A domestic research team has developed artificial intelligence (AI) technology that enables robots to perform precision tasks with only a small amount of motion data. A key feature is that the robot can adjust the precision of its movements depending on the work situation.
A team led by Professor Park Dae-hyeong of the KAIST School of Computing said on the 24th it developed a multi-precision manipulation model, DiSPo. The technology is designed to finely generate a robot's movements in line with the level of performance the user requires.
Conventional robot AI needed large-scale motion data recorded at short time intervals to learn tasks such as tightening screws or fitting parts into narrow gaps. Learning precise movements required many demonstrations and long collection times, which created a heavy burden for applying the technology to real industrial sites.
The team introduced an architecture that allows a robot to predict changes in movement while learning diverse behaviors. In particular, it enabled the robot to split motions into finer segments or adjust them relatively coarsely depending on task difficulty.
To do this, the researchers combined Mamba, a state-space model that learns changes over time, with a diffusion model that generates diverse behaviors. As a result, even after training on limited data, the robot can subdivide movements as needed in real tasks to operate with precision.
In simulation, DiSPo showed a task success rate up to 81% higher than the previous state-of-the-art model. In real collaborative robot experiments, it performed tasks such as inserting parts into a narrow 2.5-millimeter gap and pressing a small smartphone shutter button. The success rate was up to four times higher than existing AI models.
The technology can be applied to fields that require high accuracy, including precision parts assembly, cable consolidation, precision machining, and surgical assistance. Because it can train robots with a small amount of data, it is also expected to help reduce data collection expense and expand the scope of automation.
Professor Park Dae-hyeong said, "This study shows that robots can learn precise movements with only a small amount of data and adjust precision according to work conditions," adding, "We will develop it into a robot learning technology that can be used across various sites, including precision manufacturing and healthcare."
The results were presented on the 1st at the International Conference on Robotics and Automation (ICRA 2026) in Vienna, Austria.
References
arXiv (2026), DOI: https://doi.org/10.48550/arXiv.2409.14719