FLASH: Efficient Visuomotor Policy via Sparse Sampling

Abstract

Generative models such as diffusion and flow matching have become dominant paradigms for visuomotor policy learning, yet their reliance on iterative denoising incurs high inference latency incompatible with real-time robotic control. We present a Fast Legendre-polynomial Action policy via Sparse History-anchored flow (FLASH), which replaces discrete action-chunk generation with continuous Legendre polynomial trajectory representation. Specifically, by fitting expert demonstrations under sparse temporal sampling, FLASH enables a single inference to cover a significantly extended action horizon. To further accelerate generation, FLASH initiates the flow matching process from history polynomial coefficients rather than uninformative Gaussian noise, shortening the transport distance and enabling accurate single-step inference. Moreover, analytic polynomial differentiation directly provides desired velocity feed-forward signals to the torque controller without numerical approximation. Extensive experiments on five simulated and two real-world manipulation tasks demonstrate that FLASH achieves state-of-the-art success rates (≥ 92% across all tasks), a per-episode inference time of 31.40 ms (up to 175× faster than diffusion policies and 18× faster than prior flow matching policies), up to 4× faster training convergence than ACT, and 5× to 7× reduction in controller tracking error compared to discrete-action baselines.

From Noise-to-Action to Coefficient-to-Coefficient

Robot motions are inherently smooth and low-frequency — a short Legendre polynomial can represent many discrete action points with just a handful of coefficients. FLASH builds on this observation in two ways:

Sparse temporal sampling. Fit polynomial coefficients at a long temporal stride, so a single inference covers a much longer action horizon without inflating the network's output dimensionality.
History-anchored flow. Initiate the flow from history polynomial coefficients rather than Gaussian noise — the transport distance becomes short enough for accurate single-step generation (NFE = 1).
Analytic velocity feed-forward. Polynomial differentiation directly produces desired velocity signals for the torque controller, with no numerical approximation and no extra training supervision.

FLASH overview: sparse sampling and coefficient-to-coefficient flow. — **Overview of FLASH.** **(a)** Expert trajectories are fitted to polynomial coefficients under sparse temporal sampling; at deployment we densely upsample, feeding both position and velocity to the controller. **(b)** Conventional generative policies generate discrete action points through multi-step denoising from noise; FLASH transports *history* coefficients to *future* coefficients in a single step.

Pipeline

Tasks

We evaluate FLASH on a Franka robot across seven manipulation tasks: five simulated tasks on the Roboverse platform — Close Box, Pick Cube, Stack Cube, Open Drawer, Pick-Place Bowl — and two real-world tasks Place Cube and Insert Cube with millimetre-level insertion tolerance.

Main Results

Success rates (%) on five simulated tasks at a shared training budget of 10k optimiser steps. NFE denotes the number of generator function sampling steps at inference. Bold = best, underline = second best. Each entry is averaged over 50 independent rollouts.

Method	NFE	Close Box	Pick Cube	Stack Cube	Open Drawer	Pick-Place Bowl
Score-UNet	100	44	58	24	2	8
DDPM-UNet	100	84	68	44	76	90
DDIM-UNet	40	80	74	42	76	92
DDPM-DiT	100	68	78	36	50	76
FM-DiT	10	70	92	46	36	68
FM-UNet	10	78	82	30	70	76
ACT	1	70	98	20	80	0
VITA	6	94	86	86	68	86
A2A-Noise	1	98	90	92	88	92
FLASH-G	10	70	94	82	62	88
FLASH (ours)	1	100	98	96	92	98

With single-step inference (NFE = 1), FLASH achieves ≥ 92% success on every task, beating the homologous FLASH-G (which differs only by starting from Gaussian noise) by 17.6 pp on average — isolating the contribution of the history-anchored flow mechanism — and the strongest single-step baseline A2A-Noise by 4.8 pp on average.

Training Efficiency

Beyond a higher performance ceiling, FLASH exhibits significantly faster convergence. On Pick Cube, FLASH reaches 96% success in just 2,500 steps — about 4× faster than ACT, the strongest baseline. On Stack Cube, FLASH is 30 percentage points above the next-best policy at 6,250 steps. On Pick-Place Bowl, FLASH stabilises at 100% success after 5,000 steps.

Inference Speed

All policies are run on the same machine (NVIDIA RTX 5090). FLASH finishes a successful episode in 31.4 ms on average: 175× faster than Score-UNet (5,476 ms), 5.1× faster than the homologous FLASH-G (159 ms, sparse sampling but noise-initialised), and 2.2× faster than the strongest baseline A2A-Noise (69 ms).

Inference time, tracking error, and post-hoc speed modulation in simulation. — **Left:** total episode inference time on *Stack Cube* across 11 policies; inset zooms the top-4 fastest. **Middle:** sum of 7-joint absolute tracking errors on *Pick Cube*; dashed lines mark task completion times. **Right:** post-hoc speed modulation by sweeping the evaluation stride k_eval; green/red bands denote the high-SR plateau and failure zone.

Real-World: Insert Cube

The polynomial parameterisation gives the low-level controller an analytic velocity feed-forward signal and a natively high-frequency output. On the real-world Insert Cube task — with millimetre-level tolerance — FLASH attains 100% success, leading seven baselines by an average of 47 pp. Across five real-world rollouts, FLASH's joint tracking MAE is 0.274 ± 0.004°, compared with 0.460 ± 0.028° for FM-DiT.

BibTeX

@article{flash2026,
  title   = {FLASH: Efficient Visuomotor Policy via Sparse Sampling},
  author  = {Bai, Jiaqi and Jia, Jindou and Hu, Yuxuan and Li, Gen and
             Chen, Xiangyu and An, Tuo and Zuo, Kuangji and Yang, Jianfei},
  year    = {2026},
}