Domestic researchers have developed robotic technology that helps to accurately determine the state of deformed objects with only incomplete visual information. It is expected to contribute to intelligent automation in various industries and service sectors, including cable or wire assembly, handling soft components, and organizing and packaging clothing.
Professor Park Dae-hyung and his research team from the Korea Advanced Institute of Science and Technology (KAIST) announced on the 21st that they have developed artificial intelligence (AI) technology called "INR-DOM" that enables robots to skillfully handle objects that change shape like elastic bands and are visually difficult to distinguish. This research was presented at the top international conference in robotics, "Robotics: Science and Systems 2025," held in Los Angeles, USA, in June.
Technology for robots to handle objects that deform freely, such as wires, clothing, and rubber bands, has long been considered a key task for automation in the manufacturing and service industries. However, these deformed objects have irregular shapes and unpredictable movements, making it challenging for robots to accurately recognize and manipulate them.
The research team utilized "latent neural representation" technology to fully reconstruct the entire shape of deformable objects using only partial 3D information observed by the robot. This method allows the robot to recompose the entire shape of the object, including the unseen parts, based on the information it has observed. Just like humans, robots will also be able to imagine and understand the overall appearance of objects.
Based on this, the team enabled the robot to learn how to manipulate objects. In particular, they introduced a new two-step learning structure that combines reinforcement learning and contrastive learning, allowing the robot to efficiently differentiate between subtle differences in the current state of the object and the target state, leading it to discover the optimal actions required for task completion.
The technology developed was embedded in a robot and tested, resulting in significantly higher success rates in three tasks within a simulation environment: fitting a rubber ring into a groove, installing an O-ring onto components, and untangling twisted rubber bands, compared to existing top-performing technologies. The most challenging task of untangling reached a success rate of 75%, which is about 49% higher than the existing top technology (ACID, 26%).
In real-world environments, the robot also performed fitting, installing, and untangling tasks with success rates exceeding 90%, and it recorded a 25% higher success rate in the visually indistinguishable bidirectional untangling task compared to existing image-based reinforcement learning techniques.
Song Min-seok, a master's student at KAIST and the first author of the research paper, said, "This research demonstrated the possibility for robots to understand the overall appearance of deformed objects with only partial information and perform complex manipulations based on that understanding." He noted that it will greatly contribute to the advancement of robotic technologies that cooperate with or replace humans in various fields, including manufacturing, logistics, and healthcare.
References
arXiv (2025), DOI: https://arxiv.org/abs/2505.00500