NVIDIA Robots Automate GB300 Tester Tray Assembly
NVIDIA’s Seattle Robotics Lab, in collaboration with its Isaac engineering team and contract manufacturer Foxconn, has successfully demonstrated robotic automation for assembling GB300 superchip tester trays, marking a significant step toward intelligent manufacturing for AI infrastructure. The project addresses two highly dexterous tasks long considered resistant to full automation: heavy busbar fastening and precise multi-connector insertion. These operations demand exceptional adaptability due to manufacturing variances, flexible cable dynamics, and stringent industrial requirements for a 99.5 percent success rate and cycle times under twice that of human workers. Confronting Moravec’s paradox, the lab abandoned a rigid reliance on end-to-end learning in favor of a modular, hybrid architecture that applies classical engineering where effective and deploys machine learning only when necessary. For busbar assembly, a traditional perception-planning-control pipeline utilizing NVIDIA FoundationPose and high-performance impedance controllers achieved over 95 percent success. The system divided operations across multiple robot arms, optimizing workflow without compromising safety or part integrity. Connector insertion proved more complex due to tight clearances and deformable cables. Initial attempts using generalist vision models and large-scale imitation learning failed to meet precision thresholds or data-scale requirements. Instead, the team developed DOPER, a specialized pose-estimation framework trained on synthetic CAD data and fine-tuned with real-world neural reconstructions. This enabled reliable cable handling and connector alignment. To bridge the simulation-to-reality gap, researchers integrated model-free reinforcement learning within NVIDIA Isaac Lab, later augmenting policies with real-world force-torque feedback to correct minor insertion errors. Physical design also played a critical role; custom 3D-printed gripper fingers were engineered to constrain part motion during contact, effectively reducing uncertainty without adding tactile sensors or complex manipulation routines. Underpinning these efforts is NVIDIA’s Task and Agent Lifecycle Orchestration System and a suite of containerized robotics services, which streamline development, ensure sub-millisecond control loops, and enable seamless integration of perception, planning, and execution modules. The initiative yielded several strategic insights for the robotics community. The team identified an inverse bitter lesson, noting that in manufacturing environments with scarce interaction data, expensive replacement components, and rapidly degrading parts, scalable learning paradigms are often impractical. Traditional methods and specialized AI models consistently outperformed broad foundation approaches. Additionally, the project underscored that cycle time must be prioritized alongside success rate, and that real-world deployment demands rigorous statistical validation far beyond typical lab benchmarks. NVIDIA now plans to transition these systems from laboratory trials to Foxconn production lines. The next phase involves resolving domain mismatch, accommodating live factory variables, and establishing continuous learning loops where human interventions during initial deployment generate the real-world data required for robust model refinement. By merging mechanical intelligence, structured AI pipelines, and scalable robotics infrastructure, the project lays groundwork for a self-reinforcing cycle of automated hardware production. As skilled manufacturing labor faces projected shortages, intelligent assembly systems like this could fundamentally reshape how next-generation AI compute infrastructure is built, tested, and scaled globally.
