Robotics preprint reports policy corrections with 3.8 ms overhead

Researchers report a method that corrected robot manipulation actions with 3.8 ms of mean overhead per control step across four simulated tasks. Their preprint says it achieved the highest success rate among the evaluated baselines. The method learns offline how small action changes affect task geometry, then calculates corrections using one model evaluation and one constrained optimization step. Policy and model weights remain frozen during operation. The reported overhead excludes base-policy sampling, and the evaluation covers simulated manipulation rather than hardware experiments.