ScaleNav hits 94.9% on HabitatGS-LN using internet video pretraining

ScaleNav, a navigation policy whose first training stage runs on internet egocentric video rather than teleoperation or simulation, reports an 86.5% success rate on a single-view adaptation of Short-Horizon OVON and 94.9% on the authors' own HabitatGS-LN benchmark, according to a preprint on arXiv. The scope is narrower than the headline suggests: the authors define the task as one where "a robot must navigate to a text-specified target visible in its initial observation and stop at the described location."

The authors frame internet video as a way around a familiar data bottleneck. "End-to-end navigation policies require diverse training data to generalize across scenes, yet real-robot data is costly to collect and simulated data leaves a sim-to-real gap," they write, arguing that "Internet egocentric videos offer a scalable source of real-world navigation experience." ScaleNav is trained in three stages: instruction-conditioned video pretraining to learn how observations evolve as a camera approaches an instructed target, an action-alignment pass on action-labeled simulated data to turn the model into one that emits executable action chunks, and a joint fine-tune.

The paper then takes the model off the benchmark and onto hardware. ScaleNav was "deployed and evaluated on a real quadruped robot, with results supporting the model's ability to generalize to real-world navigation," the abstract states. The abstract gives no quadruped success rates, no per-scene breakdown on the two benchmarks, and no description of the video corpus used in stage one.