NVIDIA and Foxconn Use Robots to Assemble GB300, Target Success Rate of 99.5%

On October 4, 2026, the technology news outlet Tech New Report reported that NVIDIA and Foxconn are testing a robotic assembly process for the Grace Blackwell GB300 test tray, with a joint objective of raising the success rate of critical assembly tasks to above 99.5%. The report highlighted that the two technology leaders are combining their expertise to push the limits of automated manufacturing for high‑performance server components.

Development Background and Objectives

NVIDIA's robotics laboratory located in Seattle, together with its Isaac engineering team, has recently entered into a partnership with Foxconn, the contract manufacturer of the GB300, focusing on high‑precision, long‑duration automated assembly technologies. The two parties have jointly defined a production‑line threshold that requires a minimum success rate of 99.5%, zero collisions with the tray, a bus‑assembly time not exceeding 124 seconds, and a connector‑insertion time not exceeding 72 seconds. These quantitative targets are intended to ensure that the automated line can meet the demanding reliability and speed specifications required for next‑generation data‑center hardware.

The core of this collaboration is the integration of NVIDIA's Isaac Lab reinforcement‑learning platform with Foxconn's extensive manufacturing resources, aiming to create a robotic system that can maintain high stability even when operating in a changing environment. By leveraging simulated learning and real‑world feedback, the partnership seeks to produce robots that adapt quickly to variations in parts, lighting, and mechanical wear.

Robotic Assembly Process and Current Status

The assembly task for the GB300 test tray is divided into two major focus areas. The first area involves placing a long and heavy bus and then using a robotic arm to fasten sixteen screws with precise torque. The second area requires the insertion of four sets of high‑tolerance connectors, consisting of two large and two small connectors that must be seated accurately to meet electrical performance standards. These two tasks impose strict requirements on the robot's positioning accuracy, force control, and sensor feedback, because any misalignment could compromise the functionality of the final product.

To date, the success rate for bus placement and screw fastening has already surpassed 95%, while the complete end‑to‑end process—including both the bus and the connectors—averages roughly 160 seconds. This average remains above the target times of 124 seconds for the bus operation and 72 seconds for the connector insertion, indicating that further optimization is needed to close the gap between current performance and the defined production thresholds.

The connector insertion operation combines Isaac Lab's reinforcement‑learning algorithms, real‑world manufacturing data, and custom‑designed grippers. In a simulated environment, the robot learns multiple insertion strategies by exploring variations in angle, speed, and force. During physical testing, the system continuously refines its approach based on sensor readings, thereby improving its tolerance to the tight tolerances of the high‑precision connectors and reducing the likelihood of mis‑feeds or damage.

Future Production Path and Challenges

Both parties explicitly state that the current testing phase represents only a development milestone and does not yet demonstrate readiness for full‑scale mass production. The key challenges ahead include raising the bus‑assembly success rate above the 99.5% threshold, shortening the overall process to meet the 124‑second and 72‑second targets, and ensuring that the robot does not cause any tray collisions during prolonged operation. Each of these challenges requires advances in hardware reliability, software robustness, and real‑time monitoring.

To overcome these challenges, NVIDIA plans to continuously expand the training data set for Isaac Lab, incorporating additional failure cases and edge‑scenario simulations. Simultaneously, NVIDIA and Foxconn will work jointly to optimize the mechanical arm's hardware architecture and sensor layout, improving rigidity, reducing latency, and enhancing feedback resolution. Foxconn, in turn, will supply additional real‑world production‑line test samples to verify the robot's consistent performance across different product batches and to validate the scalability of the solution.

Current test results indicate that the robot already exhibits considerable stability under high‑load and high‑tolerance conditions, but further breakthroughs are required in precision calibration and dynamic obstacle avoidance before the 99.5% success‑rate benchmark can be reliably achieved. In particular, fine‑tuning of the vision system and the implementation of predictive collision‑avoidance algorithms are seen as critical steps toward meeting the stringent reliability goals.

In summary, the collaboration between NVIDIA and Foxconn has already yielded significant progress in robotic assembly of the GB300 test tray, particularly in the automation of bus handling and connector insertion, demonstrating promising potential for future scaling. If the remaining timing and success‑rate hurdles can be cleared, the joint solution could provide a new level of automation for the production of high‑end servers and AI accelerator cards.

The present shift is that both parties are accelerating the verification and optimization stage, with an expectation to complete validation of all defined thresholds within the next several months; at that point, a comprehensive assessment of full‑scale production feasibility and scheduling will be undertaken.