In one sentence: at WAIC 2026, Tencent released the Hy-Embodied series of three embodied foundation models that systematically connect the perception-cognition-action loop for the first time — and the VLA model has already run real trials in a daily-chemicals factory: 95%+ success rate, 6 seconds per item, and new-SKU adaptation in under 3 days.
The “Brain in a Vat” Problem of Embodied AI
Today’s smartest large models are, in essence, brains in vats — superb at reasoning, generation, and Q&A, but ignorant of where objects sit, how spaces relate, what an object affords, and unable to adjust while acting in a continuously changing environment.
Tencent chief scientist Zhang Zhengyou put it precisely on the WAIC stage: “Language is not the whole of cognition — it is just one channel through which intelligence expresses itself outward.” Real intelligence requires wiring language, vision, spatial cognition, body control, and environmental feedback into one circuit, validated inside a perception-body-action loop. That is the question Tencent’s five new results set out to answer.
Three Models, Three Roles
The three embodied foundation models from Tencent’s Robotics X lab together with Hunyuan map onto three layers of embodied capability:
Hy-Embodied-VLM-1.0 — the “right brain,” for seeing. It solves three increasingly deep problems: seeing an object and knowing its use (depth, orientation, affordance points); seeing a scene and knowing what changes (judging the next action and its variations); and seeing far and finding the way back (remembering history across long-horizon tasks and replanning after failure). The key figure: at only A3B scale — roughly one-tenth the compute — it approaches the previous A32B flagship’s performance, which means it can deploy directly on a robot’s own hardware.
Hy-Embodied-RxBrain-1.0 — the “brain,” for thinking. It unifies reasoning with imagination: given a task, it produces linguistic step constraints while generating a target image of what each completed step should look like. The downstream action model doesn’t just “hear what to do” — it “sees what done looks like.” A new cognitive paradigm: anchoring the chain of reasoning in visual imagination.
Hy-Embodied-VLA-0.5 — the “little brain,” for doing. Its competitive edge is not a bigger model but a complete learning stack where data, model, training, and deployment reinforce each other: sub-millimeter UMI data capture → over 10,000 hours of human demonstration data → reinforcement learning training. It took first place in overall simulation on the RoboDojo evaluation, leading specifically on long-horizon and memory tasks.
From Assembling Modules to Building a Loop
The mainstream approach to embodied AI has been module assembly — take a vision model, a language model, a control model, and stitch them. The problem: each module optimizes itself, and the seams between them leak. Tencent’s approach is native closed-loop design — three models designed from day one to work together, coordinated by two agent layers: Apexio, a continuously online agent, and TairosAgent, a native framework. Apexio’s three-tier architecture deserves particular attention:
-
Cognition layer: wakes on demand for deep thinking;
-
Perception-action layer: ingests multimodal input and adjusts at roughly 15 Hz;
-
Execution layer: runs at even higher frequency, handling collisions and imbalance like reflexes.
Add the dual throughlines of “goal-driven” (complete the task) and “survival-driven” (safety, energy, state management), and the robot keeps perceiving and self-preserving even without external commands.
What 95% in a Real Factory Means
The most important numbers come from a production floor. In live trials at a daily-chemicals plant, HyVLA-0.5 delivered results that would have been dismissed as slide-ware two years ago — and each number targets a distinct reason factory pilots fail:
-
Task success rate above 95%;
-
Cycle time under 6 seconds per item;
-
New-SKU data collection and model post-training in under 3 days.
Together these three numbers say something specific: embodied AI has stopped being a lab demo and can handle real production lines characterized by high mix, small batches, and frequent SKU turnover. The 95% answers the reliability objection, the 6-second cycle answers the economics objection, and the 3-day SKU adaptation answers the rigidity objection — historically the exact place robotics pilots went to die. A line that retools in three days keeps pace with the consumer-goods industry’s iteration rhythm, which means the robot stops being a fixed asset and starts behaving like flexible capacity.
Embodied AI’s Android Moment?
TairosAgent’s design implies a bigger ambition: through a unified hardware adaptation layer and standardized Skill interfaces, it aims to connect different models to different robot forms. The logic mirrors early Android, whose standardized interfaces let heterogeneous hardware vendors plug in. Tencent named the platform “Tairos” — the titanium screw — positioning it as “the titanium screw connecting models, bodies, tools, and applications.” If that position holds, embodied AI may be approaching its Android moment: not one robot winning, but a platform on which every robot runs.
What To Do With This
Robot developers: Hy-Embodied-VLM-1.0 is open source and runs on-device at A3B scale — evaluate now whether it fits your robot form factor.
Manufacturing decision-makers: 95% success plus 3-day SKU adaptation means embodied AI has crossed the “can it work” threshold into “is it worth investing.” High-mix, low-batch settings — daily chemicals, consumer goods — are the natural entry point.
AI practitioners: study the co-design methodology across data, model, training, and deployment. Embodied competition is not about any single model — it is about the closed-loop efficiency of the entire learning stack, and Tencent’s sub-millimeter capture hardware feeding 10,000 hours of demonstration data feeding RL feeding on-device deployment is currently the most complete public example of that stack working end to end.
