%ef%bb%bf%44%61%65%68%79%75%6e%67%20%50%61%72%6b
-
KAIST Develops Robot Learning Technology Capable of Precisely Imitating Even “Rough” Demonstrations
Robots with increasingly precise dexterity are becoming essential in everyday life and industrial settings, from assembling tiny smartphone components to assisting doctors in surgery. However, teaching robots delicate human movements has traditionally required collecting vast amounts of data at extremely fine time intervals, resulting in significant costs and time burdens. KAIST researchers have developed a robot artificial intelligence technology that can perform sophisticated tasks by autonomously adjusting precision according to the situation, even when trained only on coarsely (sparsely) sampled demonstrations.
KAIST, led by President Kwang Hyung Lee, announced on the 24th that a research team led by Professor Daehyung Park of the School of Computing has developed DiSPo, a multi-granularity manipulation model that generates fine-grained robot motions tailored to a user’s desired level of precision, even from rough human demonstrations.
Existing robot learning methods, such as Behavior Transformer and Diffusion Policy, are limited by their dependence on the time intervals of the data used during training. As a result, learning precision manipulation tasks such as screw fastening or component insertion has required collecting large volumes of high-frequency data at very short time intervals. This has significantly increased data collection costs and slowed down the inference speed of robot AI models.
To overcome these limitations, the research team combined Mamba, a state-space model capable of predicting time intervals, with a diffusion model that enables rich action representation. The team also introduced a new Step-scale factor mechanism, which allows users to directly control the time intervals used by the robot.
As a result, even when trained on only low-frequency (coarse) demonstration data, the robot can generate high-precision motions during inference without additional training by autonomously subdividing actions through a discretization process.
DiSPo achieved up to an 81% higher task success rate compared to state-of-the-art models in simulation environments. In real-world experiments using a collaborative robot, DiSPo stably performed challenging tasks such as passing a clamp through a narrow gap with only a 2.5 mm radial clearance and accurately pressing a small shutter button on a smartphone. This performance was up to four times higher than that of existing AI models.
The technology is expected to make a significant contribution to automation in a wide range of everyday and industrial service fields that require high precision, including precision component assembly, cable connection, medical surgery, and precision machining.
“This study demonstrates that robots can learn precise motions from coarse demonstrations and autonomously adjust their level of precision according to the task situation,” said Professor Daehyung Park. “Moving forward, this technology is expected to dramatically reduce data collection costs while serving as a general-purpose robot learning technology for various industrial fields, including precision assembly and medical applications.”
The study was led by Nayoung Oh, a master’s student at the KAIST Graduate School of AI, as the first author, and was presented on June 1 at the 2026 IEEE International Conference on Robotics and Automation, or ICRA 2026, one of the world’s most prestigious robotics conferences, held in Vienna, Austria.
Paper Title: DiSPo: Diffusion-SSM based Policy Learning for Coarse-to-Fine Action Discretization
DOI: https://doi.org/10.48550/arXiv.2409.14719
2026.06.24 View 1386 -
In KAIST, Robots Now Untie Rubber Bands and Insert Wires Like Humans
The technology that allows robots to handle deformable objects such as wires, clothing, and rubber bands has long been regarded as a key task in the automation of manufacturing and service industries. However, since such deformable objects do not have a fixed shape and their movements are difficult to predict, robots have faced great difficulties in accurately recognizing and manipulating them. KAIST researchers have developed a robot technology that can precisely grasp the state of deformable objects and handle them skillfully, even with incomplete visual information. This achievement is expected to contribute to intelligent automation in various industrial and service fields, including cable and wire assembly, manufacturing that handles soft components, and clothing organization and packaging.
KAIST (President Kwang Hyung Lee) announced on the 21st of August that the research team led by Professor Daehyung Park of the School of Computing developed an artificial intelligence technology called “INR-DOM (Implicit Neural-Representation for Deformable Object Manipulation),” which enables robots to skillfully handle objects whose shape continuously changes like elastic bands and which are visually difficult to distinguish.
Professor Park’s research team developed a technology that allows robots to completely reconstruct the overall shape of a deformable object from partially observed three-dimensional information and to learn manipulation strategies based on it. Additionally, the team introduced a new two-stage learning framework that combines reinforcement learning and contrastive learning so that robots can efficiently learn specific tasks. The trained controller achieved significantly higher task success rates compared to existing technologies in a simulation environment, and in real robot experiments, it demonstrated a high level of manipulation capability, such as untying complicatedly entangled rubber bands, thereby greatly expanding the applicability of robots in handling deformable objects.
Deformable Object Manipulation (DOM) is one of the long-standing challenges in robotics. This is because deformable objects have infinite degrees of freedom, making their movements difficult to predict, and the phenomenon of self-occlusion, in which the object hides parts of itself, makes it difficult for robots to grasp their overall state.
To solve these problems, representation methods of deformable object states and control technologies based on reinforcement learning have been widely studied. However, existing representation methods could not accurately represent continuously deforming surfaces or complex three-dimensional structures of deformable objects, and since state representation and reinforcement learning were separated, there was a limitation in constructing a suitable state representation space needed for object manipulation.
To overcome these limitations, the research team utilized “Implicit Neural Representation.” This technology receives partial three-dimensional information (point cloud*) observed by the robot and reconstructs the overall shape of the object, including unseen parts, as a continuous surface (signed distance function, SDF). This enables robots to imagine and understand the overall shape of the object just like humans.
*Point cloud 3D information: a method of representing the three-dimensional shape of an object as a “set of points” on its surface.
Furthermore, the research team introduced a two-stage learning framework. In the first stage of pre-training, a model is trained to reconstruct the complete shape from incomplete point cloud data, securing a state representation module that is robust to occlusion and capable of well representing the surfaces of stretching objects. In the second stage of fine-tuning, reinforcement learning and contrastive learning are used together to optimize the control policy and state representation module so that the robot can clearly distinguish subtle differences between the current state and the goal state and efficiently find the optimal action required for task execution.
When the INR-DOM technology developed by the research team was mounted on a robot and tested, it showed overwhelmingly higher success rates than the best existing technologies in three complex tasks in a simulation environment: inserting a rubber ring into a groove (sealing), installing an O-ring onto a part (installation), and untying tangled rubber bands (disentanglement). In particular, in the most challenging task, disentanglement, the success rate reached 75%, which was about 49% higher than the best existing technology (ACID, 26%).
The research team also verified that INR-DOM technology is applicable in real environments by combining sample-efficient robotic reinforcement learning with INR-DOM and performing reinforcement learning in a real-world environment.
As a result, in actual environments, the robot performed insertion, installation, and disentanglement tasks with a success rate of over 90%, and in particular, in the visually difficult bidirectional disentanglement task, it achieved a 25% higher success rate compared to existing image-based reinforcement learning methods, proving that robust manipulation is possible despite visual ambiguity.
Minseok Song, a master’s student and first author of this research, stated that “this research has shown the possibility that robots can understand the overall shape of deformable objects even with incomplete information and perform complex manipulation based on that understanding.” He added, “It will greatly contribute to the advancement of robot technology that performs sophisticated tasks in cooperation with humans or in place of humans in various fields such as manufacturing, logistics, and medicine.”
This study, with KAIST School of Computing master’s student Minseok Song as first author, was presented at the top international robotics conference, Robotics: Science and Systems (RSS) 2025, held June 21–25 at USC in Los Angeles.
※ Paper title: “Implicit Neural-Representation Learning for Elastic Deformable-Object Manipulations”
※ DOI: https://www.roboticsproceedings.org/ (to be released), currently https://arxiv.org/abs/2505.00500
This research was supported by the Ministry of Science and ICT through the Institute of Information & Communications Technology Planning & Evaluation (IITP)’s projects “Core Software Technology Development for Complex-Intelligence Autonomous Agents” (RS-2024-00336738; Development of Mission Execution Procedure Generation Technology for Autonomous Agents’ Complex Task Autonomy), “Core Technology Development for Human-Centered Artificial Intelligence” (RS-2022-II220311; Goal-Oriented Reinforcement Learning Technology for Multi-Contact Robot Manipulation of Everyday Objects), “Core Computing Technology” (RS-2024-00509279; Global AI Frontier Lab), as well as support from Samsung Electronics. More details can be found at https://inr-dom.github.io.
2025.08.21 View 5505