Summary of Key Points
Tencent has not been late to enter the AI arena; rather, it has been making long-term investments in the field of embodied intelligence—where machines can perceive, make decisions, and act like humans. Its approach encompasses a complete ecosystem ranging from large-scale foundational models to intermediate intelligent platforms, all the way to application-layer products. The core innovation lies in its hierarchical and heterogeneous embodied intelligence architecture, which differs from the industry's prevailing trend of using a single model for everything. Through the Tairos platform, Tencent empowers hardware manufacturers with a “universal brain” that has already been successfully implemented in industrial settings (with an operation success rate of over 95% in chemical plants). This solution addresses key industry challenges related to data and computing power. In the future, Tencent will prioritize scaling up applications in industrial contexts, while addressing safety concerns before entering the home market.
1. Why is Tencent’s Embodied Intelligence Different?
Most embodied intelligence solutions in the industry rely on a single-end-to-end model that must handle perception, decision-making, and action control simultaneously. As a result, these systems either suffer from slow responses (due to the need for more processing time) or are unable to handle complex tasks (due to trade-offs in speed). Tencent, however, has adopted a three-tier heterogeneous architecture modeled after the human brain:
- Upper Cognitive System: Uses large models to determine goals (e.g., “setting out dinnerware”). These models do not need to run continuously and can be activated on demand.
- Middle Perception and Action System: Monitors the environment and adjusts actions (e.g., comparing the current position of the dishes with the desired arrangement and adjusting hand movements).
- Lower Reflex System: Handles unexpected situations quickly, similar to reflexes (e.g., stopping immediately upon encountering an obstacle to prevent falling).
Why this approach? Zhang Zhengyou, Tencent’s Chief Scientist, explains it using the “split-brain experiment”: The human brain’s left hemisphere handles language and logic, while the right hemisphere processes spatial information. Language is merely a tool for expression and not the entire basis of cognition. For example, describing how to set out dishes in words is less effective than showing a picture; by dividing tasks among different layers, each module can focus on what it does best, resulting in both efficient thinking and quick responses.
2. Tairos: Not Just Robot Hardware, but a “Universal Brain”
Tencent’s focus is not on developing robot hardware itself but on creating a universal brain for embodied intelligence, similar to the Android operating system:
- For Hardware Manufacturers: It simplifies adaptation. For instance, it took three months for Yuzhu Robotics to adapt their products to Tairos for guided tours, but with the standardized interface, they added new features in just five days; Yuejiang Robot Dogs required only one day.
- For Application Developers: They can directly utilize Tairos’ intelligent capabilities without having to adapt to various hardware configurations.
Why this approach? Chinese robot hardware is already highly developed (with companies like Yuzhu and Yuejiang being leaders). Instead of reinventing the wheel, Tencent leverages its AI and cloud technology to lower the barriers to industry adoption—just as Android allows smartphone manufacturers to focus on hardware production without writing their own operating systems.
3. Industrial Applications Take the Lead: 95% Success Rate, New Products Adapted in Three Days
Tencent’s HyVLA-0.5 model has been tested in a chemical plant:
- Achieved an operation success rate of over 95% for repetitive tasks such as bottle cap screwing and packaging.
- New products can be integrated in just three days (compared to weeks in traditional methods).
- It scored first place in authoritative simulations, especially in tasks requiring long sequences of actions and memory (e.g., assembling multiple parts sequentially).
Industrial applications are easier to implement because the environment is structured, tasks are standardized, and data requirements are well-defined, meeting production-level standards.
4. Data and Computing Power Challenges: Tencent’s Solutions
The biggest challenges for embodied intelligence are data and computing power. Tencent’s approach includes:
- Data: A four-tier system is used—large-scale data for initial training and high-precision data for fine-tuning in specific scenarios. The company also guides data collectors to gather “high-value” information to avoid waste.
- Computing Power: AI agents assist in the migration of algorithms to domestic chips. Previously, manual migration took five days; now, with human-machine collaboration, it takes only one day. Tencent Cloud provides integrated storage and computing services, addressing the rapid growth in computing power (200%-300% annual growth).
5. Future Directions: Industrial Scaling First, Home Security as a Barrier
The industry is currently highly homogenized, with many companies focusing on similar robots and applications. Zhang Zhengyou believes this is normal in the early stages, as intense competition will accelerate technological advancement. The future roadmap includes:
- Industrial Applications First: Simple environments and standardized tasks make these applications profitable immediately.
- Home/Aging Care Applications Later: Safety is a critical factor; technologies like tactile and force sensing are still needed (e.g., robots must be able to adjust their grip gently when assisting the elderly). These advancements have not yet been fully realized.
Tencent’s goal is not to create technological gimmicks but to build the foundational infrastructure for AI—connecting models to applications, intelligence to hardware, and developers to industries, just as the internet connects people.