The Machines that Make the Machines | NVIDIA Technical Blog

This post details the application of robotics and AI to assemble GB300 tester trays, focusing on busbar assembly and multi-connector insertion. For busbar assembly, it describes a classical modular pipeline using FoundationPose for perception, Lissajous-curve manipulation primitives for planning, and a high-performance impedance controller, achieving over 95% success. For multi-connector insertion, it introduces Deep Object Pose Estimation Revisited (DOPER) for specialized pose estimation and the use of custom 3D-printed gripper fingers for mechanical constraint. Connector insertion policies were pretrained using sim-to-real reinforcement learning in NVIDIA Isaac Lab and refined with real-world residual policy training via the SPARR method. The post also highlights the use of robotics services and the Task and Agent Lifecycle Orchestration System (TALOS) for reusable components.