Task Convergent Flow Matching for Generative Robotics

Lorenzo Paiola, Maximilian Stölzle, Giorgio Grioli, Daniela Rus, Cosimo Della SantinaarXiv 2026 · arXiv

TL;DR

This paper introduces an inference-time refinement framework for generative robot policies that ensures task completion by evaluating candidate motions against "performance funnels." By optimizing the source seeds of flow matching models through parallel physics simulations, the authors achieve massive jumps in success rates (e.g., from 31% to 88% on PushT) while keeping robot behavior consistent with human demonstrations.

What Problem Does This Paper Address?

Generative models like Diffusion Policies and Flow Matching have revolutionized robot learning by capturing complex, multimodal behaviors from human demonstrations. However, these models have a "short-sighted" problem: they generate plausible motions (motions that look like the training data) but don't necessarily guarantee progress.

In complex tasks like folding a shirt or stacking blocks, a robot might execute a perfectly realistic-looking grasp only to undo its previous work or stall indefinitely. Current methods often try to fix this by adding constraints at the end of a motion "chunk," but they ignore what happens during the motion. A robot might reach a good final state but cause a catastrophic mess in the middle of the trajectory.

Methods

The authors propose a task-centric refinement framework that acts as a "filter and polisher" for generative policies at inference time.

  1. Task-Level Performance Funnels: Instead of just looking at the goal, the authors define a "funnel"—a time-varying envelope for task error. This funnel allows for temporary setbacks (e.g., moving a hand away to get a better pushing angle) but strictly requires net progress by the end of each action chunk.
  2. Inference-Time Refinement via CMA-ES: When the robot needs to act, the policy proposes several candidate "seeds" (the starting noise for the generative model). The system runs these candidates through parallel physics-based simulations (rollouts). It then uses the Covariance Matrix Adaptation Evolution Strategy (CMA-ES) to search for a seed that produces a motion satisfying the funnel conditions.
  3. Radial Quantile Bands: To prevent the optimizer from finding "adversarial" motions that satisfy the math but look robotic and weird, the search is constrained to a "radial band" of the source distribution. This ensures the refined motion stays close to the learned prior (i.e., it still looks like the human demonstrations).
  4. Continuous-Time Verification: Since simulations only provide data at discrete time steps, the authors use interval-wise Lipschitz bounds to mathematically guarantee that the task error doesn't spike uncontrollably between the sampled points.
Overview of the task-funnel refinement framework

Datasets

The framework was evaluated on three challenging manipulation benchmarks:

  • PushT: A planar task where a robot must push a T-shaped block into a target zone.
  • Three-Cube Stacking: A robosuite environment requiring the robot to stack three blocks in a specific order.
  • Two-Nut Assembly: A contact-rich task from MimicGen where a robot threads a square nut and a round nut onto pegs.

Key Insights and Ideas

  • Plausibility $\neq$ Productivity: A key "aha" moment is that generative priors provide a great "vocabulary" of motion, but they need a "grammar" of task convergence to be reliable.
  • The Power of the Interior: Constraining only the endpoint of a motion is insufficient. By bounding the interior of the trajectory with a funnel, the system significantly reduces "drawdown"—the loss of already achieved progress.
  • Staying in the "Typical Set": High-dimensional Gaussians (used as seeds) concentrate on a thin shell. By constraining the search to this shell (the radial band), the authors found they could maintain high success rates while keeping the Out-of-Distribution (OOD) rate of motions similar to held-out human data.
Optimization objective for seed refinement

Key Results

The quantitative improvements are substantial across all tested environments:

  • PushT: Success rate jumped from 31.3% (nominal) to 88.5% with refinement.
  • Three-Cube Stacking: Success increased from 77.0% to 86.0%.
  • Progress Retention: On PushT, using a learned funnel lowered the average maximum loss of progress by 19% compared to just constraining the endpoint.
  • Distribution Fidelity: With the radial band constraint, only 11.3% of motion windows were considered out-of-distribution, compared to 34.1% for an unconstrained search.
Qualitative comparison of refined vs. nominal episodes

Limitations and Future Work

  • Hand-Designed Potentials: Currently, the "task potential" (the measure of error) is designed by hand using privileged simulator state. Future work aims to learn these progress measures directly from pixels or language.
  • Computational Latency: Refinement takes several seconds per chunk (0.8s to 8.3s depending on the task). This is too slow for high-speed, reactive real-time control but suitable for many industrial or service tasks.
  • Simulator Dependency: The system relies on a physics simulator as a "world model." Moving this to physical systems would require either highly accurate system identification or learned world models that can run fast enough for the CMA-ES loop.

Summary updated

Abstract

Successful manipulation requires the manipulated objects to progress toward a desired task outcome. Generative policies can produce plausible robot motions without ensuring this progress, allowing successive action chunks to stall or undo earlier changes to the scene. We introduce a task-centric inference-time refinement framework for flow matching policies that evaluates candidate action chunks through parallel physics-based rollouts and searches for motions satisfying sufficient conditions for task convergence. Task-level performance funnels bound temporary regressions while requiring net progress over each chunk. We establish sufficient conditions under which repeated funnel satisfaction yields convergence throughout execution, and use interval-wise Lipschitz bounds to verify continuous-time funnel satisfaction from sampled rollouts. To accommodate contact-rich hybrid dynamics, we refine the flow's source seeds using derivative-free Covariance Matrix Adaptation Evolution Strategy (CMA-ES), constrained to radial quantile bands of the source distribution to help retain the learned motion prior. This formulation supports both Euclidean and pose-valued Riemannian Flow Matching (RMFM) policies. On PushT and three-cube stacking, refinement raises success from 31.3% to 88.5% and from 77.0% to 86.0%. On PushT, whose coverage score responds continuously to the motion of the block, a learned funnel lowers the average maximum loss of achieved progress by 19% relative to constraining the endpoint of each chunk alone, at a similar success rate. With the radial band, 11.3% of the executed motion windows are out of distribution, a rate within the range of held-out demonstrations, compared with 34.1% for an unconstrained search.
BibTeX
@misc{paiola2026task,
  author = {Lorenzo Paiola and Maximilian Stölzle and Giorgio Grioli and Daniela Rus and Cosimo Della Santina},
  title = {{Task Convergent Flow Matching for Generative Robotics}},
  year = {2026},
  eprint = {2611.22609},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.22609},
}
Figure