FlashVLA: Streaming Action Decoding for Fast and Asynchronous VLA Inference

Summary

Abstract Summary

Vision-Language-Action (VLA) models are increasingly promising for robotic manipulation, yet their real-world deployment remains bottlenecked by high inference latency and unstable asynchronous execution. This challenge is particularly pronounced in flow-matching-based VLA models, where action decoding requires multiple iterative steps conditioned on the VLM context. While efficient inference methods improve control frequency and asynchronous methods reduce execution idle time, existing approaches often fail to jointly achieve low-latency inference and accurate, temporally consistent asynchronous execution. We introduce \textbf{FlashVLA}, a streaming action decoding framework that addresses both challenges in a unified formulation. FlashVLA maintains a streaming action buffer with multiple chunks at different noise levels and decodes them using chunk-wise causal attention. This design allows FlashVLA to produce one executable action chunk per inference step. Moreover, its chunk-wise autoregressive formulation implicitly preserves action continuity, enabling smooth asynchronous execution without extra future-state conditioning. Across extensive simulated and real-world experiments, FlashVLA substantially improves inference speed while maintaining strong task performance. It can achieve $\geq$30\,Hz control frequency on a single GPU with smooth asynchronous inference in real-world deployment.

Key Points

Vision-Language-Action (VLA) models are increasingly promising for robotic manipulation, yet their real-world deployment remains bottlenecked by high inference latency and unstable asynchronous execution.
This challenge is particularly pronounced in flow-matching-based VLA models, where action decoding requires multiple iterative steps conditioned on the VLM context.
While efficient inference methods improve control frequency and asynchronous methods reduce execution idle time, existing approaches often fail to jointly achieve low-latency inference and accurate, temporally consistent asynchronous execution.
We introduce \textbf{FlashVLA}, a streaming action decoding framework that addresses both challenges in a unified formulation.
FlashVLA maintains a streaming action buffer with multiple chunks at different noise levels and decodes them using chunk-wise causal attention.

Method

We introduce FlashVLA, a streaming action decoding framework that realizes the central idea from § 1 : decode action chunks jointly rather than in isolation. FlashVLA is a drop-in modification to any flow-matching-based VLA model, requiring only light architectural changes and a fine-tuning pass.

We describe the chunk-wise autoregressive formulation (§ 3.1 ), the streaming inference algorithm (§ 3.2 ), and the multi-buffer joint fine-tuning procedure used to adapt pretrained VLA models (§ 3.3 ).

Method — FlashVLA: Streaming Action Decoding for Fast and Asynchronous VLA Inference — Figure from the paper. Source: FlashVLA: Streaming Action Decoding for Fast and Asynchronous VLA Inference.

Simulated Experiments

We evaluate FlashVLA against three claims it inherits from § 1 and § 3 : (i) at fixed task quality, decoding chunks jointly is substantially faster than decoding them in isolation; (ii) under asynchronous execution, chunk-wise causal attention preserves trajectory continuity without an explicit future-state module; and (iii) the chunk-level memory created by the streaming buffer translates into measurable long-horizon gains.

Real-World Deployment

We deploy FlashVLA on a 7-DoF Franka arm with a single RTX A4000 GPU and ask whether the simulated latency and asynchronous-continuity benefits transfer to a real robot at 30 Hz control.

We evaluate three tasks chosen to span the horizon range observed in simulation: pick and place (short: pick a fruit, place it in a basket), whiteboard wiping (medium: pick up a brush, wipe all marks), and table cleaning (long: clean three objects one by one into a basket).

Real-World Deployment — FlashVLA: Streaming Action Decoding for Fast and Asynchronous VLA Inference — Figure 4: Reaction speed under a 30 Hz target. FlashVLA achieves 1.7$\times$ lower time-to-first-action (TTFA) and 1.6$\times$ lower time-to-react (TTR) than FASTER.

Sources and Demonstrations

Figures are reproduced from Li et al. See the full paper for experimental details and the project page, when the paper links one, for demonstrations and videos.