FastOPD: On-Policy Distillation for Lightweight VLA Deployment

Abstract

Vision-Language-Action (VLA) foundation models have scaled rapidly to enhance manipulation performance and generalizability, but this scaling incurs high computational costs that render real-world deployment increasingly challenging. Existing approaches typically mitigate this issue by designing smaller architectures or reducing the iterative denoising steps in flow-based policies.

In this work, we propose FastOPD, a foundation-to-lightweight VLA framework that enables the practical deployment of large-scale VLAs through efficient on-policy distillation. Specifically, FastOPD adapts a flow map for single-state teacher supervision and combines it with a self-consistency objective to construct a compact student that learns the teacher dynamics. Furthermore, we theoretically demonstrate that minimizing this objective allows the distilled student to recover a distribution on par with that induced by an ideal few-step teacher model.

We evaluate FastOPD across diverse foundation policies in simulation and real-world experiments. On LIBERO, FastOPD retains 84% of the performance of π_0.5 with only two inference steps, reducing inference latency by 78.1% while outperforming existing few-step distillation baselines in average success rate. With LingBot-VLA as the teacher, FastOPD improves the single-step success rate over the base student by 15.9 percentage points on RoboTwin 2.0. We further demonstrate its applicability to a World Action Model (WAM) and deploy a compact student distilled from MolmoAct2 on a real robot.

Motivation

How can we apply on-policy distillation in an efficient manner?

On-policy distillation: the student rolls out a multi-step trajectory from noise and the teacher velocity is queried at every intermediate state. — (a)

FastOPD: the student jumps directly from noise to an intermediate state and the teacher velocity is queried only at that single state. — (b)

(a) Conventional OPD evaluates the teacher at every denoising step, resulting in substantial computational overhead.
(b) FastOPD evaluates teacher at a single sampled state reached by a flow-map jump from noise.

Method

FastOPD overview: (a) flow map architecture with a 350M vision-language model, a 102M action expert with a time projection layer conditioned on s and t; (b) on-policy flow map distillation loss matching the student velocity to the teacher velocity at a state reached by a single jump from noise; (c) self-consistency loss aligning a direct jump with two shorter steps through a midpoint.

To match the teacher only at sampled on-policy states and propagate this local velocity supervision to finite-step flow maps, we combine On-Policy Flow map Distillation (OPFD) loss with Self-Consistency (SC) loss. We freeze the VLM backbone and fine-tune only the action expert with additional time projection layer.

Benchmark Results

LIBERO Simulation

Our method outperforms all few-step baselines (DMD, CTM, and iMF) and the initial student model at every denoising step in average success rate (%). FastOPD further achieves the best success rates of 81.8% with only 2 steps.

Success rate (%) on LIBERO during training. Hover over a chart to see the values at each training step.

LIBERO Simulation

Method	Param.	NFE	Spatial	Object	Goal	Long	Avg.
Diffusion/Flow-based Policies
DP^†	–	100	78.3	92.5	68.3	50.5	72.4
Octo-Base^†	93M	10	78.9	85.7	84.6	51.1	75.1
π₀^†	3.3B	10	97.4	98.4	97.6	93.0	96.6
Teacher & Base Student Models
π_0.5 (teacher)	3.3B	10	98.2	99.4	97.4	94.8	97.5
SmolVLA	450M	1	74.6	73.2	77.4	51.2	69.1
		2	76.8	74.4	79.8	54.4	71.4
		4	79.0	74.8	79.2	58.0	72.8
		10	77.2	75.0	79.4	54.6	71.6

Method	Param.	NFE	Spatial	Object	Goal	Long	Avg.
Few-step Distillation
DMD	450M	1	82.4	87.2	75.6	47.2	73.1
CTM	451M	1	78.0	86.6	70.8	41.4	69.2
		2	75.4	84.2	70.4	43.4	68.4
		4	68.4	77.0	71.2	45.0	65.4
		10	72.0	81.6	72.4	44.4	67.6
iMF	451M	1	70.8	78.6	75.6	30.6	63.9
		2	76.0	90.6	75.6	46.4	72.2
		4	68.2	84.8	67.0	38.0	64.5
		10	65.2	83.6	66.6	35.2	62.7
FastOPD (Ours)	451M	1	84.2	94.2	82.6	48.2	77.3
		2	83.6	96.8	87.2	59.6	81.8
		4	83.0	95.0	85.6	57.4	80.3
		10	80.4	94.0	84.4	57.2	79.0

RoboTwin 2.0 Simulation

Our method consistently achieves performance gains in the few-step regime (1,2 steps), improving the 1 step success rate of the base student from 35.3% to 51.2%, and outperforms all few-step distillation baselines at every sampling step.

Method	Param.	NFE	Clean	Randomized	Avg.
Diffusion/Flow-based Policies
DP^†	0.46B	10	28.0	0.6	14.3
RDT^†	1.2B	10	34.5	13.7	24.1
π₀^†	3.3B	10	46.4	16.3	31.4
π_0.5	3.3B	10	70.7	46.0	58.4
Teacher & Base Student Models
LingBot-VLA (teacher)	4.2B	10	86.8	84.9	85.8
Fast-WAM (teacher)	6B	10	91.9	91.8	91.9
SmolVLA	450M	1	35.2	35.4	35.3
		2	55.0	56.1	55.6
		4	59.8	61.0	60.4
		10	59.4	61.3	60.4

Method	Param.	NFE	Clean	Randomized	Avg.
Few-step Distillation
DMD	450M	1	32.5	32.1	32.3
CTM	451M	1	15.5	15.6	15.6
		2	4.8	15.8	10.3
		4	18.7	16.0	17.3
		10	36.1	34.5	35.3
iMF	451M	1	43.2	42.4	42.8
		2	51.6	51.1	51.4
		4	50.8	51.0	50.9
		10	50.6	51.4	51.0
FastOPD (Ours) LingBot-VLA teacher	451M	1	51.5	50.9	51.2
		2	57.1	58.3	57.7
		4	58.6	61.2	59.9
		10	56.9	59.6	58.2
FastOPD (Ours) Fast-WAM teacher	451M	1	51.1	49.8	50.5
		2	55.5	57.2	56.3
		4	58.5	59.2	58.8
		10	58.3	59.6	58.9

Training and inference efficiency

Training efficiency

FastOPD supervises the student at a single sampled on-policy state instead of across all denoising steps.

Training time per 1,000 iterations (h)

Standard OPD9.51 h

FastOPD (Ours)1.43 h

This carries over to convergence: FastOPD reaches 78.2% 1 step success in just 10h, while OPD needs roughly 57h to reach a similar level (79.1%), a 5.7× speedup at a fraction of the compute.

Success rate versus wall-clock training time on LIBERO: FastOPD reaches 78.2% 1 step success within 10 hours, while OPD needs about 57 hours to reach 79.1%, a 5.7x speedup.

Inference efficiency

The distilled student cuts latency by 78% relative to π_0.5 on LIBERO (301 → 66 ms) and runs 10× faster than LingBot-VLA on RoboTwin 2.0 (571 → 57 ms).

End-to-end latency on a single RTX 3090 (bf16, batch size 1).
Benchmark	Model	Steps	Latency
LIBERO	π_0.5 teacher	10	301 ms
LIBERO	FastOPD	2	66 ms
RoboTwin 2.0	LingBot-VLA teacher	10	571 ms
RoboTwin 2.0	FastOPD	1	57 ms
RoboTwin 2.0	Fast-WAM teacher	10	503 ms
RoboTwin 2.0	FastOPD	1	57 ms

Real-World Deployment

From MolmoAct2 (5B) to a compact student (451M) on a physical robot. FastOPD shows faster and more accurate performance than the 10 steps base student, improving the success rate from 44% to 50% while cutting the execution time from 19.32s to 17.38s.

Real-world robot demonstrations on the pnp-plate task. Teacher (NFE=10), Stage 1 and Stage 2 (NFE=4).

(a) Success rate (%) ↑

78.0%MolmoAct210 steps

44.0%SmolVLA10 steps

42.0%SmolVLA4 steps

50.0%FastOPD (Ours)4 steps

(b) Completion time (s) ↓

17.95 sMolmoAct210 steps

19.32sSmolVLA10 steps

17.20sSmolVLA4 steps

17.38 sFastOPD (Ours)4 steps

Real-world results on the pnp-plate task. Completion time is averaged over successful episodes.

Few-step control on a YAM robot

Using MolmoAct2 as the teacher, the distilled student achieves 50% success with 4 steps, compared with 42% for the base SmolVLA at the same step count.

Compared with the 10 steps base student, task completion time decreases from 19.32s to 17.38s.

BibTeX

@misc{oh_fastopd,
  title = {FastOPD: On-Policy Distillation for Lightweight VLA Deployment},
  author = {Oh, Yoojin and Kim, Jeongsol and Seo, Yeonwoo and
            Park, Jangho and Jin, Seonghyun and Park, Sunwoo and
            Kim, Youngmin and Jun, Youngjun and Choi, Kyumin and
            Ye, Jong Chul},
  note = {Preprint},
  url = {https://fastopd.github.io/}
}