Introducing LightNav-0, our first general-purpose navigation brain. Open-sourced starting today. Trained entirely in simulation, so it scales. See scalable real2sim2real transfer across robots, tasks, and scenes.

One brain. Any robot. LightNav-0 generalizes across embodiments. Here, we simultaneously deploy the same model to humanoid, quadruped, aerial, and wheeled robots.

One brain. Any task. LightNav-0 exhibits a deep understanding of objects, directions, and spatial relationships, enabling it to follow a wide range of language instructions for open-world navigation tasks.

One brain. Any world. No game-specific navigation data. No domain-specific fine-tuning. The same LightNav-0 model also achieves zero-shot navigation in digital worlds such as Minecraft.

One brain. Ten benchmarks. SOTA. LightNav-0 achieves state-of-the-art monocular success rates across 10 public simulation settings, spanning instruction following, object-goal navigation, and embodied visual tracking.

Behind LightNav-0 is our scalable Real2Sim2Real data engine. It turns 2,000+ internet-sourced real-world scenes into 4,000+ hours of diverse navigation experience in simulation. And it continues to scale.

Read our full technical blog for details: At Light Origins, we believe the path toward physical AGI runs through three scaling paradigms: scalable pre-training, scalable alignment, and scalable deployment. With LightNav-0, our first step toward scalable alignment through Real2Sim2Real, we demonstrate that alignment at scale is key to achieving zero-shot generalization in the physical world. Code: Model:

nice work!

Thank you!

nav is the one task where this works though. walking through a static reconstruction is basically geometry, the physics barely matter. the second you touch something sim stops being free. thats why nav gets to eat internet scenes and manipulation doesnt

Greatness 🫡

So cool. Amazing you open sourced this! Curious how many different data streams it supports