Legged Robot Hikes Swiss Alps Using AI 'Common Sense'
Zurich, Switzerland – A legged robot hiked kilometers of alpine and urban trails on its own using a system that gives it "common sense" from an artificial intelligence model, researchers reported on Wednesday, November 18.
In the study, which was released with accompanying video, Takahiro Miki and colleagues at ETH Zurich present ROAM, short for Robust Open-world Autonomous Mobility, which combines a vision-language model, or VLM, with walking and route-planning skills the robot learned through reinforcement learning in simulation. A module turns the model's labels and reasoning about its surroundings into a real-time risk map that the robot's control policies use to decide where to go.
The team tested the system in 2024 on a 10.1-kilometer hike near Zermatt and a 2.38-kilometer loop around Seealpsee in the Swiss Alps. When its path led through a patch of bushes, ROAM found an alternate route, while the same robot relying only on a basic geometric policy got stuck in the bushes, the researchers said.
"For the current generation of foundation models, abstracting the VLM's role to a high-level semantic advisor is a more effective strategy for navigation than tasking it with generating concrete, low-level robot actions, which requires spatial reasoning," the authors wrote.
Clean HD download
Register a media account or sign in to buy & download this story
Sign in to download Apply for media access