A preprint describes WayTrace, which links web-map walking directions to robot motion through language. A vision-language model reads the directions, a satellite crop and camera images and issues one command at a time, such as crossing the road to the left. An onboard model draws the path in the image and a reinforcement-learning policy follows it.
In simulated long missions in a Zurich old-town model it kept to crosswalks and sidewalks, with about ten times fewer violations than point-goal baselines when GPS is noisy. On a wheeled-legged robot it ran four real missions in Zurich between 248 m and 1.37 km at about 1 m/s.
Three city missions (248 to 426 m) needed no operator intervention. On the 1.37 km mission the operator replaced the command twice, after turns at a campus entrance and on an open plaza, giving 97.2% of the distance without intervention; the authors say this shows a limit when commands refer to features no longer visible. At a tram crossing the operator also paused the robot, because it cannot yet react to trams. The results are the authors' own and a preprint.
Limits
Preprint; authors' own tests, a small number of real missions; the orchestrator model runs on a remote server.