Tairos

Scroll sideways for more →

Tairos (Chinese: 钛螺丝, "titanium screw") is an embodied-AI platform for robotics built by Tencent. It bundles foundation models, agent frameworks, simulation and development tools, and cloud services so that robot makers can add perception, reasoning, and action software to their machines without building those layers themselves. Tencent frames Tairos as a platform play rather than a hardware play: the company does not sell its own commercial humanoid, and positions itself, in the words of Robotics X Lab director Zhang Zhengyou, as a partner to robot manufacturers rather than a competitor to them.^[1][10][14]

What Tairos is

Tairos is a software and cloud layer that sits between robot hardware and robot applications, described in Tencent's own platform document as "an embodied AI platform for robotics applications."^[2] Tencent describes it as China's first embodied-intelligence software platform to provide large models, development tools, and data services in a modular way, opening those capabilities to the robotics industry through standardized interfaces and software development kits (SDKs).^[10][12] The stated goal is to give a robot body what Tencent calls a "brain": the ability to perceive its surroundings, plan tasks, and make decisions, with the underlying models and cloud services supplied by Tencent while the physical machine, actuators, and end applications are supplied by partners.^[10][11]

Two things are worth separating at the outset. First, Tairos is the platform and brand; the Hy-Embodied models are the foundation models that run on it. Second, "open" in Tencent's framing mostly means "open to developers" through SDKs and cloud APIs, plus a subset of models released under a genuine open-source license, rather than the entire stack being open-source. The cloud services, including the Embodied-AI-as-a-Service (EaaS) offering, are commercial products sold through Tencent Cloud.^[1][3][4]

Background: Tencent Robotics X and a platform-not-hardware strategy

Tairos comes out of Tencent Robotics X Lab, the company's robotics research group established in 2018 and led by Zhang Zhengyou (张正友), Tencent's chief scientist. Zhang is a computer-vision researcher best known outside Tencent for "Zhang's method" of camera calibration, and he also directs Tencent's Futian Lab in Shenzhen.^[9][17] Before Tairos, Robotics X had built research prototypes including quadruped robots, dexterous hands, and eldercare-oriented robots, but Tencent kept these as research projects rather than turning them into commercial hardware.^[15]

The platform strategy is deliberate. Zhang has said that after surveying dozens of robotics companies, Tencent concluded that many firms are strong at hardware but under-invest in software because software is, in his phrasing, an endless money pit; Tairos is meant to fill that gap for the industry rather than for a single Tencent product.^[14] He has repeatedly described current large models as a "brain in a vat" (缸中之脑): rich in knowledge but without a body or a closed loop of interaction with the physical world, which he argues is the missing ingredient for real intelligence.^[9] Tencent's public messaging is that it wants to "develop the entire robot industry as quickly as possible" and that Tairos is not primarily aimed at profit, a claim that should be read as strategic positioning rather than a verified financial fact.^[14][15]

This puts Tencent in the same broad camp as other large Chinese technology companies that have chosen to supply models and cloud services to robot makers rather than build robots themselves. Coverage of the sector has grouped Tencent with JD.com as "infrastructure provider" plays, in contrast to the many startups racing to ship humanoid hardware.^[15] It also fits Tencent's wider AI approach, where the Hunyuan model family and Tencent Cloud provide the underlying model and compute layers.

Architecture: a three-tier cloud plus a layered platform

There are two distinct "three-layer" structures in Tairos, and Tencent's own materials sometimes blur them, so it helps to keep them apart.

The first is Tencent Cloud's infrastructure for embodied intelligence, which the company describes as a three-tier architecture spanning computing infrastructure, model services, and perception and interaction.^[1] Chinese coverage of the WAIC 2026 launch mapped concrete Tencent Cloud products onto those tiers:^[5]

Scroll sideways for more →

On top of that cloud, Tencent launched the Embodied-AI-as-a-Service (EaaS) offering, which it calls its first cloud-based embodied-AI service, developed together with industry partners to move embodied intelligence from technical development toward broader real-world adoption.^[1]

The second structure is the Tairos platform itself, which Tencent's product site organizes into three layers: embodied foundation models, embodied agents, and platform tools.^[3] The table below lists the named components most often cited across Tencent's WAIC 2026 announcement and product materials, each with its role. It also includes ADP 4.0, which appears in the same announcement but is a separate product, for clarity.

Scroll sideways for more →

Tencent describes TairosAgent's design as native to embodied tasks, built around a perception-reaction layer, a cognition-decision layer, and an exploration-accumulation layer, with a three-part memory covering the environment, the robot's own body, and its tasks.^[5] Apexio is presented as a multi-frequency online agent: a cognitive system that runs on demand, a perception-and-action system operating at roughly 15 Hz, and an execution system that handles high-frequency reflex responses.^[5] These are Tencent's own architectural descriptions of announced systems; independent benchmarks of the agents in third-party deployments are not yet available.

The relationship between Tairos and ADP

Because both were announced at WAIC 2026, Tairos and ADP are easy to conflate, but they are different products. Tairos is the embodied-intelligence platform for physical robots. ADP (Agent Development Platform) 4.0 is Tencent Cloud's platform for building enterprise software agents, and the news at WAIC 2026 was its global or overseas edition, which adds integrations such as LINE and Telegram messaging, Google Workspace connectivity, timezone and language localization, and support for international large language models.^[7][19] ADP builds software agents that live in apps and workflows; Tairos builds intelligence for machines that move in the world. They share Tencent Cloud as a foundation and a common "agent" vocabulary, but ADP is not a component of the Tairos embodied stack.

The Hy-Embodied models on the platform

The models that provide Tairos's intelligence are the Hy-Embodied family, built on Tencent's Hunyuan technology and covered in detail in Hy-Embodied. In brief, Tencent presented three at WAIC 2026, each mapped to a part of a robot's "nervous system":^[1][5][16]

• Hy-Embodied-VLM-1.0, a vision-language model for perception, which Tencent likens to the "right brain" for understanding images, space, and scenes. Tencent says it reaches near-flagship perception quality at roughly one-tenth of the compute of its previous flagship, citing an A3B-scale model performing close to an A32B-scale one.^[1][5]
• Hy-Embodied-RxBrain-1.0, the reasoning engine, described as the embodied "brain" that unifies cognition, planning, and imagination of future states, combining language reasoning with visual understanding to support decisions. It functions in the spirit of a world model for the robot.^[1][16]
• Hy-Embodied-VLA-0.5, a vision-language-action model that connects the "cerebellum" and body, turning high-level goals into continuous, correctable actions. Tencent says it was trained on more than 10,000 hours of high-quality human demonstration data collected with sub-millimeter-precision systems in the style of the Universal Manipulation Interface, and that it can be deployed across different robot types.^[1][9][16]

One member of the family, HY-Embodied-0.5-X, is a publicly released model rather than a purely cloud-hosted one. Its GitHub repository describes a 4-billion-parameter mixture-of-transformers model with about 2 billion active parameters, positioned as an edge-friendly foundation model, and released under the Apache-2.0 license, which permits commercial use.^[4] Tencent's repository claims it reaches state-of-the-art results on ten mainstream embodied task-planning benchmarks and ranks first among edge-side domain models on seven of them; these are the developer's own reported numbers and have not been independently reproduced here.^[4]

Simulation and tooling: PlaygroundX

PlaygroundX is Tairos's simulation framework, used to train and test robot behavior in a virtual environment before it runs on real hardware. Tencent's product materials describe the platform's simulation as built on NVIDIA's Isaac Sim and accessible as a one-click cloud service, which places PlaygroundX in the same family of physics-based robotics simulators used across the industry.^[3] The open-source HY-Embodied-0.5-X model integrates with PlaygroundX to demonstrate closed-loop task planning: given a household instruction such as "put the tomato in the fridge" or "throw the potato into the trash," the model produces a full plan and then adjusts on the fly, following a loop of reason, execute, detect failure, and replan.^[4]

Alongside simulation, Tairos exposes a hardware abstraction layer (HAL) and a connectivity layer that Tencent calls RoboFusion. The HAL standardizes the low-level differences between robots so that a skill built once can, in principle, run on different bodies without being rewritten, which is the mechanism behind Tencent's "plug-and-play across form factors" claim for TairosAgent.^[3] These tools, together with data-management and corpus utilities, are what let a developer move from a model in the cloud to behavior on a physical robot. Independent verification of how smoothly this works across heterogeneous hardware is limited, since the platform is new.

Supported robots and partners

Tencent does not sell the robots; it supplies the software that runs on other companies' machines. At the WAIC 2025 launch, Tencent named a first group of partner robot companies whose platforms Tairos supports.^[10][12]

Scroll sideways for more →

At WAIC 2026 Tencent's newsroom described Tairos as having been deployed across robot platforms from Unitree, PaXini, Leju, Dobot, and Engine AI.^[1] Chinese coverage of the 2026 event named an overlapping but broader set of collaborators, including AgiBot (智元) and Songyan Power (松延动力, also known as Noetix Robotics), plus industrial partners such as Bosch; these appear in secondary reporting rather than in a single authoritative Tencent list, so the exact roster is best treated as an ecosystem in flux.^[5] The brief that seeded this article suggested "Engine AI" might be the mobile-robot firm Songling (松灵); the primary launch coverage instead identifies the humanoid maker EngineAI / Zhongqing (众擎), so that is the company recorded here.

WAIC 2026 launch and the Xiaoliu demo

Tairos first launched on 2025-07-27 at WAIC 2025, jointly released by Tencent Robotics X Lab and Futian Lab,^[13] and at WAIC 2026 (2026-07-17 to 2026-07-20 in Shanghai) Tencent upgraded it into what it billed as a full-stack embodied-intelligence solution, presenting the three Hy-Embodied models, TairosAgent, Apexio, and the EaaS cloud as one system for the first time.^[1][5][10] Tencent said the platform would remain open and continue to provide models, agents, and development tools to lower the barrier along the chain from model to robot body to application.^[5][16]

The most-watched demonstration was a robot named 小六 (Xiaoliu, "Little Six"), the new demonstration robot from Robotics X Lab and Tencent's Futian Lab. Xiaoliu is the iteration of 小五 (Xiaowu, "Little Five"), the lab's fifth-generation experimental robot released on 2024-09-24; both share a design philosophy of coexisting safely with people and moving in a human-like way, and Xiaoliu adds tendon-and-cable (键绳) driven transmission.^[6][17] Xiaowu itself, a wheeled-legged machine with large-area tactile skin and a multi-finger dexterous hand, had appeared at WAIC 2025 as a home assistant that could understand speech and greet visitors.^[12]

At WAIC 2026, Xiaoliu performed traditional Chinese-medicine massage (tuina) on visitors as part of an eldercare and wellness demonstration, running on the Hy-Embodied model stack. Tencent says the massage technique was captured with a proprietary data-collection system that fuses force, tactile, and visual sensing, then trained with reinforcement learning to reproduce both the trajectory and the pressure of the technique precisely; Tencent has said it plans to teach Xiaoliu more two-arm massage methods and to iterate the robot in real settings such as nursing homes.^[16][17] The demo drew crowds and became one of the talked-about exhibits of the conference; the fuller technical treatment of the robot's model stack is in Hy-Embodied. As of the announcement, Tencent had not disclosed Xiaoliu's full specifications (degrees of freedom, force limits) or any commercialization timeline, so the demonstration should be read as a capability showcase rather than a shipping product.^[6]

Significance and reception

Tencent's move has been widely read as a relatively quiet entry into a crowded field.^[20] While many exhibitors at WAIC brought humanoid hardware, Tencent brought a software platform and framed it as long-term infrastructure investment rather than a product line, an approach Chinese technology press described as Tencent "entering the game low-key."^[14] The bet is that as the sector matures, the scarce ingredient will not be another humanoid body but the software brain that makes any body useful, and that a neutral supplier of that brain can serve the whole industry.

Tencent executives have leaned on an engineering, rather than model-size, argument for why a platform matters. Tencent vice-president Lin Songtao said at WAIC 2026 that the gap between models is shrinking while the gap in engineering is widening, and that engineering has become the decisive factor, arguing that models set an application's ceiling but engineering sets its floor.^[8] For embodied AI specifically, that translates into the unglamorous work of simulation, data pipelines, hardware abstraction, and real-time control that Tairos tries to package.

Zhang Zhengyou has been candid about the limits of the current moment. He has acknowledged that early demonstrations across the industry look homogeneous, expressed the view that a shakeout will follow as in any young industry, and estimated that genuinely hard applications such as eldercare are perhaps another three years away from technical maturity.^[9][14] He has also questioned the industry's fixation on the human form, noting that bipedal locomotion evolved for ancient natural environments rather than modern flat interiors, and describing Tairos as deliberately morphology-agnostic so that it can serve bipeds, quadrupeds, and arms alike.^[14]

How it compares to other embodied-AI platforms and stacks

The closest analogue to Tairos is NVIDIA's robotics stack. NVIDIA pairs the Isaac Sim simulator and Isaac Lab with the Isaac GR00T foundation-model program, including the GR00T N1 humanoid model, to offer simulation, models, and tools that robot builders can adopt without training everything themselves. Tairos occupies the same "we supply the brains and tools, you build the robot" position, and, tellingly, Tencent's own simulation layer is built on NVIDIA Isaac Sim, so the two are partly complementary rather than purely competitive.^[3] The main differences are ecosystem and geography: NVIDIA's stack is tied to its GPUs and has a global developer base, while Tairos is tied to Tencent Cloud and is, for now, centered on the Chinese robotics industry and its partners.

Tairos also sits alongside a growing set of open robot-learning efforts. Open vision-language-action models such as OpenVLA and the pi-zero line, and open robot-learning frameworks more broadly, give developers models and recipes without a single vendor's cloud. Tairos is more integrated and more cloud-centric than those community stacks, trading some openness for a managed end-to-end pipeline, though its release of HY-Embodied-0.5-X under Apache-2.0 puts at least part of the stack on the same open footing.^[4] Compared with humanoid-first companies that build both the body and the brain, such as those behind commercial humanoid programs, Tencent's platform is broader in reach but thinner in vertical integration: it depends entirely on partners for the hardware it runs on.

Limitations and open questions

Tairos is early and, so far, largely unproven at scale. Almost every capability claim about the platform originates with Tencent's own newsroom, product site, or model repository; independent, third-party evaluation of the models, agents, and end-to-end pipeline in production robots is not yet available. The distinction between announced and shipped matters here: the WAIC 2026 event demonstrated capabilities and released one open model, but it did not establish a track record of robots running Tairos in daily commercial use.^[1][4]

Several open questions follow. The platform's success depends on adoption by robot makers who may prefer to build their own end-to-end models rather than accept a modular outside brain, a tension Zhang Zhengyou himself has acknowledged.^[15] "Open" is doing a lot of work in the marketing: one model is genuinely open-source under Apache-2.0, but the broader platform is an open-to-developers cloud service sold commercially, and the licensing and pricing of most components are not fully public.^[3][4] The stack is Tencent-Cloud-centric and, for now, China-focused, which shapes both its data and its likely customer base. And because Tairos does not build hardware, its usefulness is bounded by the quality and variety of the partner robots it runs on. None of this makes the platform strategy wrong, but it does mean the thesis is still a bet on an ecosystem forming, not a finished result.

ELI5

Imagine robots are like bodies without brains. Some companies are really good at building the body: the arms, legs, hands, and motors. But a body is not useful unless it can see what is around it, understand what you ask, and decide how to move. That thinking part is hard and expensive to build.

Tencent, the company behind WeChat, decided it would rather build the brain than the body. Tairos is that brain, plus a toolbox and an online workshop. A robot company can take Tencent's models (the parts that see, think, and act), test their robot in a computer simulation first, and then plug the brain into their real robot. Tencent even showed a robot named Little Six that learned to give a proper Chinese massage by watching experts and practicing. Tencent is not trying to sell its own robot; it wants to be the brain and toolkit that lots of different robots use.

See also

• Hy-Embodied
• Embodied AI
• Tencent
• Hunyuan
• Isaac GR00T
• Robot foundation model
• Unitree
• World Artificial Intelligence Conference

References

1. "Tencent Unveils Full-Stack Embodied Intelligence Solution; ADP 4.0 Launches Globally at WAIC 2026," Tencent (tencent.com), July 2026. https://www.tencent.com/tencent-unveils-full-stack-embodied-intelligence-solution-at-waic-2026/ ↩
2. "TAIROS: An Embodied AI Platform for Robotics Applications," Tencent (platform white paper PDF). https://cdn.tairos.tencent-cloud.com/downloads/tools/pdf/tairos.pdf ↩
3. Tairos 具身智能开放平台 (official platform site), Tencent. https://tairos.tencent.com/ ↩
4. "HY-Embodied-0.5-X: An Enhanced Embodied Foundation Model for Real-World Agents," Tencent-Hunyuan GitHub repository (Apache-2.0 license). https://github.com/Tencent-Hunyuan/HY-Embodied-0.5-X ↩
5. "直击WAIC丨腾讯发布多款具身智能基座模型与智能体," Sina Finance (finance.sina.com.cn), 2026-07-19. https://finance.sina.com.cn/hy/hyjz/2026-07-19/doc-iniiimtc7924002.shtml ↩
6. "腾讯 Robotics X 实验室新一代示范性机器人'小六'首次公开亮相，系'小五'迭代之作," Sina Tech (finance.sina.com.cn), 2026-07-17. https://finance.sina.com.cn/tech/digi/2026-07-17/doc-iniickay1505497.shtml ↩
7. "Tencent Unveils Full-Stack Embodied AI Strategy, Launches ADP 4.0 Overseas Edition, and Brings WorkBuddy App to HarmonyOS," BigGo Finance, 2026-07-18. https://finance.biggo.com/news/3ab476ff-f4e4-4219-9f74-92485a07fb2b ↩
8. "Tencent Unveils Embodied AI Full-Stack Solution at WAIC; Lin Songtao Says AI Engineering Gap Becomes Decisive Factor," BigGo Finance, July 2026. https://finance.biggo.com/news/acdbeb35-c150-463b-a334-0532e3ed1144 ↩
9. "对话腾讯首席科学家张正友：AI仍是'缸中之脑'，腾讯Tairos欲破具身落地难题," NetEase (163.com), 2026. https://www.163.com/dy/article/L291B0RK05198NMR.html ↩
10. "腾讯联合福田实验室发布具身智能开放平台Tairos，打造机器人领域的'钛螺丝'," Tencent News (news.qq.com), 2025-07-27. https://news.qq.com/rain/a/20250727A03TN400 ↩
11. "腾讯具身智能平台Tairos登场 欲做机器人厂商'大脑'," Sina Finance (finance.sina.com.cn), 2025-07-28. https://finance.sina.com.cn/roll/2025-07-28/doc-infhzyak4624922.shtml ↩
12. "腾讯Robotics X 具身智能开放平台Tairos正式发布，与多家行业领军机器人公司展开合作," China Fund (chnfund.com), 2025-07. https://www.chnfund.com/article/ARc55ff0b4-8b80-c9a5-056a-3a1b5d7491c0 ↩
13. "腾讯联合福田实验室发布具身智能开放平台Tairos," Economic Information Daily / Xinhua (jjckb.xinhuanet.com), 2025-07-28. http://jjckb.xinhuanet.com/20250728/ea7fd8dbba184e32884c5dafc10a766a/c.html ↩
14. "具身智能，腾讯低调入局," GeekPark (geekpark.net), 2025. https://www.geekpark.net/news/351962 ↩
15. "Tencent and JD.com's Robot Strategy: Focus on Platform Building Instead of Hardware Development," 36Kr (eu.36kr.com), 2025. https://eu.36kr.com/en/p/3403965013069441 ↩
16. "2026WAIC丨让具身智能听懂人，腾讯一次发布三大具身基座模型," Yangcheng Evening News (news.ycwb.com), 2026-07. https://news.ycwb.com/ikimvkltjc/content_54237207.htm ↩
17. "腾讯机器人按摩师走红WAIC：'要送进养老院'," NetEase (163.com), 2026-07. https://www.163.com/dy/article/L25FK4K105129QAF.html ↩
18. "Unitree Technology and Tencent Robotics X Lab Form a Strategic Partnership," Gasgoo (autonews.gasgoo.com), 2025. https://autonews.gasgoo.com/articles/market-industry/unitree-technology-and-tencent-robotics-x-lab-form-a-strategic-partnership-2008804138454663168 ↩
19. "2026 WAIC | 腾讯升级发布具身智能全栈方案，ADP 4.0海外版正式上线," Zhonghong Net (zhonghongwang.com), 2026-07. https://www.zhonghongwang.com/show-140-466717-1.html ↩
20. "Tencent's AI Strategy and WAIC Product Launches," AI Proem (aiproem.substack.com), 2026-07. https://aiproem.substack.com/p/tencents-ai-strategy-and-waic-takeaways ↩