Tencent open-sources three embodied foundation models, chief scientist details 'three-layer brain' for faster robot response

At WAIC 2026, Tencent officially open-sourced three embodied foundation models — Hy-Embodied-VLM-1.0, Hy-Embodied-RxBrain-1.0 and Hy-Embodied-VLA-0.5 — and launched the always-on embodied agent Apexio together with the native agent framework TairosAgent. Zhang Zhengyou, Tencent's chief scientist and head of the Robotics X Lab, spoke with media including InfoQ in Shanghai, reviewing a year of progress since the Tairos open embodied-intelligence platform debuted.

The lineup addresses a fact that big-model narratives often obscure: the world a robot faces does not run at a single speed. Zhang argues that a common pain point in today's VLA (vision-language-action) approaches is that visual understanding and action generation execute at the same frequency, forcing cognition, perception and execution to share one compute chain. In the physical world, though, tactile feedback arrives in about 1 millisecond and low-level motor control runs at 500Hz to 1000Hz: running a large model at control-loop frequencies is computationally unaffordable, while making low-level actions wait on the model turns latency into a safety risk.

Tencent's answer is "layered operation, closed-loop training." The top cognition system wakes on demand for complex task understanding and deep reasoning; the middle perception-action system continuously ingests multimodal information at about 15Hz and adjusts actions quickly; the bottom execution system runs at much higher frequency, handling collisions and loss of balance like reflexes. Zhang expects the architecture to stay stable for a long time: "I think layering is a reasonable architecture, and it will not change much over a very long period."

The design grew out of a failed experiment. After the OpenClaw "lobster" framework became popular, the Robotics X team wrapped robot capabilities into Skills for such a general agent framework to schedule — with "very poor" results, clearly worse than the guide system they built for Unitree last year. "The body OpenClaw controls is like a browser, not a robot": browsers can wait, robots cannot. The team then built TairosAgent natively for robot task scheduling and model coordination, cutting response time to 2-3 seconds.

The three models divide the work: Hy-Embodied-VLM-1.0 handles spatial and scene understanding; Hy-Embodied-RxBrain-1.0 unifies cognition, planning and imagination of future states — generating not just language steps but predictions of what the physical world should look like after each subtask; Hy-Embodied-VLA-0.5 converts high-level goals into continuous, correctable actions, trained on more than 10,000 hours of human demonstration data and ranking first overall in the RoboDojo evaluation.

Production results are already emerging. Tencent says Hy-Embodied-VLA has entered a daily-chemicals factory for live testing: on a production line with high mix, small batches and frequently iterating SKUs, it achieves a success rate above 95% with a cycle time faster than 6 seconds per piece, leaving less than 3 days for data collection and post-training when a new SKU arrives. "A demo scoring 80 or 90 points, without real deployment, is almost zero," Zhang said — grabbing an object once in a lab and completing tens of thousands of tasks on a line are entirely different tests.

For scenarios, Tencent is using the hardest case — elderly care — to push the technology. The "Xiao Liu" robot performs massage with two arms; the team records professional techniques with its own vision, force and tactile collection system and reproduces them with reinforcement learning. So far it has learned only four two-arm massage techniques, and the next step is iterating in real nursing homes. Zhang says Tencent does not plan to sell elderly-care robot bodies, because physical contact with humans concentrates the hardest problems in embodied intelligence: "Entering unstructured environments like homes, safety must be guaranteed 100%."

Compute demand is rising fast: Chen Yudong, Tencent Cloud's director of heterogeneous computing R&D, says customers that once trained models on hundreds of GPUs are now asking for thousand- and ten-thousand-card clusters, with related demand growing roughly 200-300% and some models iterating a new version every two or three days. What to watch next: whether Tencent can replicate the layered architecture across more robot bodies and scenarios, and whether its cross-embodiment adaptation — UMI data transferring across arms via unified action representation — holds up.