According to Reuters, the Chinese startup Spirit AI has supplied dozens of Moz1 robots which work at the CATL battery manufacturer's mill and at JD.com locations. These machines belong to the first wave of humanoids, which are entrusted with real tasks previously performed by humans.
Moz1 is simply a humanoid robot that moves on wheels but has torsos and 2 arms designed to manipulate objects. The company states that the structure has 26 degrees of freedom, not counting the hand, and the full body force control system. This plan is designed to let work in places where the robot must scope the station, admit the object and make precise movements of the arms.
Co-founder and chief starter scientist, Gao Yang, predicts that by mid-2027 AI models operating robots can scope the phase he compared to the "GPT-3 moment". At this phase of technological development, the robot should accept the command in natural language, arrange a reasonable series of physical activities and effort to execute the task.
Today the body precedes the soul. Robots already have very efficient mechanisms with different freedom of movement, but their limitations are software that inactive makes these machines implemented for circumstantial work (in a factory, shop, cafe) "mentally" inactive at the level of industrial robots.
CATL: Robot at the end of battery production line
The most detailed implementation of the Moz1 robot takes place at the CATL Zhongzhou production base. Robots work there on final battery kits tests, referred to as EOL and DCR processes. Test connectors are inserted into the appropriate battery pack points at this stage. due to the fact that the strategy works under a voltage of hundreds of volts, there is simply a hazard of an electrical arc dangerous to human wellness and life.
Therefore, as shortly as possible, a robot was introduced to this place, replacing a man. The usage of an average industrial robot prevented the nature of the work carried out in that position. Different models of sets are mounted on 1 production line in short series, and the position of the elements in the sets may change.
Unlike an industrial robot, Moz1 recognizes the position of an element, can adjust its own movements (sequences of movement) and choice the force needed in the operation of an elastic wiring beam.
Spirit AI states that the efficiency of inserting the connector exceeds 99 percent, and the rate of work presently corresponds to the level of experienced employee. Thanks to the introduction of the humanoid, regular performance with continuous production of many variants of batteries increased 3 times.
Moz1 is not a universal robot. It carries out 1 repetitive task in an environment that is variable but has clearly defined objects, tools and safety rules. From specified industrial implementation to the robot of general purpose, the road is inactive far away.
JDcom: shy attempts to work in JD MALL stores
JD.com is not only the place of implementation, but besides the investor Spirit AI. In 2025, JD led a Pre-A+ startup circular worth nearly 600 million RMB. In March 2026 the companies concluded a contract for the years 2026–2029. This includes the improvement of consumer products, commercial applications, data collection and distant control of robots.
The first described application in this cooperation was Moz in the stationary JD MALL store. The robot carried out tasks (show character) of preparing coffee: collecting a cup, grinding grains and burning.
In this scenario, the operation of the device is supported by distant operator control. During this time, during robot work, data from cameras, sensors, joints and force control systems are collected.
Data from specified operations is utilized for further software training. In the case of conventional automation, any workstation of the machine, robot, is usually designed for a well described activity. In the case of the embodiment model AI, or software for a humanoid robot, the strategy is to learn the connection between what the robot sees, the command it receives, and the movement needed to accomplish the goal. As a result, a device supported by the embodiment model AI can execute customized activities resulting from a peculiar situation and the request to accomplish the stated objective. That's far beyond a typical industrial robot.
VLA model combines image, command and movement
Moz1 uses a VLA kind model, or vision-language-action. The first associate means an image from cameras and another data about the environment, the second is simply a command written or spoken by man, and the 3rd is simply a physical action performed by a robot. The strategy transforms the situation information into a series of chassis, arm and gripping movements.
The company develops its own Spirit models. In materials related to the Spirit v1.6 version, it declares a combination of the planet model with end-to-end control, a strategy learning straight the relation between input data and action. In practice, the robot inactive needs training on tasks and data obtained under conditions akin to those in which it will carry out its tasks. Each fresh item, each fresh command will require the trainer/operator to prepare the device properly.
Gao Yang indicates that access to data from actual interactions is presently a regulation for specified models. Reuters reported that about a 1000 Spirit AI associates throughout China usage traffic recorders in homes and factories. The company uses these recordings to teach robots. According to Gao, simulations cope well with rigid lumps, but they little pattern deformable objects, specified as electrical wires.
This problem concerns the full industry. The Chinese Data Administration has started work on common requirements for data sets utilized in the embodiment AI, as described in the text ‘China announces standards for the embodiment AI“. The data must combine image, motion, sensor readings and task consequence in a way that allows them to be utilized for training and model comparison.
A akin function is performed by centers collecting data from physical robot exercises. An example is described by us Beijing Humanoid Robot Innovation Center, where motion trajectories, force sensors signals and the consequence of the task are recorded in simulated spaces of home, store or factory.
Forecast for 2027, limitations today
Gao Yang called the robot brain the weakest part of the current robotic technology set. In an interview with Reuters, he predicted that by mid-2027 models could scope a level allowing the robot to accept the command in a average language and carry out a string of activities needed to implement it.
At the same time, the company indicates the current limits of its systems. Spirit AI reports about 90% of the success of simple tasks in an orderly surviving room-like environment. The actions requiring very precise handling, specified as unscrewing the bottle cap, and tasks that the model has not done before, stay difficult. According to Gao, industrial applications are a short term, in fact useful home robots will appear on the marketplace no sooner than in about 8 years.
However, it is worth remembering that Gao Yang evaluates future advancement of humanoid robots from the position of his own company. Meanwhile, the number of companies active in humanoid robotics is increasing very rapidly in China. In this number there are most likely already all the largest Chinese technological corporations specified as Alibaba, Tencent, or Xiaomi. And that means that mature technological home robots can be sold much earlier than in 8 years.
Source:
- Spirit AI, "The robot of embodiment intelligence enters the production of CATL battery kits": https://www.spirit-ai.com/news/22
- Spirit AI, description of the Moz robot: https://www.spirit-ai.com/product
- Tsinghua University, "Building a Universal Robot Brain" — Gao Yang profile, September 11, 2026: https://www.tsinghua.edu.cn/info/1182/128169.htm
- GeekPark, "Spirit AI and JD conclude trade implementation agreement", 18 March 2026: https://www.geekpark.net/news/361345
- CATL, ‘CATL and JD Group conclude comprehensive cooperation agreement’, 15 October 2025: https://www.catl.com/news/8905.html
- Pandaily, "Spirit AI Deploys Moz1 on CATL and JD Lines; Founder Eyes Mid-2027 Robot-Brain Breakthrough", 20 September 2026: https://pandaily.com/spirit-ai-moz1-catl-jd-mid-2027-robot-brain
- Reuters, ‘Chinese ‘robot brain’ startup sees ChatGPT-style breakthrough as shortly as next year’, 18 September 2026: https://www.reuters.com/world/asia-pacific/founder-chinese-startup-spirit-ai-says-robot-brains-set-2027-breakthrough-2026-09-18/
Email: [email protected]
![Ostrołęka ma dwóch nowych prokuratorów. Nominacje odebrali z rąk Żurka i Korneluka [ZDJĘCIA]](https://www.eostroleka.pl/luba/dane/pliki/zdjecia/2026/fad63e92-3c7d-48a8-8cc8-8b4259c570e4.jpeg)