Spirit AI’s Moz1 at CATL: A Battery-Line Deployment, Not a Home-Robot Breakthrough

Spirit AI’s Moz1 wheeled humanoids are being put to work on defined tasks, rather than presented as dependable general-purpose helpers. The clearest account of their work is at CATL’s battery PACK line; much less is established about what they do at JD.com. 643

What Moz1 does at CATL—and what is known about JD.com

Spirit AI says tens of Moz1 robots have been deployed at CATL and JD.com. At CATL’s Zhongzhou battery PACK line, their reported work includes inserting and removing test connectors while adjusting motion and force to handle flexible wiring. The available reporting does not establish the robots’ precise duties at JD.com. 43

Spirit AI reports more than 99% success for connector insertion at CATL, with cycle times comparable to those of skilled workers. That is a company-reported result for a particular operation—not an independently established measure of whole-line reliability or the robot’s ability to perform unfamiliar work. 6

A separate figure needs its own context: Gao has described roughly 90% success on simple tasks in structured living-room environments. It is not the CATL connector-insertion result. 24

How the deployment fits Spirit AI’s training strategy

Spirit AI is developing robot “brains” using data from physical activity. Reports describe about 1,000 contractors collecting real-world training data, while Gao identifies the shortage of suitable data as a bottleneck. Factory deployments show where a robot can be assigned a repeatable task; they do not, on their own, show that its training will transfer reliably to every setting. 13

That distinction matters because fine manipulation and unfamiliar situations remain difficult for the robots. Those limitations, alongside the safety demands of working around people, constrain how quickly a capability demonstrated in a structured environment can be used more broadly. 1

What Gao’s mid-2027 forecast actually predicts

Gao’s proposed “GPT-3.0 milestone” by mid-2027 is a forecast for robot intelligence: a person could give a natural-language instruction and a robot could attempt a sequence of physical actions in response. It is not a claim that Moz1 will be dependable at every task by that date. 343

Gao describes the next one to two years as an initial window for industrial applications, with simpler commercial-service work expected before household use. He puts useful deployment in homes at least eight years away. The CATL result and the 2027 forecast therefore mark different things: reported performance on one factory task today, and an anticipated step toward more flexible behavior later. 143