Event

Coupling Between Execution Horizon and Sampling Depth in Flow-Matching VLA

1 reports · 1 sources

Jiabing Yang and other authors published a paper revealing that in flow-matching vision-language-action (VLA) models, the numerical solver depth and the execution horizon are coupled, and this coupling manifests as a systematic sign reversal: coarse-grained numerical integration reduces accuracy under fast replanning, but instead improves task performance when open-loop execution is extended. This conclusion was validated on LIBERO, RoboTwin 2.0, and real Franka robotic arm manipulation, covering representative VLA backbone models such as π0, π0.5, and GR00T; the authors state that the method requires no training at all and no modification of the model architecture.