With only 5 million hours of data demand met out of 1 billion hours, how can the "data infrastructure" for embodied intelligence be supplemented?

Training a human-level robot "brain" requires 1 billion hours of real-world operational data, yet the current global effective supply is only around 5 million hours. As humanoid robots become a star track chased by capital, a Zhejiang company called "Embodied Training Ground" has taken a less-traveled path: instead of manufacturing robots, it positions itself as a supplier of "data fuel" for the robotics industry.

1. The 200x Gap: The "Data Famine" Choking Embodied Intelligence

Behind this choice lies a tough problem that is choking the entire industry.

Large language models have no shortage of text data — massive amounts of text are already available on the internet for crawling and training. But embodied intelligence — that is, enabling AI to learn to operate and interact in the physical world — faces a completely different dilemma: an extreme lack of operational video data from real-world scenarios.

There is a widely cited figure in the industry: training an embodied intelligence brain close to human level theoretically requires 1 billion hours of real human operational data. At present, the available global effective supply is only about 5 million hours. This 200-fold gap has made the "data famine" a more urgent bottleneck than algorithms.

Why is data collection so difficult? In the past, obtaining robot training data mainly relied on two methods: first, manual operations in laboratories using expensive robotic arms, motion capture equipment, and engineers, which was high-cost and low-efficiency; second, generating data in simulation environments, but there is always a "gap" between the virtual and the real world, making robots prone to "maladaptation" once they leave the lab. A practitioner once calculated that a teleoperator can only produce 1 hour of valid data per day, and even with optimized processes, the upper limit is only 8 to 10 hours.

2. The Crowdsourcing Model: Making Everyone a "Teacher" for Robots

Embodied Training Ground wanted to try a different approach.

They have built a data crowdsourcing platform for ordinary people — which is likely the first embodied data platform in China open to consumer users that has successfully operated the full "task issuance — collection — review — settlement" process. The core logic is simple: in your daily life or work scenarios, you put on lightweight collection devices, complete required operational actions, upload videos, and get paid once the content passes review.

In other words, every ordinary person can become a "remote coach" for robots.

The platform currently adopts an EGO-EXO 4D full-domain collection technical solution: the collector wears a first-person camera to record operation details, and three additional fixed cameras are set up around to synchronously capture environmental information from different angles. After the four perspectives are precisely aligned on the timeline, a complete record with 3D spatial information and action sequences is formed. Compared with single-lens collection, this method allows robots to not only "see" an action, but also understand the physical environment and interaction logic in which the action takes place.

Data provided by the team shows that this model reduces collection costs to about 0.5% of traditional methods, with a daily output of over 100,000 data entries.

3. Devices, Platform Building, Data Accumulation: A Closed-Loop Business of "Hardware + Platform + Data"

The business logic of Embodied Training Ground can be summarized as a "hardware + platform + data" closed loop.

On the hardware side, they have launched a lightweight collection device priced at around 4,000 yuan. Every unit sold means an additional steady data producer. On the platform side, it connects consumer-side collectors on one hand, and robot companies and large model teams on the enterprise side on the other, matching data supply and demand for transactions. As data volume continues to accumulate, a tradable "data marketplace" will eventually take shape.

At present, they have reached strategic cooperation with multiple industry players such as Galaxy Universal, Lingxin Qiaoshou, and Qianxun Intelligence, and are in negotiations to launch data service centers in over 300 cities across China. In terms of revenue structure, standardized dataset licensing and customized data collection services form the main profit sources. The team states that they currently have healthy cash flow and relatively high gross margins.

4. "I Never Thought the Work I Do Every Day Could Also Be a Teacher for AI"

In a manufacturing city in the Yangtze River Delta, a part-time collector uses off-hours to shoot operation videos in factories, earning an extra 2,000 to 3,000 yuan per month. He was quite surprised: "I never thought the work I do every day could also be a teacher for AI robots."

This is precisely the social force that Embodied Training Ground aims to leverage — to enable the "tacit knowledge" scattered across all walks of life and scenarios to be structurally extracted in a low-cost way, and turned into standardized data that AI can learn from.

Founder Tang Chunhua has nearly 20 years of combined experience in finance and AI. Talking about the original intention of starting the business, he put it plainly: "The entire industry is flocking to robot ontology manufacturing. We chose to work on the most difficult and fundamental underlying infrastructure, because data is an indispensable missing piece for the industry. Without sufficient real data, even the most advanced large models cannot perform tasks in the physical world."

According to the team's timeline, the key priorities for the rest of this year are to increase total data collection hours and the pass rate of submitted videos, and promote the iterative development of more lightweight collection devices.

Of course, this path also faces practical challenges. Can the crowdsourcing model consistently guarantee data quality? Can the incentive mechanism for collectors remain stable in the long run? How to properly implement data compliance and privacy protection? These are all questions that must be answered before large-scale expansion.

Nevertheless, when the entire industry is still excited about trivial feats like "robots doing somersaults", the fact that someone is willing to step back and solve the more fundamental problem of "who robots should learn from" is in itself a sign that the industry is maturing.