Suppliers of "Wolverine" and "Game of Thrones" Spin off Robotics Data Business, with Top Talents from Tencent and ByteDance Joining | Exclusive Report from Intelligence Emergence
Text | Qiu Xiaofen
Editor | Su Jianxun
At the end of 2023, Dai Ruoli, the co-founder of "Noitom", the world's largest motion capture equipment provider in terms of global shipments, once thought he was deceived.
He received two mysterious calls. The callers claimed to be from a US robotics unicorn, Company A, and an automotive giant, Company B (names have been changed at the request of the interviewee), saying that they needed to purchase 110 sets of motion capture equipment.
"Is this a ploy to get the lowest price?" Dai Ruoli thought to himself. With his more than a decade of experience in the industry, he had never seen an order demand for over a hundred units.
The motion capture industry is characterized by the fact that although the customers are all large enterprises, only a few pieces of equipment are needed to produce a major production. Since its establishment in 2012, "Noitom" has captured 70% of the global professional motion capture market share.
The motion capture equipment behind Hollywood blockbusters such as "Game of Thrones", "X-Men Origins: Wolverine", "Star Trek", and "Star Wars" all comes from this company.
However, in the era of embodied intelligent robots, "Noitom" has encountered an unexpected opportunity - data shortage is one of the biggest challenges for robots at present. Motion capture equipment is an effective tool for data collection.
For the aforementioned two US companies, the multi-modal and precise data collected by motion capture equipment is the "fuel" for training robots, which directly determines the intelligence level of the robots.
Specifically, motion capture technicians attach inertial sensors or optical markers to several key skeletal positions, including the head, hands, legs, and waist, and perform various delicate movements.
△The PNDbotics Adam-U robot collects data through the Noitom Robotics teleoperation system
The collected real - human data is mapped into various sensing and motion control information of the robot body through processes such as teleoperation, and then "fed" to the robot. From then on, it has become a reality for humans to "teach" robots to work "hand in hand".
△The data collection process of Noitom Robotics' humanoid robot teleoperation
A set of comparative data shows the current shortage of robot data: Google DeepMind once collaborated with 21 top research institutions including Stanford University to integrate 60 databases to create the world's largest robot dataset, Open X - Embodiment, which contains 1 million real data entries - six orders of magnitude lower than GPT - 5 (hundreds of trillions).
After seeing the industry trend clearly, Dai Ruoli decided to "go all in on robots".
At the beginning of 2025, "Noitom" split its established robot team into a new company with independent operation and financing. The new company is also named "Noitom Robotics", and its main business focuses on robot data.
The core team of "Noitom Robotics" was formed several months ago. Dai Ruoli, who graduated from the University of Science and Technology of China and the Chinese University of Hong Kong, has considerable influence in the industry. People in the circle call him "Senior Brother Dai" and "Dr. Dai". "Intelligent Emergence" has learned that the current core members of the team include:
Han Lei (T13), the former head of the Embodied Intelligence Center at Tencent Robotics X Lab, serves as the company's chief scientist;
Xu Anmin, the former hardware head of ByteDance's Xinshi Laboratory (Level 4 - 1) and the hardware head of Luo Yonghao's "Thin Red Line", serves as the company's hardware engineering head;
And former senior executives from large companies such as SenseTime, Baidu, and Alibaba serve as the heads of product, development, sales, BD and other businesses.
△Dai Ruoli
"Intelligent Emergence" has exclusively learned that "Noitom Robotics" has completed a tens of millions of yuan angel - round financing led by well - known institutions such as Alpha Square Group and Matrix Partners China. Recently, their Pre - A round of financing is in progress.
Selling motion capture equipment to robot manufacturers, with tens of millions in signed orders this year
When embodied intelligent robots became extremely popular and 99.99% of the companies in the industry were struggling to achieve commercialization, "Noitom Robotics" took the lead in achieving commercialization by discovering the value of data collection.
In terms of revenue, "Noitom Robotics" has signed orders worth tens of millions in the first eight months of 2025, a five - fold increase compared to the whole of last year. "Almost all domestic robot manufacturers have purchased our equipment for data collection."
Dai Ruoli introduced their business model to "Intelligent Emergence" with two letters: Project T (Tele - operation) and Project D (Data factory).
In the Project T route, Dai Ruoli designed two models: data collection and deployment support.
Data collection, as the name suggests, refers to the sale of motion capture equipment for tele - operation data collection. Dai Ruoli told "Intelligent Emergence" that after the orders from the aforementioned two leading US robot manufacturers were finalized, in the past two years, they have also received orders from nearly 30 domestic robot manufacturers, including DAMO Academy, Zhipu AI, Qianxun, ByteDance, XPeng, and Tencent.
In addition, there are also robot research institutes at many universities, including those with Li Feifei involved, and robot co - construction bases led by local governments, etc.....
△Some customers who have cooperated with Noitom Robotics in robot tele - operation and data research
For deployment support, "Noitom Robotics" plans to lay the groundwork for the future implementation of robots.
Dai Ruoli believes that, similar to the self - driving industry, when humanoid robots truly enter factories and households, a certain number of safety seats will be required. Safety seats form the safety baseline for the implementation of robot applications, used to take over the robot in case of emergencies and handle complex scenarios, ensuring human safety and the operation of the robot.
In the future, "Noitom Robotics" also plans to provide a full - set of services for robot implementation, including remote operation terminal equipment in safety seats, corresponding management and control software, and associated BPO human - resource services (i.e., business process outsourcing), etc.
In summary, the business model of Project T is to sell hardware and services.
But it doesn't stop there.
"Noitom Robotics" is also preparing to build a data collection factory (Project D), hoping to engage in the business of selling data and licenses. In comparison, data has a greater marginal effect than hardware. Meta's acquisition of nearly half of the equity of Scale AI (a data service company) for $15 billion is sufficient proof of the value of data.
Dai Ruoli revealed to "Intelligent Emergence" that they are currently selecting locations for the factory in Shenzhen and other places. Dai Ruoli predicts that in the future, the normal annual revenue of Project T may reach the level of hundreds of millions, while Project D is expected to reach the level of billions.
△The schematic diagram of Noitom Robotics' data factory
However, to become the Scale AI in the robotics field, the core question that "Noitom Robotics" cannot avoid is: What kind of data do robots need to collect?
The disputes over robot data have not yet subsided
"Data" is currently a major pain point in the field of embodied intelligent robots. The industry's differences can be roughly divided into two camps:
Most people draw lessons from the development path of large language models and believe that the scaling law also applies to embodied intelligence. The reason is that autonomous driving, as a subset of embodied intelligence, has already proven the magic of data.
However, people represented by Wang Xingxing call out that the industry should not blindly pile up a large amount of data when the models are not good enough or unified.
An industry insider told "Intelligent Emergence" that autonomous driving relies on millions of cars to collect data, and after collecting data for several years, it is still only at the L2++++ level. "If the robotics field imitates autonomous driving and collects a large amount of real - machine data, how much time and money will it cost? It's just too stupid!"
Another industry insider summarized to "Intelligent Emergence" that the biggest problems with real - machine data are poor quality, low efficiency, and high cost.
The aforementioned insider revealed that a large domestic robot manufacturer needed two people to tele - operate for a whole day, with each person collecting data for eight hours, to produce one hour of qualified real - machine data in the first half of last year.
Taking Tesla as an example, the hourly wage of US tele - operators is as high as $50, and they also have to work in motion capture suits and VR equipment for more than 7 hours a day.
To solve the various constraints of real - machine data, many manufacturers in the industry choose to expand the data scope. For example, they directly "feed" the robot with Internet video data, use AI to synthesize a series of simulation data, or directly use motion capture equipment to collect real - human data, etc.
Before the industry differences are resolved, Dai Ruoli tries to use the "pyramid structure" to solve the current dilemma in the robotics field.
He divides the data required by robots into four layers:
The top of the pyramid is still real - machine tele - operation data. He believes that real - machine data is the highest - quality data, and its role is like the 100,000 kilometers a car runs after leaving the factory - used to collect the "intrinsic characteristics" of the car.
However, in addition to the aforementioned shortcomings, the biggest problem with real - machine data is that it cannot cross different robot bodies. Just as it is difficult to smoothly apply the data of Unitree robots to Zhipu AI's robots, there is a significant loss in between.
Therefore, he also opposes blindly pursuing generalization at the level of real - machine data. "Just like a car only needs 100,000 kilometers of verification after leaving the factory. It's not necessary to pursue a quantity level as high as one billion kilometers. It's very strange to demand a large - scale tele - operation requirement."
Dai Ruoli instead believes that real - human data is the key to the generalization of robots. In the second and third layers of the pyramid, he believes that it should be high - precision and low - precision real - human interaction data. The difference between real - machine data and real - human data lies in who generates the data. The former comes from robots, while the latter comes from humans.
△The pyramid of embodied intelligent data. Source: Noitom Robotics
The biggest feature of real - human data is its high - dimensionality, multi - modality, and high - precision. In his opinion, mapping high - dimensional human data to low - dimensional robots is more effective and can better solve the problem of cross - body application.
"Because you don't know what the precision of the robot's sensors will be in the future. Mapping from low - dimension to high - dimension is not feasible and may lead to insufficient data volume," Dai Ruoli told "Intelligent Emergence".
However, human behavior is boundless and complex. How to collect it?
Dai Ruoli tries to "atomize" human movements. He abstracts human movements into: grasping, holding, throwing, pinching, pushing, rolling, rubbing, twisting, etc.
In the future, "Noitom Robotics" plans to design some long - term tasks in their data collection factory to cover real - human data collection tasks. Dai Ruoli said that they are creating a dataset of up to hundreds of thousands of hours and plan to partially open - source it to the industry.
The third layer of the pyramid is low - precision human - robot interaction data. Dai Ruoli's view on this is also quite against the consensus.
He believes that the best device for collecting this part of the data is not the robot, but the currently popular AI glasses. In the future, as cameras gradually become a standard feature of AI glasses, the data collected by the glasses will have diverse scenarios, generalized targets, a large quantity, and extremely low cost.
△The schematic diagram of virtual tele - operation for cross - body mapping
Of course, Dai Ruoli does not oppose the use of Internet information in the robotics field. At the bottom of the pyramid, he believes that the vast amount of Internet - synthesized data and simulation data are also indispensable, which will greatly expand the data boundary. In the future, each layer of the robot data pyramid will be "two orders of magnitude larger" than the layer above it.
As robot data becomes the decisive factor for future robot manufacturers, Dai Ruoli predicts that in the future, nearly half of the funds in China may be used to solve the problem of robot data. In the first eight months of 2025, the financing amount in the robotics field reached 38.6 billion yuan, twice that of the whole of 2024.
Although there are many non - consensus views in the robotics field, it is certain that the bottleneck