A robot arm inserted 173 of 237 pegs in a real-world test using only the motion of its own joints to find each hole. The MIT team trained the policy in simulation, without camera images, force readings or tactile sensors; it succeeded on eight peg shapes, including three that were outside the training set.
The result suggests a robot can use the small changes in its joint motion when parts touch as a practical contact signal. It does not show that vision can be removed from an entire assembly line: each trial began with the peg already touching the fixture near the hole.
Joint motion becomes a contact cue
Robot arms already measure each joint's position. When a held part bumps a surface, that contact changes how the arm moves in response to its command. The policy learns to read those changes, a form of proprioception, instead of looking at the part through a camera or sensing contact with a force or touch sensor.
In the study, a seven-joint arm's policy used a 0.6-second history of joint positions and speeds, the gripper pose calculated from those readings, and its previous action. It was not given the hole's location, the part's shape or a force estimate. Reinforcement learning trained the policy entirely in simulation while the team varied the parts, grasps, contact properties and fixture positions.
One simulated skill transfers to real peg insertion
The researchers tested the frozen peg-insertion policy on eight 3D-printed shapes with 0.5 to 1.5 millimeters of clearance per side. It succeeded in 173 of 237 physical trials, or 73 percent, with no real-world fine-tuning. Three shapes were outside the training distribution, including a tighter fit, a wider cross-section and a taller handle.
In simulation, matched policies using wrist-camera images generalized less reliably to unseen part shapes and sizes, although the joint-sensing policy took longer to seat a peg. The same approach also learned gear meshing and nut threading in simulation; the paper reports a physical transfer test for pegs, not those other tasks.
The skill starts near the hole
The policy is designed for the last part of assembly, after another process has brought the peg to the fixture. It starts with the part already touching the surface near the opening, then searches locally by sliding and adjusting its motion. The researchers say it cannot recover if the peg slips from the gripper and lands far from that area.
That makes the study a bounded demonstration of contact-guided manipulation, not evidence that a factory robot can find, pick up and assemble arbitrary parts without other sensors or planning. It also shows a useful design option for tightly specified tasks: simulation-trained control can use existing joint encoders to close the final alignment gap.
Related coverage: A separate study on reusing imperfect robot demonstrations focused on which training data helps a connector-insertion task.
Sources
- Bronars, Park, Kedia and Agrawal, “Blind Precise Assembly,” arXiv:2610.23849 (Oct. 16, 2026) — preprint; full paper and CC BY 4.0 license.
- Authors' project page — task demonstrations and result details.
- Image: Figure 1 by Antonia Bronars, Younghyo Park, Kushal Kedia and Pulkit Agrawal, arXiv:2610.23849v1, CC BY 4.0.