Humanoid Robots Not Yet Mature, Physical AI Turns to Robotic Arms

September saw two converging moves in robotics. On September 10, U.S. robot model company Skild revealed it is collaborating with Nvidia and Foxconn to deploy models on dual-arm robots for precision assembly of Blackwell systems.

Four days later, Danish collaborative robot maker Universal Robots launched a new platform, redesigning the control box and arm-end connections to make it easier to integrate cameras, sensors, and AI applications.

A model company is borrowing mature robotic arms; a robotic arm company is modifying its products for new intelligence. Making robots smarter doesn't necessarily require waiting for a completely new body.

Canadian company Sanctuary AI, developer of the Phoenix humanoid robot, announced a route adjustment in June. It will deploy Physical AI on existing and next-generation industrial robot platforms, no longer waiting for humanoid hardware to reach large-scale commercialization.

The company disclosed a proof of concept with an automotive Tier 1 supplier: an industrial arm manipulating soft wire harnesses to insert them into moving targets. According to the company, the task success rate exceeds 99.5%, with a cycle time of 2.54 seconds, meeting the customer's existing production line requirements. This work does not require solving how a robot walks or maintains full-body balance. Sanctuary can put its AI on mature arms to first validate whether it can complete a process. Emphasizing both success rate and the 2.54-second cycle time shows the requirement to be both accurate and fast enough for production.

Factories want to hand over to new models the tasks that still rely on human judgment. Some processes already have sufficient arm strength and precision, but the objects vary each time. Japanese automotive wire harness supplier Sumitomo Wiring Systems is validating this possibility. On September 17, it announced joint development of a Physical AI robot with Skild. The collaboration announcement cited specific reasons: wire harness manufacturing involves many complex processes that traditional robotics has long struggled to replace, thus still relying on manual labor.

The trouble with wire harnesses is that they rarely maintain the same state. Rigid parts can be fixed with fixtures, but soft harnesses bend and tangle. Grasping one part changes the others. Even if a robotic arm accurately returns to the same coordinate, it doesn't mean a plug is there to grasp. For such work, the ability to accurately repeat motions is not enough. The robot must first assess the current state of the harness, then decide where to grasp and how to move.

To enable new models to truly participate in operations, industrial robot manufacturers are also modifying control systems and interfaces. Japanese industrial robot maker Yaskawa started with the division of labor in the control system. In July, Yaskawa and Japanese telecom company SoftBank announced a wire harness bin packing validation. The system judges the harness state based on camera images and task instructions, then selects grasping and manipulation actions. Yaskawa retained parts that traditional control can reliably handle, assigning functions like object state recognition and grasp point selection to AI, integrated as a module into the existing system.

Universal Robots' Gen 7 platform, released in September, further extends these modifications to hardware connections. The arm end can directly provide data and power for cameras and high-bandwidth sensors, reducing complex wiring.

Chinese industrial robot maker Estun is opening its control system and development interfaces. The company disclosed building the RoboBase industrial embodied intelligence open platform based on its self-developed control system, opening driver layer and application layer interfaces.

Model companies enter the market by leveraging mature equipment, while industrial robot manufacturers hope to expand equipment uses with new capabilities. More general bodies can continue R&D. For now, this brain first tries to plug in the wire.