Robot control preprint reports up to 2.3x faster policy inference

Researchers report that a training-free method called PICES cuts robot policy inference time by up to 2.3x on Jetson Orin and Thor while improving task success relative to fixed-horizon baselines. The arXiv preprint describes tests in simulation and real-world manipulation tasks. Models that translate images and language into actions ordinarily execute a fixed number of predicted actions before calculating again. PICES compares candidate action sequences with slightly perturbed counterparts, then selects both a sequence and how many actions to execute before recalculating. The reported speedup measures policy inference time rather than total task completion time.