Flow-matching robot policies can need separate step counts for accuracy and stability

Flow-matching robot policies usually run a fixed number of Euler steps chosen offline, according to a paper on arXiv. Mahindra Rautela asks whether that count, even when it tracks the intended noise-to-action path well, can still be too coarse where the dynamics around it contract strongly.

A two-coordinate exact model and a controlled augmentation of six trained flow models show that accuracy along the trajectory and local Euler stability can demand different step counts even when the learned transport stays the same.

Studying a frozen pi_0.5 checkpoint on LIBERO tasks without changing its weights, the author finds strongly contracting directions in the Jacobian across 30 policy inputs from 10 tasks.