Flow-matching vision-language-action (VLA) models such as π0 and π0.5 typically generate each action chunk by integrating a learned velocity field over about ten sequential evaluations. That makes action generation the dominant source of latency and caps how fast a robot can react. Previous attempts to fix this either compressed individual evaluations while keeping the sequential solver, or imported one-step objectives from image generation that require training from scratch, an EMA teacher, or an external critic.

SnapFlow instead adapts a pretrained flow-matching VLA to one-step generation through self-distillation. It mixes standard flow-matching supervision with shortcut samples whose target is the model's own two-step Euler endpoint under stop-gradient, and adds a single zero-initialized destination-time projection. This lets one action expert serve both objectives while preserving the pretrained predictor at initialization, with no separate teacher network or critic.

In experiments, the one-NFE SnapFlow checkpoint reached 98.75% LIBERO success on π0.5 versus 97.75% for the teacher, and on π0 it surpassed the teacher by 10.5 percentage points. The authors also report lower mean and tail action error on held-out data, about 3.3x lower batch-one model latency, and successful transfer to shelf retrieval on a dual-arm humanoid.