Skip to main content
Back to timeline
arXivSource publication:

Euler step counts for flow-matching policies need more than endpoint error: a frozen π0.5 checkpoint shows no sampled stability violation across 30 inputs

Synopsis

Flow-matching robot policies usually fix their Euler step count from offline endpoint-accuracy sweeps, yet each policy input plus noise draw traces a different noise-to-action path, so trajectory accuracy and local stability can demand different step counts. An exact two-coordinate system and six controlled trained flow models show the two requirements can separate while the learned transport stays unchanged. On a frozen π0.5-LIBERO checkpoint, no sampled contracting mode violated the Euler non-expansion condition across 30 policy inputs; five paired inputs needed 4–9 Euler steps at 5% terminal error and 87–112 at 1%.

Source-provided article image: One Trajectory, Two Requirements: Accuracy and Local Stability in Euler Sampling for Flow Matching Policies
arXiv

Interpretation

A flow-matching policy's Euler step budget faces two numerical requirements at once: endpoint accuracy along the nominal trajectory, and local stability of the Euler update against perturbations near that trajectory. The prior practice swept several step counts on a fixed evaluation set and chose one giving acceptable endpoint accuracy as a global budget. An exact two-coordinate system checked the joint step theorem over 1000 initial conditions, with four parameter pairs giving easy, accuracy-limited, stability-limited and jointly limited regimes; in the low-acceleration, strong-contraction case the first coordinate met the accuracy tolerance in one step while Euler needed at least 25 steps to avoid amplifying the contracting direction.

In trained flow models, the local contraction strength can be changed without altering the learned transport, which isolates the stability requirement. Three independent checkpoints were trained for each of two target distributions, then each received one contracting coordinate that does not feed back into the original velocity field. With the added contraction rates, all six checkpoints landed exactly on stability floors of 5, 13 and 25 steps; at the 25-step floor the full endpoint error met the 5% tolerance at 18 steps for Eight-Gaussians and about 22 steps for the bistable Langevin target, so endpoint error can miss a separate stability requirement.

On the frozen π0.5-LIBERO checkpoint, a grid-aware criterion found no sampled contracting mode violating the Euler non-expansion condition. Accuracy and stability had not previously been separated on the native action-flow field of a pretrained robot policy. Across 30 early-rollout policy inputs from 10 LIBERO-10 tasks sharing one noise seed, the sampled stability measure ranged from 0.5088 to 0.5252 with a mean near 0.513, none crossing the sufficient-condition boundary of 1; the full 320-dimensional task-00 Jacobian had 319 positive-real-part eigenvalues near the endpoint, yet the maximum amplification factor on the corresponding Euler grid was 0.506.

On these inputs, the extra steps demanded by tighter numerical tolerances come from trajectory accuracy rather than a sampled stability violation. Five paired inputs carry both accuracy sweeps and stability measures. Required Euler steps rose from 4–9 at 5% terminal error to 27–39 at 2% and 87–112 at 1%, while the corresponding sampled stability measures stayed tightly clustered between 0.5088 and 0.5171.

Perspective

The analysis targets the frozen pi05_libero checkpoint with weights and action dynamics unchanged, and is limited to 30 early-rollout policy inputs, the first three from each of 10 LIBERO-10 tasks, sharing noise seed 456; paired accuracy sweeps cover only the five context_00 inputs from tasks 00, 02, 04, 06 and 08. The controlled experiments cover two target distributions, Eight-Gaussians and a bistable Langevin target, with three independently trained seeds each.

The stability test is local and based on frozen-time Jacobian eigenvalues; it does not capture transient growth from non-normal dynamics and does not include every eigenvalue at every state or the whole time interval, so finding no sampled violation does not guarantee the condition holds along the entire trajectory. Numerical accuracy and local stability do not guarantee successful robot behavior, and the learned policy may still produce a poor action. The authors propose extending the analysis to more pretrained policies, noise draws and complete robot rollouts, with denser Jacobian evaluations near the action endpoint and direct perturbation experiments.