Researchers led by Zhengtu Liang have built UniVEP, a training method for flow-matching robot policies that learns two prediction targets together, according to a paper on arXiv. Predicting the final action holds up better while the action is still noisy, while predicting its velocity gives finer corrections once it is nearly clean, the authors found.
UniVEP starts with endpoint recovery and switches to velocity refinement as the noise drops, with one network run per step. On the RoboTwin benchmark it reached 61.36% success with a single StarVLA run, above the best single-target baseline at 16 runs, which reached 59.96%. With LiLa-WAM, two runs beat the best 16-step baseline and cut inference time by 77.1%.



