Researchers led by Takahiro Miki, with Marco Hutter among the authors, have built ROAM, a system that lets legged robots navigate for kilometres in alpine and urban settings, according to a paper on arXiv. The authors say it is their version of an article published in Science Robotics.
A vision-language model trained on internet data assesses risks, reads signs and labels parts of the scene with navigation costs. A segmentation module turns the labels into a live risk map for two policies trained in simulation, one for walking and one for local navigation. In real-world tests the robot finished kilometre-long hikes with little help and picked safe places to step. The authors say the reliability and spatial reasoning of vision-language models are still a problem.



