Researchers led by Lorenzo Paiola have built a way to make generative robot policies keep moving a task toward its goal, according to a paper on arXiv. Such policies can produce motions that look right but stall or undo earlier progress. Daniela Rus and Cosimo Della Santina are among the authors.
Before acting, the method runs candidate action chunks of a flow-matching policy in parallel physics simulations. It keeps those that stay within a progress funnel: the task may slip back for a moment, but each chunk must gain overall. An evolution strategy called CMA-ES adjusts the random starting points of the policy, kept within bands that preserve the motions it learned.
This raised success on the PushT task from 31.3% to 88.5%, and on stacking three cubes from 77.0% to 86.0%. With the bands, 11.3% of the motions fell outside what the policy had learned, close to the rate in held-out demonstrations, against 34.1% for a search without them.
Background
More stories about Lorenzo Paiola
Task Convergent Flow Matching
Robot policy method
Method that checks the motions of flow-matching robot policies in physics simulations before they are carried out.



