LightNav-0 introduces one RGB-and-language navigation model across tasks and robot bodies
On September1,2026, Light Origins introduced LightNav-0: one forward-RGB/language model for instruction following, object navigation and visual tracking. Dual-channel pointing and RVQ tokens yield ten planar waypoints for separate robot controllers. The study reports a shared checkpoint across tasks and four tested physical embodiments; physical transfer demonstrations are qualitative. LightNav-ER is the reasoning initializer, not a separately verified open checkpoint. Native navigation weights are currently verified public, but their first public day is not inferred from this announcement.
Previously reported
Pretrained VLMs supply visual/spatial priors, while navigation is often connected through task- or embodiment-specific components. The report’s starting point is the general Qwen3-VL backbone, before its ER mid-training and navigation SFT/RL. This is a design baseline, not a claim that every earlier navigation model needs depth or cannot generalize.
This announcement
On September1,2026, Light Origins introduced LightNav-0: one forward-RGB/language model for instruction following, object navigation and visual tracking. Dual-channel pointing and RVQ tokens yield ten planar waypoints for separate robot controllers. The study reports a shared checkpoint across tasks and four tested physical embodiments; physical transfer demonstrations are qualitative. LightNav-ER is the reasoning initializer, not a separately verified open checkpoint. Native navigation weights are currently verified public, but their first public day is not inferred from this announcement.
Featured as a major new model/interface offering a reusable single-RGB option across navigation tasks and tested robot bodies, relevant to sensor/controller integration choices. Not automatically featured by admission scores95/89 or by small benchmark gains. Historical model announcement only; paper revisions and documentation backfill are not separate featured news.