China’s AI industry is undergoing a strategic reorientation. The chatbot era, defined by text generation, question answering, and conversational interfaces, is giving way to what Chinese technologists are calling “physical AI” or “embodied AI”: systems that can perceive, reason, and act in the real world. According to the South China Morning Post, published on May 28, Alibaba and Tencent are leading this pivot, deploying their most capable AI models as the cognitive layer for humanoid robots and autonomous systems.
The shift is not merely technological. It reflects a strategic judgment that the commercial ceiling for conversational AI is lower than that for physical AI, that the most valuable applications of intelligence are not in answering questions but in performing tasks in the physical world.
Alibaba’s Robot Control Toolkit
Alibaba’s contribution to the embodied AI pivot is a suite of supporting models released alongside its Qwen3.7-Max flagship. The package includes a robotic gripper agent, a navigation model, and a vision-language system designed for physical-world interaction. Together, these tools allow Qwen3.7-Max to function as a digital brain that can orchestrate physical actions: navigating a space, avoiding obstacles, planning a sequence of tasks, and triggering external hardware.
The vision-language system is particularly significant. Giving a robot the ability to interpret its visual environment and reason about it in natural language is one of the core technical challenges of embodied AI. It requires the model to bridge the gap between the symbolic world of language and the continuous, noisy world of physical sensors. Alibaba’s release of these tools as an open suite, rather than keeping them proprietary, signals a strategy of building an ecosystem around Qwen rather than a closed product.
Tencent’s OpenClaw Powers the First Mass-Produced Humanoid Agent
On Tencent’s side, the AI pivot is playing out through OpenClaw, the company’s AI agent framework. Embodied AI startup Zeroth announced earlier this month that its M1 humanoid robot had become the first mass-produced robot to integrate OpenClaw. The framework allows large language models to interpret human speech and translate it directly into robotic movements, closing the loop between natural-language instruction and physical action.
The Zeroth M1 integration is commercially meaningful because it is not a research prototype. Mass-produced robots that can receive natural-language commands and execute them physically represent a qualitative shift in what humanoid robots can do in real-world deployments. As EastFrontier has reported, China’s humanoid robot price war is already compressing margins as manufacturers compete aggressively on cost. The addition of genuine language-to-action capability could redefine the value proposition of these robots and arrest the commoditization pressure.
The Data Bottleneck
Wu Bangyi, chief data officer at Tianyu Shuke, articulated the core challenge in a quote published by Securities Daily: “Physical AI is about how AI understands the real world and completes tasks in real-world environments. The core of physical AI is enabling AI to evolve from cognitive intelligence to action intelligence.” The distinction between cognitive intelligence, knowing things, and action intelligence, doing things, captures precisely what makes embodied AI technically harder than conversational AI.
A Goldman Sachs report published on May 26 identified “high-quality real-world data” as one of the primary bottlenecks in embodied AI development. The scale of the problem was quantified by Yao Maoqing, co-founder of AgiBot: GPT-5 was trained on data equivalent to approximately 10 billion hours of human experience, while the entire global robotics industry has access to only around 500,000 hours of high-quality embodied AI data. Closing that 20,000-fold gap is the defining challenge of the next phase of AI development.
Chinese companies are addressing it aggressively. X Square Robot announced last week that it had partnered with home-services platform 58 Daojia to launch robot-assisted household cleaning in Beijing and Shenzhen, a deployment explicitly designed to harvest real-world interaction data. Nearly 30 training facilities and data centers focused on embodied AI have been established or are planned across China, according to the Embodied AI Development Report 2025, jointly released by the China Academy of Information and Communications Technology and Tsinghua University’s Department of Electronic Engineering.
The data gap is China’s most urgent challenge in the embodied AI race, but it is also one where China’s scale advantages are most pronounced. With hundreds of millions of consumers, thousands of factories, and a government actively supporting robot deployment in public infrastructure, China has the potential to generate real-world embodied AI training data at a pace that no other country can match. The question is whether the data collection infrastructure, the training facilities, the partnerships with service platforms, the standardized data formats, can be built quickly enough to feed the next generation of physical AI models before the technology window closes.