Robot feels its way into a tight busbar fit at Foxconn
Lukas Ziegler posts a booth demo pairing 3D pose estimation with real time force sensing on a production line insertion job.
A robot at Foxconn days in Taiwan is doing busbar assembly, sliding rigid metal conductors into tight fits without jamming or scratching the part. Lukas Ziegler posted the writeup from the booth, calling it a job that sounds trivial and is not.
Vision gets the gripper to the right spot, but the last millimeters of a tight fit are a contact problem, and that’s where the force sensing earns its place.
Busbars are the metal conductors that carry current inside electrical and battery systems. They are stiff, the tolerances are close, and there are, as Ziegler puts it, plenty of ways for an insertion to jam or scratch the workpiece. Nudge one in slightly off axis and you either bind the part or mark it.
Sight, then touch
The system uses two senses in sequence. First NVIDIA Robotics' FoundationPose estimates the busbar's exact position and orientation in 3D, so the robot knows where to grab the part and where it is going. That handles the approach.
Then torque feedback takes over. As the part goes in, the robot tracks contact forces in real time, so it can feel when something is binding and ease off before it jams or damages the piece. That is the whole trick. Cameras are good at getting a gripper roughly to the right place and bad at the final fraction of a millimeter, where the information you need is pressure, not pixels.
Shown versus claimed
This is one person's post about a trade show booth, not a published result. Ziegler says it is a small, specific application on a real production line, which is why it caught his eye, but there is no cycle time, no success rate, and no detail on how often an insertion fails or has to retry. Nobody has posted those numbers yet.
What makes it worth a look anyway is how unglamorous it is. Most of the attention in robot manipulation right now goes to general purpose models grabbing unfamiliar objects. Contact rich insertion on a known part, repeated all day at the tolerances electronics assembly actually runs at, is a different problem, and the fix here is not a bigger model. It is a pose estimator plus a force loop.
Ziegler ends the post with a teaser, asking readers to guess which NVIDIA platform the robot is assembling.
Sound like you were at the demo
One email a day. Under three minutes. The demos, the deployments and the part the press release left out.