NVIDIA Trains Robots to Assemble GB300 Tester Trays

NVIDIA's Seattle Robotics Lab and its Isaac engineering team have designed a multi-robot system to automate the assembly of GB300 tester trays, which verify Grace Blackwell superchips before shipping. To tackle this highly precise task, the team deployed three robotic arms: two Flexiv Rizon 4S arms and one Universal Robots UR10e arm equipped with an OnRobot screwdriver. They focused on two primary challenges: busbar assembly and multi-connector insertion.

For the busbar assembly, which requires placing a heavy metallic bar and driving 16 screws, the team used a classical modular pipeline. This setup utilized FoundationPose and later a specialist model called Deep Object Pose Estimation Revisited (DOPER) for perception, cuRobo for motion planning, and FoundationStereo for depth estimation. This approach achieved a success rate of over 95% with a 160-second cycle time, approaching the manufacturer's 124-second target.

The second task, inserting two large DC-SCI and two small MCIO electrical connectors, proved more difficult due to flexible, deforming cables. The system used Segment Anything Model 3 (SAM3) to segment and grasp the cables. To grasp the connectors, the researchers trained DOPER on synthetic and real-world data. They also designed custom 3D-printed gripper fingers to mechanically guide the parts into repeatable poses without needing tactile sensors.

To insert the connectors into tight sockets, the team pretrained policies in NVIDIA Isaac Lab using sim-to-real reinforcement learning. They then refined these policies in the real world using the SPARR method, which applies force-torque inputs to train a residual policy. The entire system was managed by the Task and Agent Lifecycle Orchestration System (TALOS) and containerized robotics services, which coordinated the real-time control loops at speeds exceeding 500 Hz.