Drift-VLA: Fast Vision‑Language‑Action Policies with One‑Step Drifting

One-step SmolVLA action generation: 92.3 ±0.6% success at 53.5 ms/chunk with KeyStone.

Xingdong Zuo

Paper Code

Results Method Rollouts Ablation BibTeX

Animated comparison showing flow matching taking multiple small solver steps while Drift-VLA samples in one step. — **Same SmolVLA backbone, same data.** Flow matching spends 10 expert passes per action chunk. Drift-VLA uses one; batched KeyStone selection lifts the best result to 92.3 ±0.6%.

1 NFE inference

90.2% base success

92.3% best ±0.6

4.4× lower latency

Real robot rollout

LeKiwi shopping with one-step Drift-VLA + KeyStone

Pick the requested item, carry it to the customer basket, and recover online with fast action chunks.

1 NFE KeyStone Action-dimension drift

See the full 3×3 comparison

Repeated real-world trial

Salt bottle: 10 rollouts per policy

Same initial scene

Drift-VLA + KeyStone

9/1090%

SmolVLA-FM (10 steps)

7/1070%

SmolVLA-FM (1 step)

3/1030%

At a Glance

Drift-VLA keeps SmolVLA’s backbone, data, observations, and action chunks fixed. It replaces iterative flow matching with one-step drifting in the action expert. The winning recipe is action-dimension drift + KeyStone: 92.3 ±0.6% LIBERO-Spatial success at 53.5 ms/chunk.

Why one-step matters

Iterative VLA samplers spend K expert passes per action chunk. Drift-VLA asks whether one pass can keep FM-level success.

Stacked latency bars per method: shared VLM prefix cost plus action-expert cost times K.

The shared VLM prefix is paid once; the repeated action-expert calls are what one-step drifting removes.

Base Drift-VLA runs at 50.6 ms/chunk; FM-10 runs at 232.9 ms/chunk. That headroom pays for KeyStone while staying at 53.5 ms/chunk.

Two Action Samplers

Same observation context, same action chunk. The difference is how the action expert is trained and sampled.

Flow matching

Learns a velocity field. Inference integrates it over multiple solver steps, so higher quality means repeated action-expert evaluations.

Drifting

Trains the generator directly. Inference is one forward pass of the action expert: no solver loop and no distillation.

Training signal

Animated flow-matching training view: noise samples are paired with demonstration actions and a velocity direction is learned. — **Flow matching learns a direction field.** Training pairs noise with demonstrations; inference follows the learned field over multiple solver steps.

Animated drifting training view: model samples are pulled toward demonstrations while being kept separated. — **Drifting trains the generator directly.** Samples move toward demonstrations while staying diverse, so the action expert can sample in one pass.

Method

Only the action expert objective changes. The backbone, observations, action chunks, and data stay fixed.

The winning loss is action-dimension drift: each real action dimension’s chunk trajectory is matched separately.

Training objective

From iterative transport to one-step drifting

same backbone + data

SmolVLA-FM

\[ x_t = t\epsilon + (1-t)a,\qquad \mathcal{L}_{\mathrm{FM}} = \lVert v_\theta(x_t,t) - (\epsilon-a) \rVert^2 \]

Learn a velocity field; sample by integrating multiple expert steps.

Drift-VLA

\[ \mathcal{L}_{\mathrm{drift}} = \left\lVert g/s - \operatorname{sg}\!\left(g/s+V\right) \right\rVert^2 \]

Train the generator directly; sample the full chunk in one expert pass.

\(V\) pulls generated samples toward demonstrations and separates sibling samples. Drift-VLA applies it per action dimension across the full chunk.

Animated Drift-VLA pipeline: observations are encoded once, the action expert samples, drifting updates the action expert, and the prefix remains shared. — **End-to-end training pipeline.** SmolVLA’s observation encoder and action interface stay intact; the action expert trains with action-dimension drifting.

Results

Animated KeyStone inference pipeline: Drift-VLA samples multiple one-step action chunks, finds a consensus candidate, and executes the selected action chunk. — **KeyStone at inference time.** Drift-VLA samples multiple one-step chunks, selects the consensus candidate, and executes it. No retraining or new data.

Inference selector

KeyStone consensus over one-step samples

8× Drift-VLA

\[ \hat a = \operatorname{medoid}\!\left( \operatorname{largest\ cluster} \left(\{ f_\theta(\epsilon_k) \}_{k=1}^{K}\right) \right) \]

KeyStone chooses the representative chunk from the dominant candidate cluster.

SmolVLA-FM (10 steps)

Success: 88.0 ±1.5%
Latency: 232.9 ms
NFE: 10

Drift-VLA

Success: 90.2 ±1.3%
Latency: 50.6 ms
NFE: 1

              Drift-VLA + KeyStone
              Success92.3 ±0.6%
Latency53.5 ms
Sampling8×1

            

Base Drift-VLA already exceeds FM-10 while running 4.6× faster. KeyStone spends a small part of that headroom and still runs 4.4× faster.

Method	NFE	Latencyms / chunk	Success (%)LIBERO-Spatial · 200 eps × 3
SmolVLA-FM (10 steps)	10	232.9	88.0 ± 1.5
SmolVLA-FM (1 step)	1	50.1	84.0 ± 2.0
Drift-VLA (action-dimension drift)	1	50.6	90.2 ± 1.3
Drift-VLA (action-dimension drift) + KeyStoneBest observed	8×1	53.5	92.3 ± 0.6

Real-World Shopping Rollouts

Representative LeKiwi rollouts across three shop items. Drift-VLA + KeyStone is shown beside the flow-matching baselines under the same task framing.

Matcha tin

Drift-VLA + KeyStone 1 NFE

SmolVLA-FM 10 NFE

SmolVLA-FM 1 NFE

Salt bottle

Drift-VLA + KeyStone 1 NFE

SmolVLA-FM 10 NFE

SmolVLA-FM 1 NFE

Chocolate bar

Drift-VLA + KeyStone 1 NFE

SmolVLA-FM 10 NFE

SmolVLA-FM 1 NFE

Drift Grouping Ablation

Action-dimension drift is the strongest and most stable grouping: match each action dimension’s chunk trajectory separately.

Rank	Drift grouping	20k SR (%)	30k SR (%)
1	Action-dimension drift one matching unit per action dimension	89.0 ±1.4	90.2 ±1.3
2	Chunk-level drift joint matching over the complete action chunk	86.5 ±2.1	77.8
3	Timestep drift (8 samples) per-timestep matching with 8 sibling samples	78.8	81.3
4	Timestep drift (16 samples) same grouping with more sibling samples	76.5	79.0
5	Timestep drift (4 samples) same grouping with fewer sibling samples	78.8	78.8
6	Shared-temperature drift one temperature shared across dimensions	70.0	78.0

Same SmolVLA setup and data; only the drift grouping changes.

LIBERO Rollouts

Three LIBERO-Spatial cases: the best Drift-VLA setup succeeds, while the comparison variants fail on the same tasks.

Case A

Drift-VLA + KeyStone success

SmolVLA-FM (10 steps) failure

Timestep drift (4 samples) failure

Case B

Drift-VLA + KeyStone success

SmolVLA-FM (10 steps) failure

Timestep drift (4 samples) failure

Case C

Drift-VLA + KeyStone success

SmolVLA-FM (10 steps) failure

Timestep drift (4 samples) failure

References

SmolVLA: A Vision-Language-Action Model for Affordable and Efficient Robotics Shukor et al., arXiv 2025.

Generative Modeling via Drifting Deng et al., arXiv 2026.

Geometry Guided Self-Consistency for Physical AI Dai et al., arXiv 2026. Introduces KeyStone.

Citation

@misc{zuo2026driftvla,
  title  = {Drift-VLA: Fast Vision-Language-Action Policies with One-Step Drifting},
  author = {Zuo, Xingdong},
  year   = {2026},
  url    = {https://zuoxingdong.github.io/drift-vla/}
}