Researchers led by Shunyang Huang have presented ERA, a navigation system that lets a robot search for objects using depth predicted from ordinary camera images instead of a depth sensor, and that needs no training, according to a paper on arXiv.

Predicted depth contains errors that build up in the map, can remove places still to explore and end the search too early. ERA's approach, which the team calls Move with Error, uses movement to get new views that correct the map: its occupancy map can be revised, it keeps exploring from reachable viewpoints when no frontier is left, and it checks the target from several views before stopping.

The researchers say ERA beats other training-free methods without depth sensors on three benchmarks, and that its map and exploration parts raised the success rate of another method, VLFM, from 10.8% to 61.7% with the same predicted depth. They also ran ERA on a real robot.

Source: arXiv Robotics (cs.RO)

Background

Shunyang Huang

Robotics researcher

Researcher at Shanghai Jiao Tong University.