NVIDIA is using industrial robots to automate the assembly of its latest AI hardware, but research into the process has found that conventional robotics can outperform more sophisticated AI approaches when tackling some complex manufacturing tasks.
Researchers at NVIDIA’s Seattle Robotics Lab have been working with the company’s operations team and contract manufacturer Foxconn to automate the assembly of tester trays for its Grace Blackwell GB300 systems.
The project focused on two tasks that are relatively straightforward for skilled factory workers but considerably harder for robots: fitting a high-current busbar and inserting four cable-mounted electrical connectors.
The researchers were investigating whether AI-based robotics could handle manufacturing tasks in which component position, geometry and mechanical behaviour vary. Instead, the work highlighted the continuing importance of conventional perception, motion planning, mechanical design and control.
The research team said the project had “challenged assumptions about where robot learning is most effective”, particularly in manufacturing environments where components are scarce, expensive or subject to wear.
NVIDIA and Foxconn set demanding production targets. Both tasks had to achieve a 99.5% success rate, while the busbar operation had to be completed within 124 seconds and the multi-connector operation within 72 seconds. Neither task could involve unintended collisions with the tray.
Those requirements exposed a gap between robotics research and factory automation. According to the research team, research systems can sometimes be considered successful at 80-90% reliability, whereas production equipment needs to operate at a substantially higher level.
For the busbar operation, NVIDIA initially built a modular robotic system combining perception, planning and control. Multiple cameras were used for object pose estimation, a waypoint planner generated robot movements, manipulation routines corrected misalignment and an impedance controller managed contact forces.
Three robot arms shared the task. Two Flexiv Rizon 4S arms handled positioning and manipulation of the fixture and busbar, while a Universal Robots UR10e equipped with an OnRobot screwdriver carried out the 16 screwdriving operations.
The system achieved a success rate of more than 95%, although that remained below the 99.5% target. Most failures involved grasping and slippage.
Cycle time was also above the required level. The complete operation took around 160 seconds, compared with the 124-second target, with the sequential screwdriver operations identified as the main bottleneck.
The result was notable because NVIDIA had originally intended to replace the conventional approach with an end-to-end learning system. Instead, the researchers concluded that the existing architecture was effective enough to retain.
The second task, inserting four electrical connectors into tight-clearance sockets, proved more difficult.
The connectors were attached to cables that could deform and change shape, while the connectors themselves were small and relatively textureless. Conventional computer vision was effective at identifying the cables and determining where the robot should grasp them, but general-purpose pose estimation struggled with the connectors.
NVIDIA therefore tested learning-based techniques including behaviour cloning and diffusion policies.
The approaches ran into a practical limitation: there was not enough high-quality training data. The specialised cables were available only in small quantities, and repeated manipulation could permanently deform them.
The research team described this as an “inverse bitter lesson”. While robotics has increasingly followed AI towards larger models and more data, physical manufacturing imposes constraints that do not apply to digital systems.
NVIDIA says its next step is to deploy the technology in the factories of its contract manufacturers and work towards the performance levels required for production. The company is also preparing a whitepaper and plans to release DOPER and its Task and Agent Lifecycle Orchestration System, TALOS, more widely.
