Learning Native Continuation for Action Chunking Flow Policies

Researchers have developed a new method called 'Legato' to improve Vision Language Action (VLA) models, addressing discontinuities in real-time action execution as announced in their paper on February 13, 2026. The research team, whose names are not specified in the abstract, identified that while action chunking enables VLA models to operate in real-time, the naive approach often creates abrupt transitions at chunk boundaries, resulting in unnatural movements. The paper introduces Legato as a training-time continuation method specifically designed for action-chunked flow-based VLA policies to create smoother, more natural trajectories. The researchers highlight that existing solutions like Real-Time Chunking (RTC) alleviate discontinuity issues but operate externally to the policy, causing problematic multimodal switching and intrinsically non-smooth paths. By embedding continuation directly into the training process, Legato represents a fundamental advancement in how AI systems execute continuous physical or digital actions.