RESEARCH · RESEARCH · #1694
NVIDIA SRL and Isaac build robotic system to assemble GB300 tester trays
The NVIDIA Seattle Robotics Lab and NVIDIA Isaac engineering team developed robotic systems to assemble GB300 tester trays, focusing on busbar assembly and multi-connector insertion. A modular stack (FoundationPose perception, Lissajous-curve waypoint primitives, and an impedance controller) achieved over 95% success on busbar assembly with a 160s cycle time; multi-connector insertion used SAM3 segmentation for cables and a new pose estimator DOPER for connectors, with sim-to-real RL pretraining in Isaac Lab and SPARR residual policy refinement, plus custom 3D-printed gripper fingers and reusable TALOS/containerized components.
KEY POINTS
- The NVIDIA Seattle Robotics Lab and NVIDIA Isaac engineering team developed robotic systems to assemble GB300 tester trays, focusing on busbar assembly and multi-connector insertion.
- A modular stack (FoundationPose perception, Lissajous-curve waypoint primitives, and an impedance controller) achieved over 95% success on busbar assembly with a 160s cycle time; multi-connector insertion used SAM3 segmentation for cables and a new pose estimator DOPER for connectors, with sim-to-real RL pretraining in Isaac Lab and SPARR residual policy refinement, plus custom 3D-printed gripper fingers and reusable TALOS/containerized components.
- This demonstrates production-relevant robotic assembly of complex AI hardware using a hybrid modular and learning-based stack, advancing sim-to-real training, pose estimation, and reusable orchestration for manufacturing.
WHY IT MATTERS
This demonstrates production-relevant robotic assembly of complex AI hardware using a hybrid modular and learning-based stack, advancing sim-to-real training, pose estimation, and reusable orchestration for manufacturing.