KAIST develops robot AI model capable of precision assembly with minimal training data

Robots on factory floors may soon be able to learn precise assembly tasks from a single demonstration by a human worker, requiring far less training data than current systems demand.

KAIST announced Wednesday that a research team led by Park Dae-hyung, a professor in the School of Computing, has developed DiSPo, a robot AI model that trains on minimal data and autonomously adjusts its own precision level to suit the task at hand — enabling it to execute highly delicate movements.

Conventional robot AI systems require vast amounts of data — recorded at extremely short time intervals — to perform precise tasks. Learning to tighten a screw or insert a component into a narrow gap, for instance, demanded painstaking collection of enormous volumes of motion data, making the process costly and time-consuming.

To overcome these limitations, the research team developed AI technology that allows robots to learn a wide range of actions by predicting changes in movement on their own. A key new feature lets the robot subdivide or scale its movements depending on the demands of the task.

The team achieved this by combining Mamba, a state-space model that efficiently learns temporal changes, with a diffusion model capable of generating diverse behaviors.

Even when trained on a small amount of data, the robot can break down its movements into finer units during actual task execution to operate with greater precision. A relatively simple human demonstration is enough — the robot can then subdivide its own actions as needed to carry out highly precise work.

In simulation environments, DiSPo recorded a task success rate up to 81 percent higher than the best-performing existing models. In experiments using an actual collaborative robot, it also reliably completed highly demanding tasks — inserting a component into a gap just 2.5 millimeters wide and accurately pressing the small shutter button on a smartphone. That represents a success rate up to four times higher than existing AI models.

The technology is expected to find applications across industries requiring high accuracy, including precision parts assembly, cable connection, medical surgery and precision machining. Because it can train high-precision robots on minimal data, it is also projected to dramatically cut robot development costs and accelerate automation in the manufacturing, medical and service sectors.

"Even without a human providing repeated, highly detailed demonstrations, the robot can fill in the gaps in its training data on its own and improve its performance," Park said. "Going forward, we will develop this into a general-purpose robot learning technology that dramatically reduces data collection costs while remaining applicable across diverse industrial settings, including precision manufacturing and medicine."