FlashVLA: Streaming Action Decoding for Fast and Asynchronous VLA Inference

The Bottleneck Is Chunk Isolation

Prior work treats these as separate problems. Efficient-inference methods make each forward pass cheaper with lightweight architectures, compression, or token pruning, but the decoding loop still blocks. Asynchronous methods condition on future states or planned actions, but still require lookahead, which re-amplifies mismatch.

Our observation is that two issues share the same cause: chunk isolation. A flow-matching VLA decodes each action chunk from pure noise, conditioned only on the current observation and decoupled from chunks already in flight. Isolation concentrates all denoising steps inside one chunk decode, creating latency, and leaves each new chunk uninformed about the trajectory it is about to join, creating mismatch.

FlashVLA: Streaming Action Decoding

FlashVLA keeps a rolling buffer of N action chunks. Earlier slots are closer to execution and cleaner; later slots are farther in the future and noisier. Each forward pass advances the whole buffer by one denoising step. After warm-up, the cleanest chunk is emitted, the buffer shifts, and a fresh noisy chunk enters the tail.

This single formulation gives FlashVLA three advantages:

• Amortized denoising. Each chunk still traverses the full denoising path, but the passes are spread across streaming steps. The robot gets one executable chunk per step instead of waiting for a full blocking decode.
• Implicit asynchronous conditioning. Future chunks attend to cleaner near-execution chunks, so the decoder is already conditioned on the trajectory the robot is following. No future-state predictor is needed.
• Chunk-level memory. Each noisier chunk attends to all earlier chunks in the buffer, giving the action expert a short structured history of recently refined actions. This is especially useful on long-horizon tasks.

Training

We adopt multi-buffer joint fine-tuning. For each observation, we pack all valid buffer states into one training sample: cold-start states with padding chunks, and the full steady-streaming state. These packed buffers share the same visual-language context, while an attention mask prevents interaction across buffer states. The objective is still the standard flow-matching loss, just summed over the packed buffer states.

Simulated Evaluation

Real-World Experiment

We deploy FlashVLA on a 7-DoF Franka arm with a consumer-grade RTX A4000 and ask whether the simulated latency and asynchronous-continuity benefits transfer to a real robot at 30 Hz. The tasks span short, medium, and long horizons: pick-and-place, whiteboard wiping, and table cleaning. We compare $\pi_{0.5}$ synchronous, naive asynchronous $\pi_{0.5}$, RTC, and FlashVLA; all asynchronous methods use a two-step delay.

FlashVLA exceeds all three $\pi_{0.5}$ configurations on every task, improving the cross-task average score from 80.0% for synchronous $\pi_{0.5}$ and RTC (75.6% for naive asynchronous execution) to 84.4%. Completion time is measured only over successful trials, so the 1.3x speedup over synchronous execution and 1.2x over RTC cannot come from failed rollouts ending early. The joint gain in score and speed shows that streaming removes chunk-boundary stalls while preserving asynchronous continuity. Relative to synchronous execution, the absolute time saved grows from 3.1 s on pick-and-place to 10.2 s on table cleaning.

On the RTX A4000, FlashVLA has an inference latency of 67.3 ms, approximately two 30 Hz control periods. A two-step asynchronous delay overlaps this latency with action execution, enabling real-time 30 Hz control on a consumer-grade GPU.

Conclusion

FlashVLA shows that VLA action decoding does not have to be a blocking loop. A streaming buffer of action chunks at staggered noise levels collapses the latency cost and the asynchronous-mismatch cost into one design. It cuts per-step action-decoding latency by up to 20x while keeping every action chunk on a full denoising trajectory.

The same mechanism fixes asynchronous mismatch. Cleaner near-execution chunks become context for noisier future chunks, so the next action is conditioned on the motion already underway. Fast inference and smooth execution come from the same design.

Citation

@article{flashvla2026,
  title = {{FlashVLA: Streaming Action Decoding for Fast and Asynchronous VLA Inference}},
  author = {Li, Zekai and Tang, Jiaming and Liu, Zhijian},
  year = {2026}
}