AI helps robots reduce latency by 17 times and work twice as fast

(Dan Tri) - A new AI method allows robots to both perform tasks and calculate the next step, thereby reducing response latency by a maximum of 17.4 times and helping some tasks complete about 2 times faster.

Research conducted by scientists at MIT, Nvidia, Caltech, University of California at Berkeley, University of California at San Diego and Tsinghua University, introduces a system called VLASH (Vision-Language-Action with Scheduled Heuristics).

This method aims to overcome a common limitation of vision-language-action (VLA) models: the robot must wait for the AI ​​to calculate the next set of commands after completing the previous set of actions, causing interrupted movements.

Predict robot state

VLASH allows the inference process to take place at the same time as the robot performs the task. The system relies on a planned sequence of actions to estimate the state the robot will reach, thereby preparing the next set of actions before the current set ends.

It is possible to imagine the robot performing this step but at the same time preparing for the next step, instead of waiting for completion before starting to calculate.

What's notable about VLASH is that the system doesn't try to predict the entire future environment. Instead, the AI ​​estimates the robot's own future state based on its current position and planned motion sequence.

This approach reduces computational time compared to building a model that predicts the entire environment. According to the research team, VLASH does not need a separate prediction model during operation.

The research team believes that this method can also be combined with world models in the future to help robots better handle complex changes in the environment.

Robots are faster, but that doesn't mean they're safer

VLASH was tested on a variety of manipulation tasks such as picking, placing, stacking and sorting objects. In one test, the robot was able to sort cubes by color and place them in boxes about twice as fast as the control method, with about 90% accuracy.

When combined with action quantization techniques, which combine many small movements into larger steps, VLASH achieves a maximum speedup of 2.03 times on real robots while maintaining the same accuracy.

The system was also tested on tasks that require quick reactions such as table tennis and Whac-a-Mole, where the target can change while the robot is performing a sequence of actions.

The research team also improved the training process. With the π0.5 model, the execution time of each training step is reduced from about 421 to 129 milliseconds, 3.26 times faster. However, this number only reflects the time of each step, not a corresponding reduction in total training time or cost.

Despite significantly reducing latency, VLASH has not proven that robots will be safer when operating in environments with humans. Factors such as reliability, behavioral predictability and safety mechanisms still need to be independently assessed.

The system has not been fully tested in harsh conditions such as low light, dust, complex terrain or when the sensor operates unstable.

According to the research team, VLASH has potential applications in tasks that require robots to move continuously and react quickly, from manufacturing to frequently changing environments.

The current results mainly show a way to reduce the time the robot has to wait for the AI ​​to make inferences. The ability to maintain this effectiveness in diverse real-world environments remains to be further verified.