At the 2026 World Artificial Intelligence Conference (WAIC) held in Shanghai from July 17 to 20, TENCENT made significant announcements in the field of embodied intelligence.
The company's Robotics X Lab, Futian Lab, and Hunyuan model team jointly released a series of new embodied AI base models and intelligent agents. This marks the first systematic effort to close the loop encompassing "perception-body-action."
Key Model Releases
Three models were released and open-sourced. The first is the next-generation embodied Vision-Language base model, Hy-Embodied-VLM-1.0. It addresses capabilities such as understanding object function, perceiving environmental changes, and spatial reasoning. In evaluations across 37 tasks, it significantly outperformed models of comparable scale. Notably, with an A3B parameter size and roughly one-tenth the computational cost, it achieves performance close to the previous flagship A32B model, making it lighter and more suitable for direct deployment on robots.
The second is the natively embodied world cognition model, Hy-Embodied-RxBrain-1.0, which unifies reasoning and imagination capabilities.
The third is the vision-language-action model, Hy-Embodied-VLA-0.5. It builds a complete learning stack that synergizes data, models, training, and deployment. It leverages an ultra-high-precision UMI data collection system to accumulate over 10,000 hours of human demonstration data.
Industrial Applications and Ecosystem Integration
These models have already begun deployment in real industrial scenarios, completing validation in tasks such as shopping guidance, elderly care services, and more. The development also integrates capabilities from various TENCENT divisions, including Tencent Cloud HAI, multi-network aggregation acceleration, TRRO, as well as Tencent Music Entertainment's (TME) vast music library and IoT ecosystem connectivity.
New Embodied Intelligent Agents
In the embodied agent domain, TENCENT introduced Apexio, a continuously online embodied agent. It features a three-layer architecture operating at different frequencies: a top cognitive layer for on-demand deep thinking, a middle perception-action layer continuously processing multimodal information at about 15Hz, and a bottom execution layer running at a higher frequency for instantaneous reflex-like handling of collisions and imbalances.
The company also released TairosAgent, an natively embodied agent framework. Its three systems—perception-reaction, cognitive-decision, and exploration-consolidation—are all designed from the ground up for embodied tasks and continuously maintain three types of memory: environmental, self-state, and task memory.
Platform Upgrade and Industry Collaboration
Based on the Hy-Embodied model series and the native agent framework, TENCENT has comprehensively upgraded the Tairos Embodied AI Open Platform (released in 2025). The upgraded Tairos will remain open-source, continuing to provide a suite of capabilities and services including models, agents, and development tools to lower the development barrier across the entire chain from model to embodiment to application.
Over the past year, TENCENT's Robotics X Lab has collaborated with numerous robotics companies and industry partners to explore the implementation of embodied intelligence across different robots and industrial scenarios.
Next-Generation Demonstration Robot
Building upon the previously developed "Xiao Wu" robot, TENCENT has created a new demonstrator robot named "Xiao Liu." It is designed for safe human interaction and features human-like motion. Utilizing a tendon-driven mechanism, it possesses a unique "meridian massage technique." High-quality technique data is obtained through a proprietary data collection system fusing force, touch, and vision sensing, and is trained via reinforcement learning to accurately replicate massage trajectories and pressure.
Comments