Task Convergent Flow Matching for Generative Robotics

TL;DR Before executing each action chunk of a flow-matching policy, they search its noise seed with CMA-ES over simulated rollouts for a chunk whose task error stays inside a funnel that allows temporary setbacks but requires net progress. In simulation, success rises from 31.3% to 88.5% on PushT and from 77.0% to 86.0% on three-cube stacking.
What they do
Diffusion and flow-matching policies produce plausible action chunks, but successive chunks can stall or undo earlier progress. This paper adds an inference-time refinement step: simulate candidate chunks in parallel, track a task-error potential V along each rollout, and execute a chunk that meets conditions under which repeated execution drives the task error to the goal.
How it works
- Performance funnels. An envelope on V over the chunk that tolerates temporary increases but requires a lower value at the end. If every executed chunk stays in its funnel, with uniformly positive terminal contraction and a bounded transient allowance, task error provably converges. Funnels are fixed, calibrated, or predicted from the observation by a small network fitted to demonstrations.
- Verification. Interval-wise Lipschitz bounds check the funnel between simulation samples, using rates estimated from the rollout.
- Seed search. The trained flow stays fixed; CMA-ES searches its source seed, projected into a radial band of the Gaussian so decoded chunks stay close to the demonstrations. Works for Euclidean and pose-valued Riemannian flow matching.
Results
- 400 paired initial conditions, all in simulation. Constraining only the chunk endpoint gives most of the gain: 31.3% to 86.8% on PushT, 77.0% to 86.0% on stacking. The calibrated funnel reaches 88.5% on PushT.
- Requiring V to fall throughout the chunk reaches only 54.5-60.5% on PushT, where good pushes often first move the T away.
- Learned funnels retain progress: on PushT the mean drawdown of the worst 10% of episodes drops from 62.8 to 52.3 points vs endpoint-only, at on-par success; on stacking from 46.1 to 40.5.
- On PushT, banded CMA-ES reaches 86.5% vs 70.0% for L-BFGS and 82.5% for best-of-N. Removing the band gives 96.0% but raises out-of-distribution motion windows from 11.3% to 34.1%.
Limitations
Simulation only. The method needs the full scene state, a simulator that predicts execution, and a hand-designed V. Guarantees cover only chunks that pass their funnel, and verification uses estimated rather than certified rates. Planning takes seconds per chunk while the scene must stay still.
Figures



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.
Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.
