Researchers tested a mobile robot that navigates using depth estimated from one camera. In a simulated indoor benchmark, the standard setup collided in 63% of 120 episodes because the depth model overestimated the distance to nearby obstacles.
They added a small module that corrects each depth estimate and gives an uncertainty bound. It halved depth error and cut near-range overestimation from 59.8% to 12.0%. Navigation success rose from 37% to 70% and collisions fell to 3%, close to the 2% when the robot is given true depth.
A fixed safety margin set to the same coverage reached only 30% success because it made the robot stop. The authors say the gains held on glass, plain walls and dark areas. The tests were in simulation.
Background
Safe depth
Depth correction module
A small add-on that corrects depth estimates from one camera and adds an uncertainty bound for robot navigation.
