Researchers led by Raul Steinmetz have built ROMUX, which guides a robot to a goal around obstacles using cameras mounted on walls or ceilings, without knowing where those cameras are or how they are set up, according to a paper on arXiv. Such cameras see more of a room and are blocked less often than a robot's own, but earlier methods needed each one calibrated by a specialist, again after every move.
A cross-view transformer trained on large amounts of simulated data turns the camera images into a map of the obstacles around the robot, with real distances, and a small policy trained by reinforcement learning drives the robot on that map towards its goal.
In simulation, with cameras placed at random in each episode, ROMUX was as successful as a policy navigating with a camera on the robot. It works with one camera or several that each see part of the room, and it moved to a real robot, with cameras in different positions and rooms laid out in different ways, without extra training.



