In the past two years, the field of embodied AI has seen a surge in activity, with "accelerated commercialization" becoming a major theme. At the 2026 World Artificial Intelligence Conference (WAIC), robots showcased a wide array of capabilities.
However, beneath this vibrant surface, the actual capabilities of embodied intelligence still fall short of external expectations. There are also issues of homogenization in application scenarios and physical forms.
During a media exchange at WAIC, Zhang Zhengyou, Chief Scientist at Tencent, Director of the Tencent Robotics X Lab, and Director of the Futian Lab, noted that embodied AI is still in its infancy, with very few truly implemented applications. Consequently, everyone is searching for scenarios where embodied AI can genuinely be useful, and the final scenarios identified may end up being quite similar.
Nevertheless, Zhang believes this is not necessarily a bad thing and that confidence in practical deployment is warranted. "In the early stages of any industry, a multitude of companies participate, but the process naturally filters out the less viable ones. I don't think this will hinder the industry's development."
Regarding application scenarios for robots, Zhang repeatedly expressed optimism about eldercare, while also acknowledging it as one of the most challenging areas.
Concerning the currently popular hyper-realistic humanoid robots, Zhang clarified to media that Tencent is not involved in that specific domain. In his view, emotional value does not necessarily have to be provided through hyper-realism; it requires thinking from first principles.
Tencent's Robotics X Lab was established as early as 2018, dedicated to the research and application of cutting-edge robotics technologies.
Despite the current fervor surrounding embodied AI, Tencent has consistently emphasized its strategic boundaries. In early 2025, Pony Ma, Chairman and CEO of Tencent, stated that the company aims to be a partner to all robot manufacturers rather than replacing them in hardware production, aligning with Tencent's overall strategic goals.
It is within this context that during the 2025 WAIC, Tencent officially launched its embodied AI open platform Tairos. This platform is the first domestic software platform for embodied AI to provide large models, development tools, and data services in a modular fashion, offering plug-and-play functionality to the robotics industry.
It is reported that over the past year, the Tencent Robotics X Lab has collaborated with robotics companies such as Unitree, Zhiyuan, Yuejiang, Songyan Power, Leju, and Huayan, as well as scenario partners including Bosch, Lansi, Jingdezhen, and Dunhuang, to explore the implementation of embodied intelligence across various robots and industrial settings.
During the 2026 WAIC, Tencent introduced a new series of embodied AI models and intelligent agent achievements, systematically establishing a closed loop of "perception-body-action" for the first time.
Zhang explained that this product line follows a central theme: first enabling the AI to understand the physical world, then allowing it to envision a target state, and finally translating that intention into stable actions. These capabilities are then integrated into a continuously online, reusable intelligent agent.
The newly released embodied foundation models include Hy-Embodied-VLM-1.0, Hy-Embodied-RxBrain-1.0, and Hy-Embodied-VLA-0.5, all built upon Tencent's Hunyuan large model.
Specifically, the VLM model functions like a "right brain," responsible for understanding images, space, and scenes. The RxBrain model acts as the embodied "brain," unifying cognition, planning, and imagination of future states. The VLA model connects the "cerebellum" and the body, translating high-level goals into continuous, error-correctable actions.
It is reported that the core strength of Hy-Embodied-VLA-0.5 lies not in training a larger model, but in constructing a complete learning stack that synergizes "data-model-training-deployment." This is supported by accumulating over 10,000 hours of human demonstration data through a sub-millimeter high-precision UMI collection system.
Zhang pointed out that true intelligence requires the integration of language, vision, spatial cognition, body control, and environmental feedback, validated within the closed loop of "perception-body-action."
This aligns with the fundamental approach of the Tencent Robotics X Lab in exploring embodied native intelligence: moving from disembodied to embodied, and from fragmented modules to an integrated closed loop.
In Zhang's view, today's most advanced AI is essentially still a "brain in a vat." While progress in large models over recent years has been astonishing, problems arise once they enter the real world: they may not truly understand object positions, spatial relationships, and manipulability; they also struggle to act, receive feedback, and make timely adjustments within a continuously changing environment.
"It's like a brain raised in a vat—possessing rich knowledge but lacking a body and the closed-loop interaction with the world. We call this state 'disembodied intelligence.'"
The scarcity of data for embodied intelligence is a key factor hindering its advancement.
Zhang noted that in the digital pyramid for embodied AI, the base layer consists of internet data, followed by first-person perspective operation videos, then data collected with sensor-equipped gloves, with the top layer being teleoperation data. Although teleoperation data is inefficient to collect, it is considered the most effective for training robots.
Furthermore, Zhang stated that the data required for robot training encompasses web videos, first-person perspective operation videos, data from sensor gloves, and teleoperation data. To make robots more intelligent through training, these four categories of data need to be integrated.
Unlike the large language model technology stack, which has converged to a widely accepted scope, Zhang indicated that there is no consensus yet on the technical approach for embodied AI models. Last year, the industry saw the rise of the VLA (Vision-Language-Action) paradigm, while this year, discussions focus on world models. Hy-Embodied-RxBrain-1.0 represents Tencent's exploration in world understanding models.
Comments