Meta and Fourier built the humanoid robot prototype Max in three months

Meta is making a foray into the humanoid robotics race, and its first prototype was built by a Chinese company.

According to foreign media reports, Meta has partnered with Chinese robotics company Fourier, completing a humanoid robot prototype named Max in just three months, and has already begun development of a second-generation product. This is also the most concrete hardware progress since Meta established its robotics team.

From AI models to humanoid robots, Zuckerberg is extending Meta's technology footprint into the physical world. Notably, in this collaboration, Meta did not choose to build a hardware system from scratch, but instead leveraged a Chinese company's robotics R&D and manufacturing capabilities to rapidly complete prototype development.

What tasks exactly does this robot undertake? Why did Meta choose Fourier? And why can't American tech giants' robotics plans avoid Chinese manufacturing?

01 A prototype built in three months

According to foreign media reports, Max is about 1.5 meters tall, weighs about 48 kilograms, and is equipped with an Nvidia Jetson Thor chip. The robot has two built-in speakers in its chest, a display on its face capable of simulating human expressions, and five-fingered hands able to pick up small bottles and toys.

Meta's requirements for this robot include not only basic movement capabilities but also human-robot interaction. According to people familiar with the matter, Max needs to be able to recognize voices from 5 meters away, recognize gestures and waving from 3 meters away, maintain eye contact, and express states such as listening, thinking, speaking, and uncertainty through different status signals.

The robot was customized based on Fourier's existing technology. Meta leveraged the hardware and R&D experience of Fourier's previous GR-series robots; the two sides then adjusted the robot's height, body proportions, head structure, display, and cameras, and modified the hands and shoes.

Zuckerberg with the Fourier GR-3 concept robot

Li-Chen Miller, Meta's Vice President of Robotics Product, worked at Fourier's Shanghai factory for nearly two months, advancing development together with the local team. At the end of August, Max was demonstrated live to some employees at an internal Meta robotics summit.

In September, the two sides began development of Max 2. The new version plans to redesign the torso, improve hand manipulation capabilities, add a projector for displaying visual content, and equip additional sensors to help the robot go up and down stairs.

A Meta spokesperson said the collaboration with Fourier is mainly for internal research and experimentation, and the company has no plans to sell these robots.

This does not mean Meta has no larger product ambitions. People familiar with the matter said the company is recruiting supplier development engineers in Shanghai and product design engineers and technical project managers in Shenzhen, with work involving product design, manufacturing, prototype development, and preparation from prototype to mass production.

From on-site R&D staff to recruiting for Chinese supply chain positions, Meta is extending its robotics project from a single prototype toward a more complete R&D and manufacturing system. However, there is currently no evidence that Meta has settled on mass production plans or a product launch timeline.

02 Giving Muse a "body"

Meta's core goal in developing Max is related to its AI agent Muse.

Meta Chief Technology Officer Andrew Bosworth has internally called the humanoid robot Muse's physical carrier. This means Meta hopes AI can not only answer questions through dialogue but also execute tasks in real environments with the help of robots.

In May this year, Meta acquired robotics AI startup Assured Robot Intelligence. The company focuses on developing robot models that help machines understand human behavior and adapt to different environments. The related technology and team have also become part of Meta's robotics R&D.

Judging from the current setup, Meta is filling in the software capabilities needed for robot AI while using Max to test how these capabilities combine with hardware.

For example, if a user asks the robot to tidy a table, the AI must first recognize the objects on the table, determine which items need to be moved, and then plan grasping and placement actions. During execution, the robot also needs to adjust its grip based on the position, weight, and shape of objects. If an item slips or the planned route is blocked, it must reassess its next action.

Meta needs a physical platform like Max precisely to test how models perform in real environments. The perception errors, action failures, and environmental changes robots encounter while executing tasks can all serve as a basis for subsequent model training and hardware improvements.

Meta's robotics business has also undergone management changes. Marc Whitten, who previously served as CEO of General Motors' autonomous driving unit Cruise, left at the end of August, only about 18 months after joining Meta. Meta subsequently appointed Ning Li as interim head of Meta Robotics Studio.

Bosworth took over Whitten's previous responsibilities and continues to be involved in the department's technical and operational work.

For Meta, whether Muse can evolve from a software agent into a system capable of executing real-world tasks depends on whether the model and robot hardware can work together. Max is currently undertaking exactly this R&D task.

03 The US does AI, China does hardware

Meta's choice of Fourier shows the importance of China's robot manufacturing capabilities in the global industrial chain.

Fourier initially focused on exoskeletons and rehabilitation robots, and later entered the humanoid robot field. Its GR series products have gone through multiple generations of iteration, accumulating R&D experience in bipedal locomotion, dexterous hands, and tactile perception.

Promotional image of the Fourier GR-3 concept robot

For Meta, partnering with a company that already has product and manufacturing experience can reduce the time spent figuring out everything from mechanical structures and components to final assembly, allowing more resources to be devoted to AI models and interaction systems.

Chinese companies' advantages are also reflected in industrial supporting capabilities and real-world application scenarios. Humanoid robots involve multiple links such as motors, reducers, sensors, structural parts, and control systems. China has a relatively complete manufacturing supply chain, as well as a large number of factories and industrial customers that can provide testing environments.

According to foreign media reports, in the first half of 2026, Chinese embodied intelligence companies secured 93.5 billion yuan in financing across 322 deals, equivalent to about 13 billion US dollars, five times the same period last year. Investment has expanded from complete machine manufacturing to robot AI models, training data, and software systems.

Chinese companies are trying to solve the problem of how robots can learn unfamiliar tasks.

Spirit AI has hired about 1,000 people to record daily life and industrial activities, accumulating data for robot training. ACE Robotics uses wearable sensors to collect motion and grip strength information from store and hotel workers.

Ant Group's robotics unit Robbyant said its model used 160,000 hours of training material, including 50,000 hours of human-guided robot motion data, and trained its software on 20 robot configurations from 17 brands.

Another company, DexForce, uses computer-generated environments to train robots, reducing the cost of collecting real-world data.

These companies' technical approaches differ, but their goal is the same: to make robots no longer completely dependent on motion programs pre-written by engineers, but able to adjust their behavior according to changes in the environment and tasks.

US companies are also advancing in this direction. Startups such as Physical Intelligence, Skild AI, and Generalist are developing general-purpose robot models, hoping robots can learn multiple skills and reduce the need for repeated training for a single task.

But Meta's cooperation with Fourier also faces policy restrictions.

In July, the US Federal Communications Commission (FCC), citing potential supply chain and cybersecurity risks, restricted imports of new Chinese humanoid robots. The rules at the time allowed companies to import up to 4,000 devices for testing and product development, but not for sale. In August, the FCC proposed reducing this exemption to 40 units unless companies obtain written approval.

If this proposal takes effect, US companies' imports of robot prototypes from China for testing, as well as continued access to related components, could all be affected.

This gives Meta's cooperation a dual reality: on the one hand, Chinese companies can provide mature hardware R&D and manufacturing capabilities; on the other hand, US restrictions on China's robot supply chain may increase the cost of subsequent cooperation.

However, for Meta, Max has also not yet provided enough public data to prove its actual capabilities. The robot's task completion rate, battery life, payload capacity, and how much human assistance is needed during execution have all not been disclosed. Adding hand and stair sensors does not mean that complex object manipulation and stair traversal have already been solved.

For home robots to truly enter the consumer market, reliability, safety, and cost issues still need to be addressed. What consumers need is a device that can consistently complete specific household chores over the long term, adapt to different home environments, and stop promptly when something unexpected happens—not just a machine that can understand commands and perform actions.

As AI begins to enter the physical world, what will determine product competitiveness is not only how smart the model is, but also whether it can translate understanding into action through a reliable body. For the robotics industry, this is harder than building a humanoid robot that can walk.

Special contributing translator Jin Lu also contributed to this article.

This article comes from "Tencent Technology," author: Zhou Xiaoyan, editor: Xu Qingyang, published by 36Kr with authorization.