Robotics / news brief
ROAM combines vision-language reasoning with learned legged-robot control
- Event
- Published

A Science Robotics paper presents ROAM, which turns vision-language model judgments into risk maps for learned locomotion and navigation policies. The authors report kilometer-scale alpine and urban trials with minimal intervention, while warning that VLM reliability and spatial reasoning remain incomplete.
Why it matters
ROAM shows a practical hierarchy in which a foundation model interprets terrain while faster learned controllers retain responsibility for physical motion.
Sources
- Bridging vision-language reasoning and learned control for legged navigation in the wild (external site)Science Robotics · primary source ·
Single source: Peer-reviewed primary publication with public code; no independent editorial coverage was verified in the collection window.