CPNav calibrates AI model scores to help robots search for objects in unseen places

Researchers led by Qizhao Chen have presented CPNav, a framework that lets a robot search for a target object in a place it has not seen before without building a map, in a paper posted on arXiv.

From colour and depth images, CPNav works out where the robot can move, generates motion options that avoid collisions and asks two vision-language models to spot the goal and score each action. Because such models are often overconfident, the scores are calibrated offline with conformal prediction, which gives statistical guarantees at a risk level the user chooses. A record of earlier choices lets the robot backtrack.

In the AI Habitat simulator, CPNav gave relative gains of 6.2% in success rate and 11.4% in success weighted by path length on the HM3D dataset compared with the uncalibrated baseline, and 12.4% and 8.8% on MP3D. On a modified Hiwonder MentorPi robot the gains were 24.9% and 15.2%.