One Trajectory, Two Requirements: Accuracy and Local Stability in Euler Sampling for Flow Matching Policies

Mahindra RautelaarXiv 2026 · arXiv

TL;DR

Flow matching robot policies generate actions by numerically integrating an ODE, but how many steps are actually necessary? This paper demonstrates that the step count is determined by two distinct requirements: following the nominal trajectory accurately and maintaining local numerical stability against "stiff" contracting directions. While current state-of-the-art models like $\pi_{0.5}$ are primarily limited by accuracy, the authors prove that stability can impose its own independent—and potentially stricter—bottleneck on inference speed.

What Problem Does This Paper Address?

In the world of Vision-Language-Action (VLA) models, Flow Matching has emerged as a powerful way to generate continuous robot actions. Unlike traditional diffusion models that require many denoising steps, flow matching learns a vector field that "transports" simple noise into complex action sequences.

To turn this learned field into a robot command, we use a numerical integrator—most commonly the simple Euler method. The big question for engineers is: How many Euler steps ($N$) do we need?

Usually, $N$ is chosen by "sweeping" different values and picking the smallest one that yields acceptable endpoint accuracy. However, this paper argues that accuracy isn't the whole story. ODEs can be "stiff," meaning they have directions that contract very rapidly. If your Euler step is too large relative to these contraction rates, the solver can actually amplify noise rather than suppress it, leading to numerical instability. The authors investigate whether this "local stability" requirement might sometimes demand more steps than accuracy alone.

Figure 1: Accuracy and local stability impose distinct numerical requirements. Accuracy (left) ensures we follow the path; stability (right) ensures we don't blow up nearby perturbations.

Methods

The researchers tackled this question in three stages:

  1. Theoretical Proof: They designed an exact two-coordinate ODE system where one coordinate controls trajectory curvature (accuracy) and the other controls local contraction (stability). This allowed them to formally prove that the required step count is the maximum of the two requirements: $N_{required} \approx \max\{N_{acc}, N_{stab}\}$.
  2. Controlled Augmentation: They took standard flow matching models trained on toy distributions and artificially added a "contracting dimension." This allowed them to dial up the stability requirement without changing the original learned path, demonstrating that stability can indeed become the primary bottleneck.
  3. Real-World Analysis: They analyzed a frozen, pretrained $\pi_{0.5}$-LIBERO model. They computed the Jacobian (the matrix of local derivatives) along real inference trajectories to see if real robot policies encounter these "stiff" stability limits.

Datasets

  • Toy Targets: Used for controlled experiments. These included "Eight-Gaussians" (a mixture of eight Gaussians on a circle) and a "Bistable Langevin" target (representing a double-well potential).
  • LIBERO-10: A benchmark consisting of 10 diverse household manipulation tasks (e.g., putting a bowl in a drawer, loading a microwave). The authors evaluated 30 different policy inputs across these tasks.

Key Insights and Ideas

  • The Stability-Accuracy Decoupling: A trajectory can be highly accurate at its endpoint while being locally unstable. This means a "coarse" solver might land in the right spot but misrepresent how the model handles small errors or noise along the way.
  • Stiffness at the Finish Line: In the $\pi_{0.5}$ model, the strongest contracting directions (and thus the highest stability demands) occur near the action endpoint ($t \to 0$). This suggests that if stability were to fail, it would likely happen just as the robot is finalizing its action plan.
  • Current Models are Accuracy-Limited: For the $\pi_{0.5}$ model, the "stability floor" is very low (requiring only ~1 step), while achieving 1% endpoint accuracy requires over 100 steps. Currently, we are paying for accuracy, not stability.
Figure 4: Results for the pretrained $\pi_{0.5}$ policy. While stability (left) is well-behaved, accuracy (right) becomes significantly more demanding as the error tolerance tightens.

Key Results

  • Stability is Safe (For Now): Across all 30 LIBERO tasks, the $\pi_{0.5}$ model never violated the Euler stability condition. The "scaled stability measure" stayed around 0.5, well below the failure threshold of 1.0.
  • The Cost of Precision: To get within 5% of the "true" trajectory, the model needs only 4–9 steps. However, to reach 1% error, the requirement jumps to 87–112 steps. This order-of-magnitude increase in cost is driven entirely by the need for trajectory accuracy.
  • Jacobian Complexity: Even though the model is stable, it is highly complex. Near the endpoint, 319 out of 320 eigenvalues in the Jacobian were positive, indicating widespread contracting directions.

Limitations and Future Work

  • Sample Size: The analysis was limited to 30 policy inputs and a single noise seed. Future work should explore a wider variety of noise draws and full robot rollouts.
  • Local vs. Global: The stability test used is "local" (looking at a single point in time). It might miss "transient growth" where errors grow temporarily before shrinking.
  • Task Success: The paper focuses on numerical metrics (error %). A critical next step is determining how these numerical errors actually translate to physical task failure or success for the robot.
  • Other Solvers: While Euler is standard, exploring higher-order or adaptive-step solvers could provide a better balance between accuracy and stability.

Summary updated

Abstract

Flow matching robot policies commonly use a fixed Euler step count selected from offline accuracy sweeps. At inference time, however, each policy input (visual observation, robot state, and language instruction) together with a source-noise draw defines a different ODE trajectory and sequence of local Jacobians. We ask whether a step count that accurately follows the nominal noise-to-action trajectory can still be too coarse for strongly contracting directions around it. An exact two-coordinate system and a controlled augmentation of six trained flow models show that trajectory accuracy and local Euler stability can impose different step requirements even when the original learned transport is unchanged. We then study a frozen $π_{0.5}$-LIBERO checkpoint without modifying its weights or action dynamics. Across 30 policy inputs from 10 LIBERO-10 tasks, the Jacobian exhibits strong contracting directions near the action endpoint, yet no sampled mode violates the Euler non-expansion condition. On five of those inputs, accuracy sweeps require 4--9 Euler steps at 5% terminal error and 87--112 at 1%. These five paired cases are therefore consistent with accuracy-limited rather than stability-limited sampling under this local criterion, while the controlled experiments show why endpoint accuracy alone cannot rule out a separate stability requirement.
BibTeX
@misc{rautela2026one,
  author = {Mahindra Rautela},
  title = {{One Trajectory, Two Requirements: Accuracy and Local Stability in Euler Sampling for Flow Matching Policies}},
  year = {2026},
  eprint = {2611.13311},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.13311},
}
Figure