Researchers led by Yang Li have built RPSI-Nav, a framework that lets a robot search for several objects in places it has never seen, without training, and get better from one search to the next, according to a paper on arXiv. Earlier methods also improve their navigation rules by reflecting on past paths, but the way they diagnose problems and write fixes is designed by people and never changes.

RPSI-Nav diagnoses its search behaviour, asks for extra views when its observations cannot tell competing causes apart, and writes targeted changes to its navigation program. Feedback from later searches is traced back to the rules that produced those changes, so the framework also changes how it diagnoses and revises. The robot carries accepted program updates and its spatial memory from one goal to the next.

The authors say benchmark experiments and qualitative tests on a real robot confirm that the method works.

Source: arXiv Robotics (cs.RO)

Background

Yang Li

Robotics researcher

Robotics researcher.