MOSU2 raises outdoor robot navigation success from 61.8% to 94.2% in real-world tests

Researchers led by Jing Liang have presented MOSU2, a long-range navigation system for robots that travel through public outdoor spaces, in a paper posted on arXiv.

Such robots must stay on terrain that suits them and decide when to go, give way or go around people on foot, cyclists, vehicles, traffic lights and temporary closures, the authors write. Vision-language models can read these cues but reason over images without depth or motion, and supplying that at every step is too slow. In MOSU2, a first stage of contextual reasoning decides when more evidence is needed and what to fetch, and only unresolved interactions trigger a second stage that adds geometric and motion information.

Building on its predecessor MOSU, it follows GPS-based routes and filters candidate paths by traversability. In real-world tests with stop gestures, construction sites, crossings and roads shared with vehicles, cyclists and pedestrians, its average success rate went up from 61.8% to 94.2%, and its accuracy in spotting conditions from 72.2% to 91.4%, while taking about 2.2 times as long on average to decide as the single-image version.