On August 1, 2026, a team from MIT, with collaborators from Nvidia, Tsinghua and three U.S. universities, published VLASH, a method that removes the pauses between a robot's moves when it is driven by a vision-language-action (VLA) model. Instead of waiting for the next computation, the robot plans its next action while still executing the previous one. According to the arXiv paper, this cuts reaction latency by up to 11.8x, with no changes to the model architecture.
Key takeaways
- VLASH is an asynchronous inference method for VLA models — full name: Real-Time VLAs via Future-State-Aware Asynchronous Inference.
- Reaction latency drops by up to 11.8x versus classic synchronous inference (per the arXiv paper).
- With action quantization, tasks finish 1.5–2.0x faster, and MIT reports reaction speedups of up to 30x.
- A cube-sorting test held 90% success while fine-tuning time was cut fivefold.
- The code is public on GitHub (mit-han-lab/vlash), and the paper is slated for the IROS conference.
The problem: a robot that freezes when it thinks
VLA models control a robot by turning a camera image and a natural-language command into a sequence of moves. Their weakness lies in how they run: classic synchronous inference?Synchronous inference: The model only computes the next chunk of moves once the robot finishes the previous one — so the robot has to stop and wait. forces the robot to stop and wait while the model computes the next chunk of actions. With large models that pause is visible — the robot moves in jerks, and fast tasks become out of reach.
VLASH targets exactly this temporal gap between prediction and execution. Instead of computing the next move from where the robot is now, the method predicts where the robot will be once the current action finishes, and uses that future state to prepare the next chunk of moves.
The method's name breaks down into three ideas:
Our motivation was to overlap the thinking process with the execution process to make the reaction speed faster.
Jiaming Tang, MIT graduate student and co-author of the paper.
The results, in context
The strongest result is exactly this latency. MIT frames it more broadly, citing reaction speedups of up to 30x in some tests, while matching — and sometimes improving — accuracy relative to earlier asynchronous methods.
The remaining numbers show the gain does not come at the cost of quality. Combined with action quantization?Action quantization: Compressing the motion representation into fewer values — less data to compute at each step., tasks finish 1.5–2.0x faster, and in the cube-sorting test the robot held 90% success at double the pace. MIT also reports a fivefold cut in fine-tuning time, with no extra compute required.
The method was tested on robotic arms doing stacking, sorting and pick-and-place, as well as reflex-heavy tasks — ping-pong and whack-a-mole — that were practically unreachable under classic inference.
How it compares
Speeding up VLA robots is a busy research front. We previously covered the PolicyTrim framework, which cut task execution time by up to 5.83x in simulation — but by pruning the control policy itself. VLASH works differently: it changes neither the model nor the architecture, only how inference interleaves with motion. That makes it complementary to compression techniques rather than a rival — the two can be combined.
The work comes from the group of Song Han, an associate professor in MIT's Department of Electrical Engineering and Computer Science known for research on efficient inference. Co-authors include Jiaming Tang (MIT) and Yufei Sun (Tsinghua), with funding from the MIT-IBM Computing Research Lab, Amazon, the National Science Foundation and Nvidia.
Why it matters
Latency is today one of the main barriers to moving large VLA models out of the lab and into real-time tasks. A method that cuts latency by an order of magnitude without changing the model or adding hardware lowers the entry bar for bigger, more capable control policies. What matters is that VLASH needs no architectural redesign — it can be layered onto existing models. If the results hold beyond a narrow task set, reaction speed will stop being a reason to pick smaller, weaker models just because they are faster.
What's next?
- The VLASH paper is slated for presentation at the IROS (Intelligent Robots and Systems) conference.
- The code is already public on GitHub (mit-han-lab/vlash), allowing independent verification of the claimed results.
- Scale remains an open question: tests covered robotic arms, not full humanoids with locomotion, where the cost of a wrong state prediction is higher.
