KAIST announced that a team led by professor Park Dae-hyung has developed a robot AI model called DiSPo that learns from only a few sparse human demonstrations yet performs ultra-precise motions such as inserting a part into a gap with a 2.5-millimeter radial clearance. DiSPo raised task success rates by up to 81 percent over the previous best-performing model in simulation, recorded up to a fourfold higher success rate in real collaborative-robot experiments, and was presented at ICRA 2026, the International Conference on Robotics and Automation, held in Vienna, Austria on June 1, 2026. This article organizes the primary figures from KAIST's announcement and Korean reporting, and weighs both the problem DiSPo solves and its limits.

Making Precise Motions From a Single Coarse Demonstration

The problem DiSPo targets is a longstanding dilemma in robot learning. Until now, for a robot to learn a precise task it needed vast demonstration data recorded by humans at very short time intervals, and that collection process was expensive and slow. DiSPo instead takes sparse, roughly shown low-frequency demonstrations as input and, at the moment of the task, generates dense precise motions on its own. In effect, the model bridges the gap between a coarse demonstration and precise execution.

A Design That Links a Diffusion Model to a State-Space Model

DiSPo's name refers to multi-precision manipulation, and its core is a structure linking two kinds of neural networks. A state-space model, Mamba, predicts time intervals to gauge when and how densely to move, while a diffusion model generates the actual robot motion. A Step-scale factor that lets the user tune the strength of precision is added on top, so the same trained model can pick coarse or fine motions to suit the situation. Treating precision as a control value at inference time rather than as data is the axis of this design.

The Real Bottleneck Was Data-Collection Cost

DiSPo matters less for its success-rate numbers than for the bottleneck it targets. In robot manipulation research, the surest way to raise performance was more and denser demonstration data, but the cost of humans producing that data one piece at a time has blocked real-world adoption. Rather than adding data, DiSPo turns toward squeezing more out of less, shifting the barrier to precise tasks from data collection to model design. That carries large practical significance for small and mid-sized manufacturing sites or Korean labs that struggle to build large demonstration infrastructure, widening their access to precise robot learning.

How to Read the 81 Percent and the Fourfold Gain

Reading the numbers requires distinguishing two settings. The up-to-81-percent improvement is a value compared against the previous best model in simulation, while the up-to-fourfold success rate came from moving to a real collaborative robot. In robotics, a simulation edge often fails to reproduce on physical hardware, so it is notable that DiSPo showed improvement on concrete real tasks such as 2.5-millimeter-gap insertion and pressing a smartphone shutter button. Still, both figures are relative values tied to specific tasks and baselines, so absolute success rates and generalization to other tasks must be checked separately.

Open Questions When Moving to the Real World

DiSPo showed strength on a narrow band of precise assembly tasks, but clear questions remain. The physical validation confirmed in the announcement covers specific tasks like part insertion and button pressing, and whether it extends to soft objects or tasks mixed with hard-to-predict environmental change needs separate verification. How broad a class of tasks the approach of reconstructing precise motion from low-frequency demonstrations holds for, and whether Step-scale tuning can find optimal values without human intervention, are also open problems. Even so, the direction of preserving precision while cutting the burden of data collection carries clear meaning as an attempt to change the cost structure of robot learning.

Source: AI Times · MIT Technology Review Korea