Hon Hai joins hands with Huida to promote AI server manufacturing automation. The two parties have locked in the two key tasks of bus assembly and multi-connector plugging of the GB300 test tray. They have introduced robots to challenge complex assembly operations and aim at a success rate of at least 99.5%. At the same time, the operation time is required to be no more than twice that of skilled workers, bringing actual AI server manufacturing needs into robot automation development.
Seattle Robotics Lab (SRL) recently revealed the research results of the GB300 test pallet robot assembly. Huida pointed out that the assembly of GB300 pallets still requires skilled manpower, so the team began to study whether smart robots can complete the assembly of such complex systems. Based on the recommendations of Huida's operations team and Hon Hai, which is the GB300 contract manufacturer, busbar assembly and multi-connector plugging were listed as two key tasks.
Among them, busbar assembly requires robots to grab and carry long and heavy busbars, place them on the pallet and lock screws at 16 locations. The process also involves the installation and removal of jigs and fixtures; multi-connector plugging requires picking up two large and two small electrical connectors from the pallet and inserting them into slots with narrow gaps.
In order to make the research results meet actual manufacturing needs, Huida said that the relevant performance requirements are formulated in direct cooperation with the operation team and Hon Hai. The busbar assembly success rate target is 99.5%, and the completion time must not exceed twice that of skilled workers, that is, 124 seconds. There must not be any unexpected collision with the pallet during the process. Even slight damage may cause the entire system to be scrapped.
Multi-connector plugging also requires a success rate of at least 99.5%. The four connectors must complete the operation within 72 seconds, and no unexpected collisions shall occur. However, while skilled workers can complete the relevant work smoothly, robots face challenges such as part position changes, wire deformation, manufacturing differences, and narrow insertion gaps, making seemingly simple assembly actions a difficult automation problem.
According to Huida's statistics, the success rate of busbar assembly solutions has exceeded 95%. Huida continues to improve the success rate. Among them, the limit fixture and busbar sub-tasks have met the cycle time requirements, but the speed of the locking screws has not yet reached the standard, making the complete operation time about 160 seconds, higher than the target of 124 seconds. The main bottleneck comes from the 16 screw locking operations performed in sequence.
Huida pointed out that general manufacturing can reduce uncertainty through fixed automation, using fixtures and transfer devices to control the position of parts with extremely high precision, and then robots repeatedly perform fixed actions; however, the GB300 test pallet faces rapid design cycles and low-volume production conditions, making fixed automation unfeasible. Robots must have higher adaptability to cope with the uncertainty of the actual manufacturing environment.
The technology development process also showed that different processes require different methods. The bus assembly finally adopted a modular architecture such as sensing, planning and control, with three robot arms working separately. For multi-connector plugging, due to easy deformation of wires, small connector sizes and narrow slot gaps, the team once tried end-to-end learning, but was limited by the scale of physical data acquisition. Finally, it returned to a dedicated posture estimation method to continue to improve the robot's ability to identify and operate actual parts.
