Full-scenario data services resolve the dilemmas of embodied intelligence, Lingling Intelligence is seeking angel round financing.

When large language models enable machines to "understand" the world, the core bottleneck that prevents robots from truly "stepping into" the physical world is now centered on data. Lingling Intelligence, a service provider focusing on embodied intelligent data collection and governance, has chosen to press the accelerator for financing at this moment. Huo Yanfeng, the founder, told 36Kr: "Data is the textbook for robots to enter the physical world." Lingling Intelligence is currently seeking angel round financing.

Models do not lack algorithms, but data that can be imported into training pipelines

The industry consensus on embodied intelligence is shifting: the bottleneck no longer lies in algorithms, but in data supply. Training a general embodied model requires tens of millions of hours of high-quality physical interaction data, while the currently available high-quality data worldwide is only about 500,000 hours, with a gap of over 99%. For model companies, self-built collection capacity is far from meeting the training needs; for ontology manufacturers, the cost of real machine data collection in specific scenarios is extremely high. According to estimates, the cost per hour can reach 1,200 US dollars, and the cycle is very long. Cross-scenario generalization requires the model to be exposed to a sufficiently diverse real environment, but the quality of data obtained through crowdsourcing and video distillation varies unevenly, making it difficult to directly enter the training pipeline.

The scarcity on the supply side is also prominent: high cost of real machine collection, insufficient production capacity without ontology, limited scenario coverage, relatively single factory scenarios, difficult to reproduce long-tail scenarios such as home scenarios, and the industry also lacks unified data quality and compliance standards. According to industry estimates, the embodied intelligence data market is expected to reach the order of 100 billion yuan in 2027, maintaining a growth rate of 5 to 10 times per year. Whoever can convert human behavior into robot-trainable data assets at lower cost and more stable quality will hold the pricing power on the supply side of this track.

Dual-track and three-mode system, turning data collection into a replicable industry

Lingling Intelligence's solution is large-scale supply based on the "dual-track and three-mode" system. On the non-ontology track, the company has deployed data collection factories in Henan, Jiangsu and Shandong, and built training bases with local colleges and universities to produce data on a large scale through standardized processes; at the same time, it has built a home-based flexible employment network, allowing trained and certified natural persons to complete collection from a first-person perspective in real home environments, covering long-tail scenarios that are difficult to reproduce in factories. On the real machine teleoperation track, the company cooperates with the team of Professor Pang Zhibo from the School of Advanced Manufacturing and Robotics of Peking University to complete task demonstration on real robot ontologies, synchronously record multi-modal information of actions, force perception and environment, and provide high-quality data that can be directly deployed for ontology manufacturers.

The operation of the dual-track system is supported by Lingling Intelligence's self-developed data governance platform. From preprocessing, spatial reconstruction, intelligent labeling to quality assessment, the platform ensures output quality through unified equipment, unified operation standards and full-process quality inspection. Its self-developed VIO pose algorithm improves manual processing efficiency by more than 10 times, and the acquisition end ensures multi-modal data synchronization through hardware-level global clock alignment. Relying on the backtesting and evaluation mechanism, the company runs the "integrated teaching, evaluation and testing" data textbook service throughout the whole delivery process, and plans a "return data flywheel": the failed cases during model execution are returned to the platform, targeted supplementary collection is carried out for long-tail scenarios, so that the data set can be continuously updated with model iteration.

In terms of positioning, Lingling Intelligence clearly states that it will not manufacture ontologies, not train models, and not bind to a single ontology manufacturer. As a neutral third-party data service provider, it provides full-case services for all embodied intelligence players in need of data.

Head customers have verified the business model, and the team has pressed the accelerator for capacity leap

In terms of business model, Lingling Intelligence provides overall data solutions for embodied intelligence model manufacturers and ontology manufacturers, with products covering four categories: standard data sets, customized collection services, data platforms and collection equipment. The company has served head customers such as Unitree Robotics, Ant Lingbo and AutoNavi, with the delivery scope covering data types such as commercial service scenarios, minority languages and first-person perspectives, and the cumulative delivery volume exceeds 30,000 hours. Customers such as Unitree Robotics have directly put the purchased data into the VLA training pipeline, and the two sides are jointly defining the industry data quality standards.

In terms of the team, Huo Yanfeng, founder and CEO, has worked in leading Internet companies such as Tencent and Chuangye Heima, and is an EMBA of Guanghua School of Management, Peking University. Chen Jia, co-founder and COO, has more than 10 years of continuous entrepreneurship experience, and is also an EMBA of Guanghua School of Management, Peking University. Pang Zhibo, chief scientist, is a tenured professor at the School of Advanced Manufacturing and Robotics of Peking University. Relying on the industry-university-research cooperation and alumni network of Peking University, the team has unique advantages in customer reach and data quality trust.

At the current stage, Lingling Intelligence has built a replicable large-scale production capacity with three factories and the college network. The company plans to raise 10 million yuan in this angel round, with a pre-money valuation of 50 million yuan. The funds will be invested in the R&D and upgrade of the data management platform and the customized development of collection equipment. According to the plan, the company will increase the registered scale of the collection crowdsourcing network to 300,000 people next year, and release an open source data set of 200,000 hours, further consolidating its industry position as the data infrastructure for embodied intelligence.