MIT's VLASH Makes Robots 30x Faster by Letting Them Think Ahead
A new framework from Song Han's lab overlaps AI reasoning with physical execution, enabling robots to handle dynamic tasks like ping-pong in real time.
MIT researchers have cracked a fundamental bottleneck in robotics: AI brains that think too slowly for their bodies to move fluidly. The new VLASH system lets robots predict their own future state and plan the next move while still executing the current one—delivering more than 30-fold faster reaction speeds.
Vision-Language-Action models have long processed visual inputs and generated motion commands in discrete chunks, creating computational pauses that produce jerky, delayed responses. VLASH eliminates this inference lag by overlapping thinking with execution, enabling robots to handle fast-paced, dynamic environments previously out of reach.
Thinking While Moving
The core innovation addresses what the team calls the misalignment problem. Previous attempts to predict future moves failed because the environment changes between observation and action. VLASH sidesteps this by training the VLA model to prioritize future state information over current observations.
"Our motivation was to overlap the thinking process with the execution process to make the reaction speed faster," said Jiaming Tang, an MIT EECS graduate student and co-lead author on the work.
The researchers developed a training-augmentation method that reuses existing training data to teach this predictive capability. This fine-tuning approach accelerated training fivefold with no additional computational overhead.
Action Quantization Trade-offs
VLASH also employs action quantization, generating coarser chunks of motion commands rather than fine-grained instructions. This technique enables robots to complete overall tasks 2-3 times faster, though with a slight dip in accuracy.
In real-world testing, the system sorted colored cubes twice as fast as baseline methods while maintaining 90% accuracy. More impressively, the framework enabled robots to perform highly dynamic tasks including playing ping-pong and whack-a-mole.
Toward Human-Like Fluidity
The work bridges a critical gap between the slow reasoning of large AI models and the real-time agility required for physical tasks. Modern VLAs create pauses between action chunks that make robots appear hesitant and unresponsive to environmental changes. VLASH's predictive approach moves robots closer to human-like fluidity in complex, fast-changing settings.
"This work sets up a good foundation for efficient, fast, accelerated, and low-cost robotics applications," said Song Han, associate professor at MIT EECS and senior author on the research.
The team presented their findings at the Intelligent Robots and Systems Conference. The paper lists Tang and Yufei Sun as co-lead authors, with Han as senior author. Code and project details are available through MIT's Han Lab repository.
For robotics to move beyond controlled industrial settings into dynamic human environments, reaction speed matters as much as reasoning capability. VLASH demonstrates that the two need not be mutually exclusive.