October 11, 2026
NVIDIA robots assemble GB300 test trays: over 95% success rate for bus bar assembly
NVIDIA robots insert the conductive busbar into the GB300 test tray with a success rate above 95%, and connectors are attached in 90–95% of cases (NVIDIA blog, Oct 7, 2026). The trays test compute modules before they ship to customers; previously this assembly was done by hand.
Manual assembly pressed the busbar down with a large protective bracket. Now a single screwdriver tightens it without a bracket: 16 screws in one process step.
Designs change fast and batches are small, so NVIDIA chose flexible adaptive automation over rigid systems with tight tolerances.
What's open today. GitHub hosts FoundationPose (object pose estimation), cuRobo (motion planning), Isaac Lab (simulation and reinforcement learning) and DROID+. The stack works with Franka FR3, Flexiv Rizon 4S, Universal Robots UR10e and SO-101 robots, and the services ship in Docker.
DOPER, the TALOS library and reference connector-insertion workflows are promised but not yet released. The cycle is still longer than the target: the busbar takes about 160 seconds against a target of 124.
The DOPER and TALOS release is promised, but the blog gives no date.