Researchers led by Zezhi Liu have built GraF-VTA, which plans robot trajectories that satisfy tasks written in signal temporal logic, a formal language for timed requirements, using camera images, according to a paper on arXiv. It fuses the structure of the logic with visual features and keeps a compressed memory of earlier frames.
A flow-matching model generates a plan, carries out its first actions and replans from fresh observations. It outperformed two earlier methods in four of five simulation environments. In a closed-loop test with a simple vehicle model it reached 64.06% success, against 12.50% and 27.34% for adapted versions of other methods.



