ScaleNav uses internet video training to guide robot navigation

Researchers report that ScaleNav, a robot navigation system pretrained on internet videos, outperformed evaluated baselines on two navigation benchmarks and was tested on a real quadruped robot. The preprint reports success rates of 86.5% on a single-view adaptation of Short-Horizon OVON and 94.9% on the researchers' HabitatGS-LN benchmark. The system learns from first-person videos how a view changes when approaching a target described in text. It then learns executable actions from labeled simulated data before joint fine-tuning. Its task is limited to reaching and stopping at a text-specified target visible in the robot's initial observation; the reported benchmark percentages do not describe its real-robot performance.