Robots take on AI server test trays: busbar assembly tops 95% success, four-cable insertion reaches 90–95%

Robots can now perform two GB300 tester-tray assembly tasks: inserting a long, heavy busbar and driving 16 screws, and inserting two large and two small cable-mounted connectors into tight-clearance sockets. These operations previously relied on skilled factory workers, who perform them reliably and fluidly with little mental load, while the same dexterity is exceptionally hard to automate. Busbar assembly exceeds 95% success, with remaining failures mainly from grasping errors and in-hand slippage during insertion; four-connector insertion reaches 90–95%, with remaining failures mainly from multi-arm coordination during connector grasping.

The approach tried classical baselines first for each subproblem and introduced learning only where those fell short, rather than training end-to-end from the start. Robotics research tends to prize technical novelty and offers little incentive to optimize baselines that may perform better, so it is easy to finish a paper without knowing why prior approaches failed. The busbar pipeline of separate perception, planning and control modules worked well enough that no pivot to end-to-end learning was needed, and its modular structure made debugging, tuning and distribution across three robot arms straightforward.

Connector grasping moved to a specialist pose estimator, DOPER, and insertion combined simulation-pretrained reinforcement learning with real-world residual learning. Generalist pose estimation was inaccurate on the connectors, end-to-end ACT or diffusion policies had low success, and the specialized cables were scarce and deformed permanently after limited use, making large-scale data hard to acquire. DOPER is first trained in simulation on rendered CAD views with automatically generated keypoint labels, then fine-tuned on pseudolabels from a 3D neural reconstruction, running at 30 Hz or faster; insertion policies trained in Isaac Lab transferred well for the large connector but were less effective for the small one because of pose-estimation error.

The demonstrated scope is laboratory performance on two GB300 tester-tray tasks: busbar assembly above 95% success at 160 seconds per cycle, and four-connector insertion at 90–95% success with about 40 seconds per cable, both short of the manufacturer's 99.5% success requirement and the 124-second and 72-second cycle-time limits. The stated next step is deployment into factories where NVIDIA contract manufacturers produce AI hardware, alongside a planned whitepaper and releases of DOPER and TALOS.

The authors note that good laboratory trials do not guarantee post-deployment performance: factory ambient and workcell conditions differ, domain mismatch is a near-universal weakness of learning-based robots, and production lines change with product design. Supporting a 99.5% success probability with a one-sided 95% confidence bound requires zero failures over at least 598 trials, and cable wear means such evaluations may need hundreds of duplicate GB300 components; accurately simulating geometrically and mechanically irregular parts remains an open challenge.