China's robot makers cite data gaps and weak AI 'brains' at WAIC 2026
Synopsis
Top executives at SenseTime spin-off Ace Robotics and AgiBot publicly admitted at WAIC 2026 that China's humanoid robot industry lacks both the real-world data and the AI 'brain' needed to build viable world models — a candid admission that mass deployment timelines could slip significantly.
Key Takeaways
Wang Xiaogang , co-founder of SenseTime and chairman of Ace Robotics , said the embodied AI industry must link hardware, data, models, and real-world scenarios into a single closed-loop system.
Most robot training data currently comes from human demonstrations, but this data is ineffective unless hardware design and data-collection methods are co-optimised.
Yao Maoqing of AgiBot said multi-modal physical-world data is far less abundant than the datasets used to train large language models, creating a bottleneck for world-model development.
Real-world robot deployment remains too limited to generate the volume and variety of data needed to improve AI models at scale, according to Wang .
The remarks were made at the World Artificial Intelligence Conference (WAIC) , which concluded on Monday, 21 July 2026 , in Shanghai .
Chinese robotics companies are struggling with two compounding deficits — insufficient real-world training data and underpowered AI models — that are slowing the development of next-generation embodied intelligence systems, industry executives said at the World Artificial Intelligence Conference (WAIC), which concluded on Monday, 21 July 2026, in Shanghai.
The core bottleneck: a broken feedback loop
The most critical challenge facing the embodied AI sector is building a system that links 'hardware, data, models and real-world scenarios into a closed-loop iterative system,' said Wang Xiaogang, co-founder of SenseTime and chairman of its robotics spin-off Ace Robotics. Speaking on the sidelines of WAIC, Wang said that robot hardware design must be jointly optimised alongside data-collection methods and physical structure — a level of integration the industry has not yet achieved.
Currently, the bulk of training data is collected through human demonstrations. Wang cautioned, however, that raw demonstration data cannot meaningfully improve robot performance unless the underlying hardware and data pipelines evolve in tandem.
Why it matters: world models need physical-world data
Yao Maoqing, partner and senior vice-president at Shanghai-based humanoid robot maker AgiBot, pointed to a structural mismatch: the volume of multi-modal physical-world data available today is far below what large language models consumed during their training runs. This shortfall directly constrains the development of so-called world models — AI systems expected to allow next-generation humanoid robots to simulate, understand, and navigate their physical surroundings.
Without richer and more diverse physical-world data, world models remain theoretical ambitions rather than deployable tools, Yao indicated.
The scale problem: too few real-world deployments
Wang identified limited real-world robot deployment as a self-reinforcing obstacle. Robots operating in constrained or controlled environments generate too little varied data to drive meaningful model improvements. 'The key question is how to unlock these scenarios and replicate them at scale,' he said.
The implication is stark: without broader commercial deployment, Chinese robotics firms cannot accumulate the data needed to close the gap with the physical-world reasoning capabilities their hardware promises.
The competitive backdrop
China's humanoid robotics sector has attracted significant state and private capital in recent years, with companies including AgiBot, Ace Robotics, and others racing to commercialise bipedal robots for industrial and service applications. Yet the acknowledgement at WAIC that both the 'brain' — the underlying AI model — and the data infrastructure remain inadequate signals that the sector's timeline to mass deployment may be longer than headline investment figures suggest.
The convergence of hardware limitations, data scarcity, and immature world models means that whichever company cracks the closed-loop data flywheel first will hold a durable structural advantage. Investors and policymakers tracking China's robotics ambitions should watch deployment scale — not just funding rounds — as the leading indicator of genuine progress.
Point of View
Causing costly timeline resets. The 'world model' bottleneck is particularly significant because it is not solvable by procurement alone; it requires sustained real-world deployment at a scale Chinese regulators and enterprise buyers have not yet enabled. Mainstream coverage tends to focus on robot form factors and investment rounds; the more consequential story is whether China can build the data commons — and the closed-loop systems — that would let its robots actually learn.
NationPress
21 Jul 2026
Frequently Asked Questions
What did Chinese robotics companies say at WAIC 2026?
Executives from Ace Robotics and AgiBot said at the World Artificial Intelligence Conference in Shanghai that China's robotics industry faces two critical gaps: insufficient real-world training data and AI models that are not yet capable enough to power next-generation humanoid robots. The conference concluded on 21 July 2026.
What is a world model in the context of humanoid robots?
A world model is an AI system designed to allow robots to simulate, understand, and navigate their physical surroundings. According to AgiBot's Yao Maoqing, developing these models is being held back by a shortage of multi-modal physical-world data compared with what was available for training large language models.
Who is Wang Xiaogang and what is Ace Robotics?
Wang Xiaogang is the co-founder of SenseTime, one of China's largest AI companies, and serves as chairman of Ace Robotics, SenseTime's robotics spin-off. He said the key challenge is building a closed-loop system that integrates hardware, data, models, and real-world scenarios.
Why is real-world deployment important for robot AI training?
Real-world deployment generates the diverse operational data that AI models need to improve. Wang Xiaogang noted that robot deployment across real-world scenarios remains too limited, creating a self-reinforcing cycle where sparse data prevents model improvement, which in turn limits deployment.
How does China's robotics data problem compare to the large language model era?
The data gap in robotics is far wider than what the LLM industry faced. AgiBot's Yao Maoqing said the amount of available multi-modal physical-world data is far from adequate compared with the datasets used in large language models, making the robotics training challenge structurally harder.
