Skip to main content
Back to timeline
arXivSource publication:

Flow-matching robot policies get a task-progress check: physics rollouts lift PushT success from 31.3% to 88.5%

Synopsis

Generative manipulation policies can produce plausible action chunks that still stall or undo earlier progress. An inference-time refinement method for flow-matching policies evaluates candidate chunks through parallel physics-based rollouts and requires each chunk to satisfy a task-performance funnel with terminal contraction, so repeated execution drives task error toward zero. On PushT and three-cube stacking, refinement raises success from 31.3% to 88.5% and from 77.0% to 86.0%; on PushT, a learned funnel lowers the mean maximum drawdown of achieved progress by 19% relative to constraining only each chunk's endpoint, at a similar success rate.

Source-provided article image: Task Convergent Flow Matching for Generative Robotics
arXiv

Interpretation

Action chunks from generative manipulation policies can look plausible without moving the manipulated objects toward task completion, and successive chunks can stall or undo earlier changes. Earlier inference-time refinement targeted collision avoidance, admissible trajectories, or stability of robot motion, and those requirements do not by themselves ensure that successive chunks advance the objects; objectives based on predicted endpoint progress or accumulated reward also do not specify how much regression within a chunk is acceptable. On PushT, the nominal policy without refinement succeeds in 31.3% of episodes, and endpoint-only refinement raises this to 86.8%; on three-cube stacking, the same comparison raises success from 77.0% to 86.0%.

Refinement leaves the policy fixed and instead searches over the flow-matching source seed, using physics rollouts to predict each candidate chunk's effect on the scene and selecting plans that satisfy the task funnel. The task funnel is a time-varying error envelope that permits temporary regression inside a chunk while requiring a net decrease by its end; the search uses the derivative-free Covariance Matrix Adaptation Evolution Strategy (CMA-ES, a population search that samples candidates and iteratively updates its distribution) inside a radial quantile band of the source distribution to retain the learned motion prior. On 200 new PushT initial conditions, band-constrained CMA-ES reaches 86.5% success, above 70.0% for L-BFGS, 82.5% for best-of-n and 79.0% for CEM; removing the band raises success to 96.0% but increases the share of executed windows out of distribution relative to the training demonstrations from 11.3% to 34.1%.

The funnel's transient bound retains more of the progress already achieved on PushT, where the task score responds continuously to motion within a chunk. Constraining only the chunk endpoint lets the interior regress sharply and recover at the end, so a favorable endpoint can conceal loss of progress within the chunk. On PushT, the learned additive funnel lowers the mean maximum drawdown from 62.8 to 52.3 points relative to endpoint-only, at a success rate statistically on par with endpoint-only; with the band, 11.3% of executed windows are out of distribution, within the 9.6% (95% interval 5.4–15.4%) range of held-out demonstrations.

Perspective

The demonstrated scope is planar PushT, three-cube stacking and two-nut assembly in simulation, where refinement reads full scene state from the simulator and rolls out candidates in that same simulator; the convergence results apply to executed chunks that satisfy their funnels.

Task potentials are designed by hand from task geometry, and refinement quality follows their quality; the search can execute a plan that does not satisfy its funnel, and continuous-time verification uses rates estimated from sampled rollouts rather than certified bounds; planning takes seconds per chunk, during which the scene must stay still.