Yingke Exclusive | Tang Wenbin's "Force Brain" and Shopping Flow Robot Company, with investments from Zhipu, SenseTime, Jiayue, and others

Null

Author | Qiu Xiaofen

Editor | Yuan Silai

Exclusive from Hard Tech: Embodied intelligence company "Yuanli Lingji" recently completed a new round of financing, mainly from several large model companies, including Zhipu, Jieyue Xingchen, and SenseTime. Additionally, industry investors such as Huqian and SAIC Hengxu continue to increase their investments.

"Yuanli Lingji" is a general embodied large model company, founded in March 2025 by Tang Wenbin, co-founder and CTO of Megvii Technology, with the core founding team being the original team from Megvii.

Interestingly, this round of financing also marks a rare "meeting" between SenseTime and Megvii, former competitors, amid the wave of embodied intelligence;

Moreover, including Alibaba, the sole lead investor in Series A+, this is also a rare gathering of four major domestic large model companies in the embodied intelligence track—previously, Zhipu only made limited investments in embodied intelligence through Z Fund, and Jieyue Xingchen almost did not participate in embodied intelligence.

This collective action also signals that as the main battlefield of large model competition shifts from tokens to actions, embodied models with physical interaction capabilities have become the next high ground for model companies to target.

Along with this financing, Tang Wenbin is also consolidating robot assets.

Exclusive from Hard Tech: "Yuanli Lingji" recently merged with "Atomix" (formerly Lingji Juhua) logistics robots through equity acquisition, aiming for large-scale deployment and global expansion of embodied intelligence.

The business gene of "Atomix" can be traced back to 2016—at that time, Tang Wenbin led the smart logistics and robot dispatching business (Hetu system) within Megvii, promoting multi-form logistics robot solutions.

Until July 2024, as Megvii's business evolved, Tang Wenbin spun off the logistics robot business from Megvii and established "Atomix."

After several years of exploration, "Atomix" has become the second-largest global seller of four-way pallet trucks, with over 500 service projects, clients including Uniqlo, Mixue Bingcheng, CATL, and annual revenue nearing one billion yuan.

As the embodied intelligence hardware supply chain approaches maturity, the industry is nearing a critical barrier: the embodied brain. Compared to the clear evolution path of language models, current embodied intelligence models lack low-cost, large-scale, high-quality data, let alone convergent training paradigms. It can be said that the entire industry is exploring in chaos.

In this context, the integration of the entity, the brain, and data may become the new norm in the embodied track.

The industry's ideal state has always been to build a true data flywheel. But in reality, the industry is stuck in a "data deadlock"—models need error data from real scenes to evolve, but without well-trained models, robots cannot operate in scenes, and thus cannot collect real data.

Therefore, some sources say that the merger of these two companies essentially opens the data deadlock through closed-loop models and scenarios.

As Tang Wenbin previously mentioned in an interview, Picking (grasping) is the "atomic task" of the embodied intelligence era—just as coding is to large models, "Atomix" is like a continuously running Picking data engine.

"Yuanli Lingji" robot making breakfast (Image source / company)

It is understood that in the future, the real data generated by "Atomix" in over 20 countries and more than 500 projects will directly fuel the training of "Yuanli Lingji" models; meanwhile, the embodied intelligence models trained by "Yuanli Lingji" will quickly achieve collaborative operation with existing "Atomix" robots.

This concept may not be just a castle in the air but built on a certain technical foundation. Prior to this, "Yuanli Lingji" had launched a general embodied large model "DM0."

In an earlier interview, Tang Wenbin mentioned that on the data level, "Yuanli Lingji" has completed the industry's first "three-category data fusion"—mixing internet semantic data, physical rules of intelligent driving, and robot operational data for training, enhancing data scale and quality.

This cross-domain hybrid training method allows "DM0" to break free from dependence on specific hardware parameters, like an experienced "driver," abstracting general physical laws from massive heterogeneous data, and capable of ignoring hardware differences to transfer across various robot configurations, achieving universal operation logic.

"Yuanli Lingji" robot making cocktails (Image source / company)

More importantly, "Yuanli Lingji" is also attempting to extend the "thinking chain reasoning" of large models into the physical space—this enables "DM0" to achieve sub-millimeter precision operations with only 2.4 billion small parameters, significantly improving success rates in long-range continuous tasks.

Through a series of combined efforts, "DM0" is trying to break the pain points of traditional embodied models: single data source, device swapping leading to failure, and bloated parameters.

After this merger and financing, China's embodied intelligence industry welcomes a strong new player. More importantly, it signals that the industry is entering the next stage—finding the scaling law for embodied models.

This is not something that can be overcome merely by stacking robot bodies.

This week, media reports that ByteDance is recruiting a head of embodied intelligence technology, targeting core technical talents from leading startups; meanwhile, overseas star embodied intelligence company Skild AI has just completed the acquisition of Zebra Technologies' automation business.

The moves of domestic and international giants are strikingly similar— as manufacturers, data assets, model developers, and scenario operators accelerate their consolidation, the industry is entering deep waters.

Homepage image source | Internet

Layout | Fan Xinya

Welcome to communicate