Highlights

Humanoid machines can walk, dance, and box, but reliably doing practical tasks in messy real-world settings remains difficult. ACE Robotics chairman Wang Xiaogang forecasts that embodied AI could achieve its own “ChatGPT moment” by the end of 2027, if advances in world models combine with far larger datasets captured from real environments. The company aims to gather tens of millions of hours of operational data within two years and to deploy its systems across many retail locations, with plans to pursue an IPO. A major emphasis is on overcoming the current shortage of real-world training data, which the company sees as essential to moving robots from demonstrations to dependable commercial use.

Sentiment Analysis

• The overall tone is cautiously optimistic, reflecting strong confidence in technical progress tempered by clear recognition of remaining challenges. The narrative highlights recent funding, industry backing, and concrete deployment plans, which contribute to a positive outlook. At the same time, limitations—especially the scarcity of high-quality real-world training hours—introduce a realistic, measured perspective on timelines and feasibility. This balance produces a moderately positive sentiment with noticeable caveats.

Article Text

Humanoid robots have demonstrated impressive physical capabilities—walking, dancing, and even mock-combat—but turning those demonstrations into dependable, useful behavior in unpredictable real-world environments remains a major technical hurdle. Wang Xiaogang, chairman of ACE Robotics, argues that a convergence of improved AI models known as world models and a significant expansion of real-world training data could push embodied intelligence to a breakthrough moment similar to ChatGPT for language models. Wang expects that turning point could arrive by the end of 2027 if current development trajectories continue and data collection accelerates.

ACE Robotics, founded in July 2025 and backed by investors including Ant Group and SenseTime, has rapidly attracted attention and capital. The company raised over $100 million in the first half of 2026 and has public ambitions to pursue an initial public offering when regulations allow. In parallel with fundraising and corporate milestones, ACE is prioritizing the collection of extensive real-world data. Wang notes that the industry to date has amassed roughly 100,000 hours of relevant robotic data—an amount he considers insufficient for training robust embodied foundation models. To address that gap, ACE plans to gather tens of millions of hours of environmental and interaction data over the next two years, a scale-up the company views as necessary for the next phase of model training.

Embodied AI refers to systems that perceive their surroundings, model the state of the world, reason about possible actions, and then execute those actions through sensors and actuators. World models augment this capability by predicting how objects and environments respond to actions: for example, estimating what will happen when a robot reaches to grasp a cup or steps over an obstacle. These predictive abilities are essential for safe and effective operation in varied, unstructured settings. Without large-scale, diverse, and realistic training data, such predictive models risk being brittle when confronted with novel situations.

Other organizations are pursuing complementary strategies. Researchers have developed devices like wearable exoskeletons to capture human motion for robot training, and major robotics and AI companies are building specialized model suites for navigation, manipulation, and environment simulation. Boston Dynamics has progressed toward a production-ready humanoid, citing AI improvements as a key enabler, while firms such as Alibaba have introduced robot-focused model collections aimed at real-world tasks.

Commercial deployment remains a central objective. ACE plans to roll out its technology across a network of stores in the coming year, using on-site operations both to provide practical services and to harvest further training data. This dual deployment-and-data strategy aims to accelerate learning loops: robots operating in real environments generate data that improves models, which in turn enable more capable and reliable robotic behavior. If successful, this iterative approach could shorten the path from laboratory prototypes to fielded products.

Nevertheless, significant obstacles persist. Collecting and curating tens of millions of hours of meaningful, labeled interaction data is a logistical and technical challenge. Ensuring safety, generalization across diverse environments, and regulatory compliance are additional hurdles that will influence how quickly commercial-grade embodied intelligence becomes widespread. The timeframe suggested by ACE—reaching a transformative ‘‘ChatGPT moment’’ by late 2027—depends on coordinated advances in modeling, data acquisition, and real-world testing.

In summary, ACE Robotics projects a near-term inflection point for embodied AI driven by larger world models and massive real-world datasets, backed by strong funding and practical deployment plans. While the outlook is optimistic, the path forward requires overcoming the persistent scarcity of high-quality training data and ensuring systems generalize reliably beyond controlled demonstrations.

Key Insights Table