One Navigation Model, Four Robot Bodies: LightNav-0 Turns 2,000+ Real Scenes Into Simulation

Liangyuan's newly published general navigation model, LightNav-0, reframes “where should the robot go” from an engineering problem solved with pre-built maps and localization into one that can be trained with data at scale. It arrives alongside two other techniques the company released over the past month: LightParkour for whole-body parkour locomotion and Light REACT for resilient humanoid control.

Data is the first obstacle. Language models inherit an internet-scale corpus; robots have no comparable “robot internet.” If every navigation experience must be collected through teleoperation on real hardware, each new environment or new robot body adds another round of expensive data collection.

LightNav-0's answer is a Real2Sim2Real data engine that converts more than 2,000 internet-sourced real scenes into reusable simulation environments, producing over 4,000 hours of vision, language and action post-training experience. A single scene can yield different goals, routes, viewpoints and tasks, and camera geometry varies during training so the model does not lock onto one fixed robot viewpoint.

One finding stands out: within the current experiments, widening environment coverage improved generalization more reliably than simply adding more trajectories inside the same environment. That hints embodied data scaling may differ from language-model scaling — diversity of worlds, not just volume of data.

Turning vision and language into motion is the second problem. Told to “walk to the right of the sofa,” a model understands the sofa easily; the hard part is locating it, defining “right,” finding a passable route and generating a trajectory. LightNav-0 therefore adds Point CoT, predicting target and traversable points in image space before emitting action tokens, so inference runs as vision-language understanding, then spatial reasoning, then action generation.

In eight public ablations reported by the company, Point CoT lifted average task success by 8.4 percentage points and SPL by 5.7 points. An RVQ action encoder then compresses continuous robot trajectories into three action tokens, letting vision, language, spatial locations and actions share one training framework.

Validation was not confined to one robot or one benchmark. Across 10 simulation settings, LightNav-0 covers instruction following, goal navigation and embodied visual tracking; on real hardware the same model transfers zero-shot to humanoid, quadruped, wheeled and flying robots. The company also released the model, code and technical report.

Read together, the three projects answer one question: how to turn “the robot can do a thing” into “the capability scales.” LightParkour starts from a single human motion seed and grows it into a skill distribution across obstacle heights through reinforcement learning and curriculum learning, then distills multiple experts into one policy; LightNav-0 tackles how experience and generalization scale; Light REACT addresses deployment resilience, inferring body state from recent interaction history after perturbations or hardware damage, with upright behaviour rising from about 42% to about 76% in public experiments.

The write-up also points to adjacent work: Skild AI's S1, Generalist AI's GEN-1.5 and RoboTTT all reduce explicit human adaptation, but their context comes mainly from externally supplied task demonstrations, whereas Light REACT reads the robot's own body state. The open question is whether this pre-training, alignment and deployment loop actually closes — whether real-world deployment data flows back into training, and whether cross-embodiment zero-shot transfer holds in messier open environments.