FOMI researchers report faster collection of robot training demonstrations

Researchers behind the FOMI preprint report that their handheld interface improves demonstration collection efficiency by approximately 29% without a robot and 18% during teleoperation, compared with the best baselines across four tasks involving physical contact. A scissor-shaped mechanism passes forces from object interactions directly to the operator's hands through mechanical links without force sensors, helping operators adjust their actions. The system combines calibration, data acquisition and synchronization, and can translate interface movement and gripper opening into robot commands. Demonstrations collected in both modes supported learned manipulation policies; the efficiency comparisons cover four tested tasks.