LLM Generates Robot Control Code Directly: Professor Sim's Team Presents at World's Largest Robotics Conference
- Views 61
- Writer 커뮤니케이션팀
- 보도일자 2026-06-16
A research team led by Professor Joo Yong Sim of the Division of Mechanical Systems Engineering has unveiled a study on using a large language model (LLM) to automatically generate and refine control code for robotic arms. The team presented its findings at ICRA 2026 (the 2026 IEEE International Conference on Robotics and Automation), held June 1–5 in Vienna, Austria.
According to the team, prior LLM-based control-code research focused largely on high-level planning and remained constrained by its reliance on predefined APIs. To address this, the researchers proposed ModuLoop, a framework that automatically generates and executes low-level control code.
ModuLoop comprises a Modular Code Synthesizer, which produces code module by module, and a Closed-Loop Debugger, which corrects errors based on execution results. The system automatically revises its own code to improve performance, achieving high accuracy across a range of task environments without any prior training.
The team validated the framework on hand-eye calibration—precisely aligning the relative positions of a robotic arm and an RGB-D camera—and on pick-and-place operations, in which an object is grasped and moved to a desired location. ModuLoop outperformed conventional methods in both code-generation success rate and task accuracy, while its iterative, feedback-driven debugging delivered fast convergence and strong stability.
The work is significant in extending the LLM's role to that of a direct participant in the robot control loop. Because it can be applied to diverse tasks without task-specific training, the study lays a technical foundation for broad use in robotic automation and intelligent manufacturing.
"By having the system generate and revise code on its own through natural language, we confirmed that more flexible and scalable robot control is possible," said Gina Yoon, a student on the team. "Going forward, we plan to expand the research to handle more complex manipulation tasks and to apply it in real industrial settings."
The paper was co-first-authored by master's students Gina Yoon and Sumin Lee of the Division of Mechanical Systems Engineering, with Professor Sim serving as corresponding author. The findings were published in IEEE Robotics and Automation Letters (RA-L, JCR Impact Factor 5.3), an international journal in the field of robotics.