Sim-to-real RL controls microfibers with 270μm precision

A team led by Aalto University researchers has introduced a closed-loop sim-to-real reinforcement learning method for controlling deformable microfibers — a notoriously hard problem due to unpredictable surface and interfacial forces at the microscale. The key innovation: they train a policy entirely in a simplified frictionless simulator, then deploy it directly onto a physical dual-gripper micromanipulation system operating at 40 Hz. No retraining, fine-tuning, or domain adaptation is required. Instead, real-time visual feedback during deployment corrects for unmodeled interactions, enabling the system to iteratively shape silk microfibers with sub-millimeter accuracy.

The policy achieved a mean point-wise shape error of 270 ± 80 μm across 24 diverse initial configurations. It handled nine test specimens covering three fiber diameters (50, 80, 120 μm) and three manipulated lengths (10, 15, 20 mm) without any retuning. This demonstrates that simplified simulation combined with closed-loop feedback can bridge the sim-to-real gap for delicate micromanipulation tasks. The work opens the door to autonomous manufacturing of microdevices, biomedical assembly, and fiber-optic alignment where precise contact-based control is critical.