The Machines that Make the Machines

The NVIDIA Seattle Robotics Lab, working with the NVIDIA Isaac engineering team, developed robotic systems to assemble GB300 tester trays, focusing on two tasks recommended by the NVIDIA Operations Team and Foxconn, a GB300 contract manufacturer: busbar assembly and multi-connector insertion.

For busbar assembly, a classical modular pipeline using FoundationPose for perception, waypoint planning with Lissajous-curve manipulation primitives, and a high-performance impedance controller exceeded 95% success, with remaining failures mainly from grasping errors and in-hand slippage during insertion. The full cycle took 160 seconds against a 124-second target, with sequential, device-specific screwdriver operations as the bottleneck. The setup used two Flexiv Rizon 4S arms and a Universal Robots UR10e with an OnRobot screwdriver, driving 16 screws.

Multi-connector insertion proved harder. Cable grasping succeeded with SAM3 segmentation, but connector grasping failed with a generalist pose model and end-to-end learning efforts. NVIDIA Research developed Deep Object Pose Estimation Revisited (DOPER), trained on synthetic data then fine-tuned with real-world pseudolabels on a 3D neural reconstruction. Custom 3D-printed multi-purpose gripper fingers mechanically constrained part motion during contact. Connector insertion policies were pretrained in NVIDIA Isaac Lab using sim-to-real reinforcement learning, then refined with real-world residual policy training via the SPARR method using force-torque inputs.

Robotics services and the Task and Agent Lifecycle Orchestration System (TALOS) supplied reusable, containerized perception, planning, and control components, accelerating development and deployment across robot platforms. TALOS executes controllers in real-time loops faster than 500 Hz and supports the Franka FR3, Flexiv Rizon 4S, Universal Robots UR10e, and SO-101.

Why it matters: Manufacturing partners requested 99.5% success rates and cycle times no more than twice those of skilled workers; the lab reports it is rapidly approaching those thresholds as of the article's publication. The work indicates where classical modular pipelines prove effective and where specialist models like DOPER are needed for real-world manipulation.