Simultaneous Learning of Bimanual Cooperation


Advancing Humanoid Robot Research

Gwangju Institute of Science and Technology (GIST) announced on April 9 that the research team led by Professor Lee Kyubin from the Department of AI Convergence has developed an "AI bimanual robot grasping technology" that enables robots to coordinate both arms like a human, maintain balance, and stably grasp and manipulate objects.


The core of this technology is that the robot autonomously learns to resolve instability caused by collisions or loss of force balance between the two arms when grasping objects, thereby achieving a stable posture. This advancement is expected to contribute to the development of humanoid robots that use their upper bodies to manipulate objects in a manner similar to humans.

Model architecture of the BiGraspFormer bimanual grasping robot system. Provided by GIST

Model architecture of the BiGraspFormer bimanual grasping robot system. Provided by GIST

View original image

Recently, there has been growing demand for "Physical AI"—AI that performs physical tasks—in various industrial fields such as logistics, manufacturing, and healthcare. As a result, "dual arm robots" capable of stably grasping and handling objects with both arms are receiving significant attention. In particular, for large or heavy objects such as furniture, handling them with only one arm is difficult, making technologies that enable both arms to cooperate and stably grasp the object essential.


However, bimanual grasping requires consideration of the position, orientation, and force of both arms, resulting in a vast number of possible combinations and necessitating much more complex control to ensure the arms do not collide and that forces are evenly distributed on the object. Previous research calculated the movements of each arm separately and then combined them, which often led to insufficient cooperation between the arms, resulting in collisions or imbalanced forces, and thus posed limitations for stable bimanual cooperation in real-world environments.


To address these issues, the research team developed an AI model, the "BiGraspFormer bimanual grasping robot system," which considers both arms as a single integrated system from the outset. This technology adopts an end-to-end approach, processing the entire sequence from input to result in one step, enabling simultaneous learning of arm cooperation without dividing the process into separate stages.


This system operates in three stages, mirroring the process a human uses to pick up an object. First, a camera is used to perceive the target object as a point cloud composed of numerous points, allowing the system to analyze both the overall shape and surface structure simultaneously. This is analogous to how a person visually assesses where to grasp an object before picking it up.


Next, the system identifies various positions that can be stably grasped using one arm and then uses these as "guides" to efficiently generate combinations where both arms can grasp the object together. This significantly reduces the number of possible cases and enables fast computation. Finally, the system selects the optimal combination in which the two arms do not interfere with each other and the forces applied to the object are balanced, thereby determining the most stable grasping posture.

Virtual Environment Evaluation Setup Diagram. Provided by GIST

Virtual Environment Evaluation Setup Diagram. Provided by GIST

View original image

The research team validated the performance of the technology through both virtual environment simulations and real-world robot experiments. In virtual experiments using 100 objects of various shapes, including chairs, shelves, and storage boxes, the system achieved a grasp success rate of 89.67%, representing an improvement of about 18 percentage points over previous technologies. Even under conditions involving external impacts, it maintained a success rate of 59.72%, about 23 percentage points higher than previous methods, confirming that the technology can operate stably in real-world settings.


Furthermore, even when the AI model trained in a virtual environment was applied to an actual robot without any additional training, it achieved an average grasp success rate of over 88% across a variety of real-world objects, ranging from heavy stepped structures to irregularly shaped chairs, demonstrating its potential for field application, versatility, and stability.


Professor Lee Kyubin stated, "This study is the first to demonstrate an integrated bimanual grasping system that learns cooperative arm movements in a unified manner and can be directly applied from a virtual environment to a real robot without additional training." He also expressed expectations that, "Going forward, the practical use of dual arm robots in areas such as large furniture transport, warehouse automation, and heavy-duty assembly in manufacturing sites—fields where single-arm robots have been limited—can be significantly advanced."


This research, supervised by Professor Lee Kyubin and conducted by Kim Kangmin, a doctoral candidate (first author), was supported by the Artificial Intelligence Graduate School Support Project of the Ministry of Science and ICT and the Institute for Information and Communications Technology Planning and Evaluation (IITP), as well as the Open Simulator Core Technology Support Project for Intelligent Application Development of the National Research Foundation of Korea.



The research results will be presented at "ICRA 2026," the world’s most prestigious robotics conference, to be held in Vienna, Austria, from June 1 to 5. The study was also pre-published on the international academic server "arXiv" on September 23, 2025.


This content was produced with the assistance of AI translation services.

© The Asia Business Daily. All rights reserved. Unauthorized AI training and use prohibited.

Today’s Briefing