Embodied navigation has long been a challenge in AI, requiring agents to translate diverse goals and visual observations into actions across different tasks, environments, and robot bodies. Traditional systems often rely on task- or embodiment-specific components, which fragment perception, reasoning, and action, limiting their generalization capabilities.
In a new paper from Hugging Face, researchers present LightNav-0, a compact generalist model that elicits the spatial intelligence of a pretrained vision-language model (VLM) and aligns it with navigation—without task-specific prediction heads. The model uses a unified token interface: dual-channel pointing expresses task-, scene-, and embodiment-agnostic spatial intent, while a residual vector-quantized action tokenizer maps that intent into precise, embodiment-specific trajectories.
LightNav-0 is trained on a corpus spanning over 2,000 scenes and more than 4,000 hours of embodied navigation data. The training pipeline includes temporally aware visual history compression, ER mid-training, supervised fine-tuning, and reinforcement learning. This formulation enables the model to handle instruction following, open-vocabulary object navigation, and visual tracking within a single architecture.
The results are striking: the LightNav-ER checkpoint, used to initialize LightNav-0, achieves the highest complete-set average across eight embodied-reasoning benchmarks. LightNav-0 itself attains state-of-the-art monocular success rates across all ten public navigation simulation settings. Real-world evaluations further demonstrate zero-shot generalization across different robot embodiments, diverse scenes, and both static and dynamic targets.
These findings establish compact VLMs as a unified and transferable backbone for generalist embodied navigation, potentially simplifying the design of future robotic systems.