LightNav-0: A Shared Navigation Model for Different Robots

LightNav-0 is Light Origins’ attempt to make robot navigation less dependent on collecting a new set of teleoperated demonstrations for every task and robot. Built on Qwen3-VL-4B, it uses one model for following instructions, finding objects described in open vocabulary, and tracking visual targets. Its approach pairs a shared navigation interface with simulated experience derived from real-world scenes. 128

How one model handles different navigation tasks

Rather than adding a separate prediction head for each task, LightNav-0 represents navigation through a shared token interface. Dual-channel pointing tokens express where the model intends to go in a form designed to generalize across tasks, scenes, and robot bodies. A residual vector-quantized action tokenizer then represents the more precise, embodiment-specific trajectory. Temporally aware visual-history compression helps the model use observations from more than one moment. 15

That separation is central to the cross-robot goal: the model can share a representation of spatial intent without assuming every robot executes the same movements. It is a design for transfer, however, not a guarantee that an unseen robot will need no adaptation. 15

Where the training experience comes from

Light Origins calls its data pipeline Real2Sim2Real. The company says it turned more than 2,000 internet-sourced real-world scenes into reusable simulated environments and generated more than 4,000 hours of vision-language-action navigation experience. This lets the team generate varied training situations in simulation instead of relying solely on separately teleoperated demonstrations. Those figures describe the company’s reported data pipeline, not a measured reduction in real-world data-collection costs. 8

The company describes three training stages: embodied-reasoning mid-training, embodied supervised fine-tuning, and online reinforcement learning. Together, they are intended to adapt the pretrained vision-language backbone to navigation, teach the shared token-based behavior, and improve the policy through interaction. 18

What the results—and the release—show

Light Origins reports state-of-the-art monocular success rates across 10 public navigation simulation settings spanning instruction following, object-goal navigation, and visual tracking. It also reports zero-shot generalization across robot embodiments and real-world scenes. These are the team’s reported evaluations; the cited summaries do not establish a universal success rate for unfamiliar robots or prove that teleoperation can be dispensed with. 7813

The release includes model weights, code, a technical report, and INSIGHT-Bench evaluation materials, with the release described as Apache 2.0 licensed. That gives others materials to inspect and test the approach. The practical proposition remains narrower than fully autonomous data collection: reusable simulated scenes and a shared policy may reduce robot-by-robot demonstration needs, while real-world performance still has to be verified for the robot and setting at hand. 2510