FlashVLA: Streaming Action Decoding for Fast and Asynchronous VLA Inference

Paper | Blog | Models

FlashVLA is a general streaming action decoding method for flow-matching VLA models, achieving fast and asynchronous execution.

Install

git clone https://github.com/z-lab/flashvla.git
cd flashvla
conda env create -f environment.yml
conda activate flashvla

This installs the core FlashVLA env for training, benchmarking, and real-robot deployment. LIBERO and RoboTwin evaluation each need extra, simulator-specific setup — see their READMEs under sim_eval/.

Evaluation

Evaluate the released FlashVLA pi0.5 checkpoints with async chunk-overlap inference.

LIBERO (z-lab/flashvla-pi05-libero):

bash sim_eval/libero/eval.sh

RoboTwin 2.0 (z-lab/flashvla-pi05-robotwin) uses a server (flashvla env) / client (RoboTwin env) split. After the one-time setup in sim_eval/robotwin/, from $ROBOTWIN/policy/pi05_flashvla/:

bash eval_server.sh z-lab/flashvla-pi05-robotwin 9999 0 current_state 1 16 true  # terminal 1
EVAL_STEP_LIM_OFFSET=48 ROBOTWIN_VENV=... bash eval_client.sh beat_block_hammer demo_clean  # terminal 2

See sim_eval/libero/ and sim_eval/robotwin/ for setup and full options (single run, multi-seed sweeps, and the RoboTwin server/client).

Performance

π0.5 vs. + FlashVLA (bold = the better result). d denotes how many steps before the current chunk ends the next inference is launched: d=0 is synchronous, while d≥1 overlaps inference with robot execution. The reported π0.5 evaluations execute 5 actions per segment on LIBERO and 16 on RoboTwin.

Time/Step is mean episode wall time divided by mean executed low-level steps, including rendering and (for RoboTwin) client-server communication. Inference latency measures one policy invocation after warm-up; it is not Time/Step. The reported up-to-20× gain applies only to the action-decoding stage in the profiling experiment.

LIBERO

Success rate (%) and Time/Step (2,000 episodes per run, averaged over five runs):

Method	Spatial	Object	Goal	Long	Avg	Time/Step (ms)
π0.5	98.8	98.2	98.0	92.4	96.9	53.8
+ FlashVLA (d=0)	98.6	99.0	97.8	96.2	97.9 (↑1.0)	–
+ FlashVLA (d=1)	98.8	99.6	97.6	95.4	97.8 (↑0.9)	22.1 (2.43×)
+ FlashVLA (d=2)	99.0	99.8	97.2	97.4	98.3 (↑1.4)	20.6 (2.62×)

RoboTwin 2.0

50-task multitask success rate (%):

Method	Clean	Random	Avg
π0.5	86.1	85.8	86.0
+ FlashVLA (d=0)	90.8	90.2	90.5 (↑4.5)
+ FlashVLA (d=1)	91.2	89.9	90.6 (↑4.6)
+ FlashVLA (d=2)	91.0	90.2	90.6 (↑4.6)

Cross-Architecture Generalization

Backbone	Method	Avg SR (%)	Time/Step (ms)	Inference Latency (ms)
SmolVLA	Baseline	80.1	41.2	19.7
SmolVLA	+ FlashVLA (d=0)	80.1 (↑0.0)	29.7 (1.39×)	10.1 (1.95×)
SmolVLA	+ FlashVLA (d=1)	79.5 (↓0.6)	28.7 (1.44×)	–
LingBot-VLA	Baseline	85.2	56.7	70.6
LingBot-VLA	+ FlashVLA (d=0)	88.6 (↑3.4)	49.5 (1.15×)	25.1 (2.81×)
LingBot-VLA	+ FlashVLA (d=1)	89.3 (↑4.1)	46.6 (1.22×)	–

Inference latency is measured separately from the overlap schedule and is reported once at d=0. SmolVLA preserves synchronous success and loses 0.6 points at d=1; LingBot-VLA improves success in both modes.

Latency Benchmark

Per-policy-invocation inference latency is measured with benchmarks/benchmark_latency.py:

python benchmarks/benchmark_latency.py --config_path=benchmarks/configs/latency_flashvla.yaml

Use --num_views=1, --num_views=2, or --num_views=3 to sweep camera views.

Inference latency (ms) on RTX 4090 / 5090 with two and three camera views, averaged over 100 samples after 10 warm-up iterations. π0.5 and FlashVLA use the same CUDA Graph and kernel-fusion optimizations.

Method	RTX 4090 (2 views)	RTX 4090 (3 views)	RTX 5090 (2 views)	RTX 5090 (3 views)
π0.5	45.8	55.4	37.0	44.8
+ Realtime-VLA	29.2	38.9	26.6	34.2
+ FlashVLA	26.7	36.8	20.3	27.1

Training

LIBERO:

bash train/train.sh train/configs/pi05/libero/pi05_flashvla.yaml

RoboTwin:

bash train/train.sh train/configs/pi05/robotwin/pi05_flashvla.yaml

Acknowledgement

This project builds on LeRobot and VLASH.

Citation

@inproceedings{flashvla2026,
  title = {{FlashVLA: Streaming Action Decoding for Fast and Asynchronous VLA Inference}},
  author = {Li, Zekai and Tang, Jiaming and Liu, Zhijian},
  booktitle = {Conference on Robot Learning (CoRL)},
  year = {2026}
}

License

Apache-2.0