Dialogue with Tencent Chief Scientist Zhang Zhengyou: AI is still a "brain in a vat"; Tencent Tairos aims to solve the challenge of embodied implementation

Author | Huang Yu

Over the past two years, embodied intelligence has been in full swing, with "accelerated commercialization" becoming a major label for industry development. At the 2026 World Artificial Intelligence Conference (WAIC), robots showcased their "eighteen skills".

But beneath the heated surface, there is still a gap between the actual capabilities of embodied intelligence and external expectations, and there are issues of homogenization in application scenarios and physical forms.

During a media briefing at WAIC, Zhang Zhengyou, Chief Scientist at Tencent, Director of Tencent Robotics X Lab and Futian Lab, told WallStreetCN and other media outlets that embodied intelligence is still at an early stage, with few genuine applications. Therefore, everyone is searching for scenarios where embodied intelligence can truly play a role, but in the end, the scenarios found are often quite similar.

However, Zhang Zhengyou believes this is not necessarily a bad thing, and everyone should have confidence in practical applications: "Whenever an industry begins, there are a large number of companies involved, but later, the market will sort things out. So I don't think this will hinder the industry's development."

Regarding practical scenarios for robots, Zhang Zhengyou has repeatedly mentioned favoring elderly care scenarios, but he also considers elderly care to be the most challenging.

As for the highly discussed ultra-bionic humanoid robots, Zhang Zhengyou told WallStreetCN that Tencent has not entered this field. In his view, emotional value does not necessarily have to be provided through ultra-bionic means; it should be thought through from first principles.

As early as 2018, Tencent Robotics X Lab was established, dedicated to frontier research and application of robotics technology.

Although this wave of embodied intelligence development is intense, Tencent has always emphasized its boundaries of participation. In early 2025, Tencent Chairman and CEO Ma Huateng stated that Tencent hopes to be a partner to all robotics companies, rather than replacing them and making hardware, which is consistent with Tencent's overall strategic goals.

Against this background, during the 2025 WAIC, Tencent officially launched Tairos "Titan Screw", an open platform for embodied intelligence. This is the first domestic embodied intelligence software platform to provide large models, development tools, and data services in a modular, plug-and-play way for the robotics industry.

Reportedly, over the past year, Tencent Robotics X Lab has cooperated with robotics firms including Unitree, Zhiyuan, Yuejiang, Songyan Power, Leju, Huayan, as well as scenario partners like Bosch, Lens, Jingdezhen, and Dunhuang, exploring the application of embodied intelligence across different robots and industries.

At the 2026 WAIC, Tencent released a series of new embodied intelligence models and intelligent agents, systematically connecting the "perception—body—action" feedback loop for the first time.

Zhang Zhengyou explained that this is a product unfolding along a main line: first letting intelligence understand the physical world, then letting it imagine the target state, and finally turning ideas into stable actions—finally, integrating these abilities into continuously online, reusable intelligent agents.

Reportedly, the embodied base models Tencent released this time include Hy-Embodied-VLM-1.0, Hy-Embodied-RxBrain-1.0, and Hy-Embodied-VLA-0.5, all built atop the Hunyuan large model.

Specifically, the VLM model acts as a "right brain", responsible for image, spatial and scene understanding; the RxBrain model is like the embodied "brain", unifying cognition, planning and imagining future states; the VLA model connects the "cerebellum" and body, transforming high-level goals into continuous, error-correctable actions. 

Hy-Embodied-VLA-0.5's core strength is not training a larger model, but building a complete learning stack with coordinated "data—model—training—deployment", accumulating over ten thousand hours of human demonstration data through a UMI ultra-high precision sub-millimeter collection system.

Zhang Zhengyou noted that true intelligence requires connecting language, vision, spatial cognition, body control, and environmental feedback, and receiving validation in the "perception—body—action" feedback loop. 

This is precisely the basic approach of Tencent Robotics X in exploring native embodied intelligence: moving from disembodied to embodied, from segregated modules toward an integrated feedback loop.

In Zhang Zhengyou's view, today's smartest artificial intelligence is essentially still a "brain in a vat". In recent years, progress in large models has been astonishing, but once entering the real world, problems arise: it may not truly understand the positions of objects, spatial relationships, or operability; and it is difficult, in a continuously changing environment, to act while receiving feedback and adjusting in real time.

"This is like a brain raised in a vat—rich in knowledge, but without a body and lacking a closed loop of direct interaction with the world. We call this 'disembodied intelligence.'"

Data scarcity is a key factor hindering embodied intelligence from becoming smarter.

Zhang Zhengyou points out that in embodied intelligence’s digital pyramid, the lowest layer is internet data, followed by first-person perspective operational videos, second is data collected via sensor gloves, and at the top is teleoperation data. Although teleoperation data is collected less efficiently, it is regarded as the best for training robots.

Additionally, Zhang Zhengyou states that robot training requires data from web videos, first-person operational videos, sensor glove data, and teleoperation data. To make robots smarter through training, these four types of data need to be integrated.

Unlike the technology stack for large language models, which has converged on an industry consensus, Zhang Zhengyou says that the technical solutions for embodied intelligence models have yet to reach agreement. Last year the industry favored the VLA (Visual-Language-Action) paradigm, but this year discussions are more about world models—with Hy-Embodied-RxBrain-1.0 representing Tencent’s exploration of world understanding models.

Risk Warning and DisclaimerThe market has risks; investment requires caution. This article does not constitute personal investment advice nor takes into account special investment objectives, financial situations, or needs of individual users. Users should consider whether any opinion, viewpoint, or conclusion in this article suits their circumstances. Investing based on this is at your own risk.