The Machines that Make the Machines

NVIDIA's Seattle Robotics Lab, working with the Isaac engineering team, describes trying to automate two hand tasks in assembling GB300 tester trays, which are used to check GB300 compute modules before they ship. The goals were set with NVIDIA's operations team and contract manufacturer Foxconn: 99.5% success, no unintended collisions with the tray, and no more than twice the time of a skilled worker.

The robots are not there yet. Busbar assembly, which includes driving 16 screws, succeeds above 95% of the time but takes 160 seconds against a 124-second target. Inserting four cable-mounted connectors succeeds 90 to 95% of the time, at about 40 seconds per cable against a 72-second target for all four.

We like this post because it reports the misses. Plenty of robotics demos stop at the clip that worked. This team lays out the target, the gap and why. It also makes a point the field needs to hear: in research, 80 to 90% success can be called solved, while a factory wants 99.5%, and proving that with confidence takes, by their math, at least 598 trials in a row without a failure.

The technical story is a useful corrective to the idea that bigger end-to-end models fix everything. A classical pipeline of perception, planning and a carefully tuned controller handled busbar assembly with no pivot to learning needed. For the connectors, imitation learning stalled because there were only a few of these specialized cables and each one deformed after limited handling. The team calls this an 'inverse bitter lesson': sometimes the data simply cannot be gathered at scale. What worked was a specialist pose estimator, 3D-printed gripper fingers that physically guide parts into place, and insertion policies trained in simulation and then refined on the real robot.

The bigger frame is NVIDIA's stated vision of robots, under human supervision, assembling the hardware that trains the next generation of robots. It cites 1.9 million manufacturing jobs that may go unfilled by 2033. We take the labor-shortage argument seriously; the tasks shown are repetitive and precise. But the same capability that fills open jobs can also replace the skilled workers shown in the post's own videos, and that trade-off deserves more than a footnote as these systems move into contract factories.

Our read: this is solid, grounded work that is candid about the distance between a lab result and a production line. I'm watching whether it reaches Foxconn's floor at the required reliability, and whether the promised tools actually ship.

The setup used two Flexiv Rizon 4S arms and a Universal Robots UR10e with a screwdriver. The team says it is preparing a whitepaper and working toward releasing its DOPER pose-estimation framework and TALOS orchestration library for community use, along with reference workflows for connector insertion.