Smoother Action Chunking Flow Policy via Prior-Corrected Orthogonal Trust-Region Guidance
Researchers have developed a new guidance method called POTR (Prior-Corrected Orthogonal Trust-Region) to improve the smoothness of action chunking in flow-matching robot policies. This method addresses discontinuities at chunk boundaries by incorporating a data-prior scale for stronger intermediate-time correction and by constraining the guidance vector's perpendicular component within a trust region. Experiments on the LIBERO benchmark demonstrated that POTR enhances success rates and significantly reduces undesirable action transitions like discontinuity, acceleration, and jerk compared to existing RTC guidance. AI
IMPACT Enhances robot control by reducing jerky movements and improving policy smoothness.