--- title: "Tencent's \"Tairos screw\" bets on robot brains, not bodies" date: 2026-10-09 category: Humanoids & Robotics site: NeuroAI canonical: https://neuroai.site/a/na-robot-tencent-robotics-x language: en --- # Tencent's "Tairos screw" bets on robot brains, not bodies > At WAIC 2025 Tencent opened Tairos, a modular embodied-AI platform for robot makers — a software play that says the future of humanoid (人形机器人) deployment is in the "brain," not the hardware. A robot walks a trade-show floor, greets visitors, and makes coffee when asked. Behind it sits not a factory, but a cloud account. That is the picture Tencent is trying to sell: the valuable part of a robot is the software that thinks, and most hardware companies are not very good at that part yet. In late July 2025, at the World Artificial Intelligence Conference (WAIC) in Shanghai, Tencent's Robotics X lab and the Futian Lab jointly released Tairos — branded "Tairos screw" (钛螺丝). It is described as China's first modular embodied-AI open platform, built to hand robot makers a ready-made "brain" through plug-and-play modules. **Why a "screw," not a robot** Tencent is explicit about what it will not do. The company is not shipping a humanoid (人形机器人) to compete with the hardware startups. Zhang Zhengyou (张正友), Tencent's chief scientist and director of both labs, repeated the line at WAIC: Tencent wants to be the partner of every robot maker, not its replacement. Pony Ma (马化腾) said much the same at Tencent's annual-report briefing in March 2025. The metaphor is deliberate. A screw is small, cheap, and invisible — but a machine falls apart without it. Tencent's bet is that as embodied AI (具身智能) moves from labs to factories and homes, the bottleneck is software: perception, planning, and the link between "I see" and "I act." Hardware companies can build arms and wheels; wiring in intelligence is harder. **What Tairos actually ships** Tairos splits into two layers: model algorithms and cloud services. - The multimodal perception model is the "right brain." It turns camera and sensor input into a layered 3D map of the environment and answers natural-language questions about what is around the robot. - The planning large model is the "left brain." It understands the robot's role and goal, breaks a task into executable steps, and can call external tools. - The perception-action joint model is the "cerebellum." It closes the gap between seeing and doing, turning understanding into safe, precise motion commands. On the cloud side, Tairos offers a simulation environment, a data platform, and development tools, all callable through standard interfaces and an SDK. The simulation environment ships with preset scenes and objects so a team can validate a model in the cloud with one click before touching real hardware. **Who is already plugged in** At launch, Tencent named an initial partner list that reads like a roll-call of China's robot makers: Dobot (越疆), Leju (乐聚), PaXini (帕西尼感知), Keenon (擎朗), Zhongqing (众擎), and Unitree (宇树). On the WAIC floor, a Unitree G1 running Tairos acted as a tour guide, a home-assistant prototype called "Xiao Wu" (小五) hugged visitors, and a Dobot arm made food after a voice command. Before Tairos, the same lab had spent years building its own research robots: the quadruped Max, the wheel-legged Ollie, the TRX-Hand dexterous hand, the TRX-Arm, and the "Xiao Wu" human-environment prototype. Those were research prototypes, not products — and that history is the point. Tencent says it learned enough building hardware to know it should stay a software supplier. **The strategic logic** The timing matters. From the second half of 2023, China's robot field changed fast: ChatGPT pulled attention to large models, and a wave of hardware startups appeared. Tencent's read is that the ecosystem now needs shared software infrastructure more than another prototype. Tairos is offered module-by-module, so a company strong in perception can license only the planning model, and vice versa. That is a different wager from rivals who sell full stacks or full robots. If robot makers standardize on Tairos the way many apps standardized on cloud services, Tencent captures the layer that is hardest to rebuild and hardest to switch away from. ## Honest limitations Figures and capabilities here are company-disclosed at WAIC 2025; independent benchmarks of Tairos are not yet public. Partner names reflect launch announcements, and commercial terms between Tencent and each robot maker have not been detailed. The "left brain / right brain / cerebellum" framing is Tencent's own metaphor, not a peer-reviewed taxonomy. Real-world reliability of the platform across unstructured environments remains unproven at scale. ## What readers can do now - If you build or buy robots, watch whether Tairos stays open or drifts toward Tencent's preferred partners — that determines its usefulness as neutral infrastructure. - Robotics engineers should test the free SDK and simulation environment against a concrete task before committing to a cloud dependency. - For anyone tracking China's embodied-AI race, treat "brain vs. body" as the real split: the platforms that own the software layer may matter more than the next flashy humanoid demo. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-robot-tencent-robotics-x Free to quote with attribution and a link to the original.