This Robot Fits Parts Without a Camera. Its Joints Guide It.

The peg is already in the robot's gripper, touching the surface near its matching hole. The arm presses, searches, and adjusts until the part slips into place. Its controller never sees the opening. It works from measurements of the robot's own movement.

In a new MIT preprint, one controller completed 173 of 237 physical insertion trials across eight 3D-printed pegs. That is 73%, with gaps of 0.5 to 1.5 millimeters on each side. An operator judged whether each peg was seated; reaching the 20-second limit counted as a failure.

The interesting result is not that cameras have become unnecessary. It is that a robot can extract useful information about a tight fit from sensors it already uses to control its joints.

Contact travels through the arm

When a peg meets an edge, the contact changes how the arm moves. The joints do not follow their commands exactly as they would in free space. Their position sensors record those differences.

The researchers trained a controller to use a rolling 0.6-second history of joint positions, speeds, and previous commands, including the gripper motion calculated from them. That history helps distinguish moving freely from pushing against a surface or getting stuck. It is called proprioception: sensing the body's own state.

No camera image, force reading, torque reading, or tactile-sensor input enters the controller. This is not a sensorless robot. The sensing happens in the joints, rather than through an added view of the workpiece or a touch-sensitive fingertip.

Nor is the arm simply repeating a spiral until something fits. The paper's analysis shows that its learned movement history carries information about contact states and affects the next movement. The search responds to what happens.

The practice holes had to be real enough

The controller learned in simulation, then transferred to a Flexiv robot arm without being fine-tuned on physical insertion attempts. The team did use real measurements to identify aspects of the arm's dynamics for its model; simulation training does not mean the hardware was irrelevant.

Training varied the parts, grip, alignment, and contact conditions. Instead of practicing one perfect approach to one familiar opening, the controller encountered many ways a fit could begin badly. Three of the eight physical test pegs also went beyond its training distribution: one had a tighter gap, one a larger cross-section, and one a taller handle.

There was a surprisingly basic obstacle to creating those practice worlds. A simulator's approximation of a hole can put material where the narrow gap should be. A virtual peg then jams for a reason the physical part would not.

The team's open-source geometry generator, mjfit, constructs the contact surfaces so the intended clearance remains open. Better learning required a faithful place to practice, not just more attempts. That is a different use of simulation from Google's VERID lab system, which checks proposed laboratory actions before carrying them out.

A finishing skill, not a blind factory

The starting position does important work. Every physical trial begins with the part already held and touching the fixture near the opening. Another action must first bring it there. If the peg falls far from the hole, this local controller cannot retrieve it and restart the job.

The paper also reports gear meshing and nut threading, but those results are in simulation. So are its direct comparisons with camera-based controllers. The blind approach coped better with unfamiliar shapes and sizes in those tests, while vision could make insertion faster. There was no physical factory contest establishing that robots generally work better without cameras.

And 73% is a research result, not production-ready reliability. The work, submitted on October 16 and first publicly announced on arXiv on October 19 in Europe, is a preprint rather than an independently replicated industrial deployment.

Other approaches improve the training examples, as in robot learning from difficult human demonstrations. This experiment finds useful feedback inside the machine itself. A camera may help an arm reach the work; its joints may supply the information needed to finish the fit.