NVIDIA's Seattle lab tested robots on GB300 tester trays, the fixtures used to verify compute modules before shipment. In a report posted Oct. 7, the company said its arms fell short of targets on both assembly tasks: fastening a busbar and inserting four cable connectors. These were lab trials, not factory deployment.
Neither task met thresholds set with NVIDIA Operations and contract manufacturer Foxconn. NVIDIA reported a busbar success rate above 95%, below its 99.5% target, and a 160-second cycle against a 124-second limit. For the full connector sequence covering all four parts, it reported 90% to 95% success. NVIDIA reported an average of 40 seconds per cable, not an overall four-cable cycle time; multiplying that average by four gives a rough 160 seconds, compared with the 72-second limit for all four. The post does not say whether either success rate also requires meeting its time limit or disclose the trial counts behind the rates.
For the busbar, the team used separate steps to find parts, plan arm movements and control contact. NVIDIA says the main remaining failures came from poor grasps or slippage during insertion. For connectors, cameras track each cable and estimate where its connector is and which way it faces, so one arm can position it for a second arm to grip. Custom-shaped fingers help hold parts consistently. NVIDIA trained insertion motions in simulation, then refined them with real-world data and force-and-torque readings because the connector is hidden during insertion. It reports coordination between the arms as the main remaining failure source for connectors.
Both tasks prohibited unintended contact with the tray because even minor damage could scrap a system. NVIDIA says at least 598 trials with zero failures would be needed to support a 99.5% success probability with a one-sided 95% confidence bound. Its post does not disclose how many trials produced the two reported rates.
