Researchers led by Yuhang Huang have built TARS, a framework that tests robots which use a large language model to understand people and a smaller model to turn that understanding into movement, according to a paper on arXiv. Meaning can get lost between what the large model understands, the commands the action model produces and what the robot then does, so a robot may act even when nobody asked it to.

TARS treats the robot as a black box, explores how commands are reasoned about and carried out, and automatically finds and confirms unintended physical behaviour.

Across four commercial robots in 44 experimental settings and 133,400 runs, it found widespread mismatches between the models and reached up to 94% coverage of verified actions, the authors report.

Source: arXiv Robotics (cs.RO)

Background

Yuhang Huang

Robotics researcher

Robotics researcher.