Researchers at Ghent University and imec used a locally run AI coding agent to adapt a robot arm to toy-manipulation tasks its instructions did not describe exactly. On a physical UR3e arm, the system completed 30 of 45 trials without an operator intervening. The experiment asks whether a robot can measure a new object and write the code it needs during a task, instead of having a programmer prepare each variation.
The team gave the agent a written procedure for one example in each of three toy games: stacking cups, fitting a block into a shape sorter and placing rings on a rod. A Qwen3.8-27B vision-language model then read camera images, measured objects and wrote short perception and motion programs for new colors, sizes, shapes or arrangements. A separate robot service handled calibrated movement and enforced speed, force and workspace limits; the model could not change those safeguards.
Cup tasks worked more often than shape sorting
Five trials were run for each of nine tasks. The arm finished 20 of 25 cup trials, seven of 15 shape-sorter trials and three of five ring trials without help. Success meant completing the task with no operator intervention, not simply making a grasp. Some failed runs involved the arm misidentifying a hole or object; others ended when a protective stop caught a bad motion. The robot's safety checks tracked the gripper's target point, but did not fully account for the space swept by the gripper body.
Resumed attempts took about half as long
The researchers also resumed 17 sessions that had already succeeded and asked the agent to repeat the same task after moving the objects. These second attempts took about half as long in total as the corresponding first attempts, while retaining the earlier conversation and generated scripts. That is reuse within a continuing session, not evidence that the model learned a lasting new skill.
Reliability beyond the tabletop remains untested
The trials took place on one fixed tabletop with children's toys, one arm and one local model running on a workstation with an RTX PRO 6000 GPU. They did not test different lighting, clutter, materials or robot designs. The result shows that a coding agent can transfer a supplied procedure to some unseen objects on real hardware. Whether it works reliably in varied scenes without resumed sessions or narrowly prepared task documents remains open.