What Makes Contact-Rich Tasks Hard
Contact-rich manipulation — insertion, screwing, polishing, peg-in-hole, cable routing — involves sustained, force-sensitive contact between the robot and objects. Small position errors lead to jamming or excessive forces. The contact dynamics are discontinuous, hard to simulate accurately, and highly sensitive to geometric tolerances.
Classical vs Learned Approaches
Classical approaches use impedance control with spiral search patterns and force thresholds. They work reliably for known geometries but require extensive per-task tuning. Learned approaches (behavior cloning, RL with force observations) can handle more variation but need high-quality demonstration data with force-torque sensing. The best current results combine both: use learning for high-level strategy and classical control for low-level force regulation.
Data Collection for Contact Tasks
Recording force-torque data alongside visual observations is essential. Use leader-follower teleoperation for intuitive force transmission. Filter demonstrations where the operator applies excessive corrective forces. SVRC's OpenArm with PaXini tactile sensors provides the sensor suite needed for contact-rich data collection.