Multi-Buffer Joint Fine-Tuning
Pretrained vision-language-action models are trained to denoise each chunk from scratch in isolation and have never seen a partially populated buffer or a chunk-wise causal attention pattern. FlashVLA adapts them to the streaming paradigm during the standard downstream fine-tuning stage.
The training objective is the standard flow-matching loss summed over all buffer configurations.
The model minimizes flow-matching loss across every valid combination of buffer contents and noise levels.
This single objective covers both cold-start and steady-streaming behavior: the model learns to denoise chunks at any noise level conditioned on any valid buffer prefix.
Real-World Deployment
FlashVLA is deployed on a 7-DoF Franka arm with a single RTX A4000 GPU. The deployment evaluates whether the simulated latency and asynchronous-continuity benefits transfer to a real robot operating at a 30 Hz control frequency.
Conclusion
FlashVLA adapts pretrained vision-language-action models through multi-buffer joint fine-tuning. This makes it a drop-in modification, but it inherits the pretrained model’s independent-chunk training objective.
Pretraining a vision-language-action model from scratch under the chunk-wise causal formulation is a natural next step and is likely to yield further gains.
The streaming buffer also requires a one-time cold-start of a fixed number of warm-up steps per episode. The cost amortizes quickly over multi-second rollouts but is more noticeable on very short tasks.
Acknowledgment
The authors gratefully acknowledge Professor Xiaolong Wang’s lab for providing a Franka robot arm.
Real-World Experiment Settings
The fine-tuned model is deployed on an RTX A4000 at a 30 Hz control frequency. All methods use the same CUDA Graph and kernel-fusion system optimizations.
The execution horizon is set to 16, and the asynchronous delay is set to two steps for all asynchronous methods. Each task score is averaged over 15 rollouts, while completion time is averaged over successful rollouts only.
Frequently Asked Questions
What is FlashVLA? FlashVLA is a streaming action-decoding approach for faster and asynchronous vision-language-action inference.
How is FlashVLA adapted to pretrained models? It uses multi-buffer joint fine-tuning during the standard downstream fine-tuning stage.
What robot hardware is used for real-world deployment? The system is deployed on a 7-DoF Franka arm with a single RTX A4000 GPU at a 30 Hz control frequency.
What limitation does the streaming buffer introduce? The streaming buffer requires a one-time cold-start warm-up during each episode, which is more noticeable on very short tasks.