LightNav-0 paper: compact VLM tops all 10 navigation sims

A team led by Shaoan Wang has posted LightNav-0 on arXiv, a compact vision-language model the authors say achieves "state-of-the-art monocular success rates across all 10 public navigation simulation settings."

The framing in the abstract is that modern VLMs "already encode spatial priors for visual grounding, spatial reasoning, and pointing, but these capabilities are rarely elicited directly for robot control." Their answer is a unified token interface: "dual-channel pointing expresses task-, scene-, and embodiment-agnostic spatial intent, while a residual vector-quantized action tokenizer maps this intent to precise, embodiment-specific trajectories."

Training draws on what the authors describe as "2K+ scenes and 4K+ hours of embodied navigation data," run through supervised fine-tuning and reinforcement learning on top of a mid-training reasoning stage. A companion checkpoint used to initialize the model, LightNav-ER, is reported to attain "the highest complete-set average across 8 embodied-reasoning benchmarks." A single model handles instruction following, open-vocabulary object navigation, and visual tracking.

The abstract publishes no per-benchmark success numbers, no parameter count, and no named baselines it beats. Real-world tests are described as showing "zero-shot generalization across robot embodiments, diverse scenes, and static and dynamic targets," without quantified figures.

It lands amid a busy quarter of robotics coverage on our tracker, which logged 152 robotics stories over the last 90 days.