New AI technique helps robots complete tasks twice as fast by letting them 'think ahead'

Scientists have engineered a new artificial intelligence (AI) system that cuts reaction delays in robots by over 10 times without using additional computing power.

Many robots controlled by vision language action (VLA) models move in fits and starts: reach, pause, adjust, pause again.

The problem largely stems from how these models are typically run: After completing one set of instructions, the robot waits for the model to calculate the next set, creating a stop-and-go rhythm.

This reaction lag is a major barrier to using VLA-controlled robots for tasks that demand continuous, real-time interaction.

A new system called "VLASH" tries to eliminate that wait by planning the next actions while the robot completes its current ones.

In tests, robots using VLASH completed some tasks 1.5 to 2 times faster while retaining most or all of their accuracy.

Maximum reaction latency fell by up to 11.8 times, depending on the computer hardware used.

Researchers from MIT; Nvidia; Caltech; the University of California, Berkeley; the University of California, San Diego; and Tsinghua University in China described the system in a paper uploaded to the arXiv preprint server and will present at the Intelligent Robots and Systems Conference this fall.

An earlier version of the paper reported a speedup of more than 30 times, but the researchers told Live Science that those tests used longer action sequences and slower hardware.

The newer experiments used shorter, more common action sequences and more powerful GPUs.

"The 11.8x figure may better represent recent practical settings," the study authors told Live Science in an email.

Robots are getting a brain upgrade VLA models act as the brains of many advanced robots.

They combine images from a camera with human instructions and information about the robot's current state, and then translate that information into physical movements.

As the robot completes one group of actions, VLASH uses its current position and scheduled movements to estimate where it will finish.