A rising star in the embodied intelligence sector has attained unicorn status in merely 3 months.
When investment in embodied intelligence in China's primary market is gradually tightening, an overseas company focused on robot data training has secured two rounds of financing within three months, and has already joined the ranks of unicorns valued at over 10 billion yuan.
Not long ago, robot data firm XDOF announced news of a new round of financing, with an estimated valuation of about 1.2 billion US dollars, equivalent to more than 10 billion yuan. It has only been about three months since the company announced the completion of its 70 million US dollar Series A financing in June this year.
Founded in 2024, XDOF was co-founded by Philipp Wu, Fred Shentu and Nemo Jin. One of the company's technical foundations is the GELLO teleoperation system developed in the research of robot learning at UC Berkeley.
What draws my attention is that this company neither develops models nor manufactures humanoid robots. Instead, it runs a seemingly unglamorous data business — collecting, sorting and processing real-world interaction data for robots. Yet this "data-selling" company has won the favor of star institutions such as a16z and Thrive Capital.
At present, this round of financing for XDOF is still a proposed transaction, and details such as the specific financing amount and whether the valuation includes new funds have not been publicly confirmed. However, no matter how the final transaction is implemented, this round of financing has at least sent a noteworthy signal: as the large model and robot industries enter a new stage of development, the data business that was once regarded as labor-intensive and lacking in imagination space is being re-examined by capital.
The Scale AI in the robotics sector
"Data quality is often the unsung hero that gets overlooked in deep learning." This is a line written by Philipp Wu when he shared a research article from his team on Twitter.
Philipp Wu is the founder of XDOF. UC Berkeley, where he works, is itself a leading global academic center for research on robot learning, reinforcement learning and embodied intelligence. The widely talked-about "Four Berkeley Returnees" in China's embodied intelligence circle also reflects the influence of this university among Chinese entrepreneurs and investors from the side.
Philipp Wu, founder of XDOF, started his research with robotics itself. According to his public LinkedIn profile, he once conducted undergraduate research at the Robot Learning Lab of the University of California, Berkeley, under the supervision of Professor Pieter Abbeel, focusing on robot control. From July 2022 to August 2024, he also served as a visiting researcher at Meta, working on multimodal robot foundation models.
Pieter Abbeel is a professor at the University of California, Berkeley, director of the Robot Learning Lab, and co-director of the Berkeley Artificial Intelligence Research (BAIR) Lab. He has long been studying how robots acquire skills through human demonstration, reinforcement learning and trial and error. For example, the complex operations that we often see robots perform today, such as folding clothes, tying knots and assembling parts, are all part of Abbeel's research areas. He also co-founded Covariant with his students to bring robot learning technology into the logistics and warehousing scenarios. In 2022, ACM Prize in Computing was awarded to Abbeel in recognition of his research contributions in the field of robot learning. Philipp Wu once worked as a research engineer at Covariant, studying learning technologies for logistics and warehousing robots.
If we compare embodied intelligence to a new industrial chain, the robot body is the "body", the model is the "brain", and the data is more like the "fuel" in the training process. Robot body manufacturers need data to train their models, model vendors need data to improve their capabilities, and data companies have the opportunity to provide services to multiple customers. Therefore, some investors call XDOF the Scale AI or Mercor in the robotics sector.
However, this analogy also needs to be treated with caution. Companies such as Scale AI operate in the internet data and model training market, while XDOF faces a far more complex physical world. Different from traditional industrial robots that rely on pre-written programs, general-purpose robots need to cope with ever-changing objects, environments and tasks. How a person grabs a pen, opens a bottle cap, or folds a piece of clothing involves complex processes of vision, movement and feedback. For robots to learn these tasks, they cannot simply rely on engineers writing rules for each action, but need a large amount of real interaction data.
This type of data does not exist naturally like internet text, and the data required by robots often has to be obtained through real operations. The pose, hand trajectory, contact feedback, failure causes when a robot grasps an object, as well as how humans adjust their movements, are all materials needed for training general-purpose robots. In addition, the quality requirements for robot training data are higher: an action that seems "almost right" may fail due to differences in strength, angle, contact position or timing. This is why robot data collection cannot be simply understood as "shooting videos of robots". It requires synchronous recording of vision, motion trajectories, robot status and task results, and also needs to consider motion mapping between different robots, data quality control, and how to extract effective learning signals from failures and adjustments.
In June this year, XDOF released the WARP-RM (Warp-Augmented Relative Progress Reward Model) solution. The research team found when training robots to fold T-shirts that after adding more demonstration data, the robots performed worse instead. The problem is not that these demonstrations ultimately failed to complete the task, but that they contained a large number of pauses, hesitations and repeated adjustment processes. For imitation learning models, these actions will also be treated as training signals.
What WARP-RM tries to solve is how to identify the parts that truly drive the progress of the task from these data. By changing the time scale of existing trajectories, the model learns to judge the speed of task progress: whether an action is pushing the task forward, stagnating, or even regressing. In this way, a piece of data that originally only has one demonstration can generate more relative progress signals for training by playing back at different speeds. The results show that in the T-shirt folding experiment, the policy trained with WARP-BC weighted training maintained a higher success rate on datasets of different qualities; on the cleanest dataset, both methods completed 20 folds, but the weighted strategy took an average of about 64 seconds, while ordinary imitation learning took about 114 seconds.
Star institutions such as a16z place heavy bets
As is known to all, the data problem is plaguing the embodied intelligence industry. Although more and more companies are using simulation data and synthetic data as a way to solve the problem, simulation can expand the scale of data, but real data determines the distance between these data and the real world. If a company can master the capability of real data collection, and further form a stable data pipeline, tools and customer network, its commercial value may be upgraded from a data service provider to a training infrastructure supplier.
In June this year, XDOF publicly released the ABC-130K dataset, which contains more than 130,000 robot operation trajectories, covering about 200 dual-arm operation tasks, and was completed in cooperation with institutions such as UC Berkeley. At the same time, the company announced the completion of 70 million US dollars Series A financing, with Thrive Capital, a16z, Lux Capital, Spark Capital and WndrCo all participating. By September, XDOF was reported to be in talks for a new round of financing, with an estimated valuation of 1.2 billion US dollars.
Behind the intensive attention from capital, the company's annualized revenue is close to 50 million US dollars, and its customers include a number of cutting-edge AI laboratories.
In fact, a number of star companies and founders focused on large model data training have already emerged in China, and their financing popularity is also very high this year. Recently, after communicating with some investors, we found that the upper limit of the data business does not lie in how many trajectories are sold. Just like some early data companies, they are easily regarded as outsourcing service providers: customers put forward tasks, the company organizes personnel to collect data, complete labeling, and then deliver according to the project. However, if the training demand for robot models continues to grow, this model cannot support a larger commercial space.
The reason is that data demand is not one-off. Model training requires continuous supplementation of new data. When the model effect deteriorates, recollecting is needed. When switching to a different robot or task scenario, new data is also required. A truly valuable company needs to connect collection, cleaning, labeling, training, evaluation and feedback to form a continuously operating data production system.
Taking XDOF's business model as an example, the company publicly emphasizes that its business includes not only datasets, but also data pipelines, collection tools and labeling systems. GELLO, which was developed by the founding team earlier, is also a set of data collection system for robot teleoperation. In the eyes of investors, the upper limit of this type of company is not necessarily to become a larger data outsourcer, but to become an infrastructure platform in the robot training ecosystem.
The financing of embodied intelligence in 2026 has clearly been concentrated at the two ends of the "brain" and "data". But this does not mean that all sub-tracks are heating up synchronously. On the contrary, capital is being concentrated in fewer projects, and investors are beginning to shift from "technology stories" to "commercialization capabilities".
In China, Photon Intelligence, a physical AI data and evaluation infrastructure enterprise, has successively completed multiple rounds of financing since the beginning of this year, and the financing pace has accelerated significantly. In March, it completed 10 billion yuan of Series A++ and A+++ rounds of financing, becoming the world's first embodied data unicorn; in May, it obtained another new round of financing, led by Ant Group, with a post-money valuation of over 2 billion US dollars (about 150 billion yuan); in June, it completed 10 billion yuan of strategic financing, with a total of about 20 billion yuan for the two rounds within three weeks.
The "Zhiyuan Robotics ecosystem" is also making efforts to build the infrastructure of the embodied intelligence data platform. In February this year, Mifeng Technology, a physical AI data service platform incubated internally by Zhiyuan Robotics, announced its establishment, and has completed hundreds of millions of yuan of financing in total by the end of August 2026. Not long ago, the 20,000th set of MEgo ontology-free data collection equipment was officially off the production line. It is understood that this is also the first time in the industry to achieve mass production of ontology-free data collection equipment.
"Garbage in, garbage out." The quality of data ultimately determines the upper limit of the model's capabilities.
As robots gradually step into the real world, more and more enterprises are entering the field of data collection, and the data supply of physical AI is accelerating. Although the current data accumulation is still far from truly supporting the emergence of robot intelligence, the failure of some data collection projects also means that some capital has paid the cost for technical routes that have not yet been proven feasible. But this does not mean that the value of data is disappearing. On the contrary, as the technical routes gradually converge, data will shift from the pursuit of quantity and quality to a set of data-centered capabilities for collection, processing and feedback. By then, the value of data companies should also be re-evaluated.
This article is from the WeChat official account "Dongshi 10 Tiao Capital", written by Wei Xianghui, and authorized for release by 36Kr.