↓	16.57	15.19

Aggregate values across maze, pillars, stairs, and pits.

Encoder scalability

Encoder scalability on one RTX 4090
Encoder	Parameters	Max. RL environments	Throughput (steps/s)
Utonia	137M	48	19
Sonata	108M	64	30
LitePT-S	12.7M	96	75
RangeNav	183k	1,920	1,469

Pretraining and architecture ablations

Pretraining source · navigation success (%) ↑
Pretraining	Success
Sonata distillation	85.06
Utonia distillation	80.98
RangeNav objectives	92.38

Selected architecture ablations · success (%) ↑
Variant	Parameters	Success
RangeNav · 3 × 9	183k	92.38
3 × 3 stem	182k	85.96
2× channel width	462k	91.53
RegNet-FPN	5.8M	70.46

The distillation and architecture ablations use a shorter 5,000-iteration RL budget and should not be directly compared with the main results.

Resources

Code and pretrained models are coming soon.