From Reach to Insert: Tactile-Augmented Precision Assembly under Sub-Millimeter Tolerances

A team of researchers has developed a tactile-augmented system for high-precision assembly tasks that require sub-millimeter clearances—a challenge where even tiny pose errors can cause jamming or excessive forces. Their two-stage approach separates the problem into a reaching phase and an insertion phase. First, Imitation Learning (IL) trains a policy that grasps the peg and positions it near the target hole, handling position variances. Then, Reinforcement Learning (RL) takes over for the insertion itself, using trial-and-error with contact-rich interactions to insert the peg and recover from failures. To improve RL performance, the researchers introduced tactile group sampling to focus training on the most critical contact regions, along with a tactile critic that more accurately evaluates policy values from force/torque signals.

The method was tested across five different hole geometries and three clearance settings (from loose to ultra-tight 0.05mm). At the most demanding 0.05mm clearance—roughly half the thickness of a human hair—the system achieved a 67% insertion success rate while keeping contact forces low. Crucially, it reduced maximum interaction force by 60% and torque by 44% compared to baselines, meaning safer operation for both the robot and the assembled parts. These results validate that combining IL for coarse positioning with RL for fine, contact-aware control—augmented by specialized tactile feedback—can dramatically improve both reliability and safety in precision manufacturing.