Task Convergent Flow Matching for Generative Robotics
Refining robot actions before execution raised PushT success from 31.3% to 88.5% in the reported tests.
The method checks candidate action chunks through parallel physics simulations, seeking motions that deliver net task progress while limiting temporary setbacks. It also raised three-cube stacking success from 77.0% to 86.0%. On PushT, a learned progress constraint reduced the average maximum loss of achieved progress by 19% compared with checking chunk endpoints alone, at a similar success rate. Constraining the search kept 11.3% of executed motion windows outside the learned distribution, versus 34.1% with unconstrained search.
ArXiv · AI/CL/LG's note
The method checks candidate action chunks through parallel physics simulations, seeking motions that deliver net task progress while limiting temporary setbacks. It also raised three-cube stacking success from 77.0% to 86.0%. On PushT, a learned progress constraint reduced the average maximum loss of achieved progress by 19% compared with checking chunk endpoints alone, at a similar success rate. Constraining the search kept 11.3% of executed motion windows outside the learned distribution, versus 34.1% with unconstrained search.
ArXiv · AI/CL/LG's note
score 5