The Tsinghua team featured on the cover of Nature has launched its startup, Miji Intelligence has successively completed three rounds of financing totaling hundreds of millions of yuan | Xinglian Portfolio Update

 DenseAI focuses on more fundamental underlying issues of world models: What is the ultimate source of physical intelligence?

DenseAI, a Physical AI foundation model company, has completed its third round of financing, with total accumulated financing amounting to hundreds of millions of RMB within half a year. Its investors include leading state-owned capital from Beijing and Shenzhen, well-known USD capital, and top AI ecosystem investment funds.

Founded in 2026, DenseAI is a top-tier AI team originating from the Department of Automation, Tsinghua University. It is committed to building the next-generation physical world foundation model, taking world models and Dense Learning as the core, to promote the evolution of artificial intelligence from understanding the physical world to autonomous learning and continuous iteration.

At the moment when world models have become a global AI competition hotspot, DenseAI has chosen a differentiated technical route: taking large-scale stock real physical data with extremely low marginal cost as the source of intelligence, using original Dense Learning technology to break through the Curse of Rarity and long-tail challenges of data, to realize high-efficiency pre-training and high-level post-training of physical intelligence, and finally enable agents to have the ability of continuous evolution in the physical world.

01

Top Tsinghua Scientific Research Team, Original Technologies of Nature Cover Level

DenseAI has gathered a founding team that combines top international scientific research capabilities, large model R&D strength and industrialization experience.

Feng Shuo, co-founder and Chief Scientist of the company, is an associate professor and doctoral supervisor at the Department of Automation, Tsinghua University. He has profound original scientific research accumulation in the fields of Dense Learning and physical intelligence, has been selected into the MIT TR35 China list, won the Damo Academy Green Orange Award, and was selected for the Class A Youth Science Fund Project of the National Natural Science Foundation of China (formerly the National Science Fund for Distinguished Young Scholars) in 2026.

Another co-founder and CEO, Shi Mengkai, holds bachelor's and doctoral degrees from the Department of Automation, Tsinghua University. He once co-founded a startup in the autonomous driving field and served as CTO, with rich experience in hard technology entrepreneurship and industrialization. The company has gathered core R&D talents from national laboratories, ByteDance and leading embodied intelligence companies, and attracted a group of outstanding young AI talents from Tsinghua University, forming complete R&D capabilities covering original theories, foundation models and embodied intelligence systems.

In 2023, DenseAI's core team published the research results of dense reinforcement learning in the world's top comprehensive scientific journal *Nature*, which was selected as the cover paper of the current issue, standing shoulder to shoulder with Google DeepMind in the academic field of artificial intelligence. This research was also specially reported by official columns such as Nature News & Views, Nature Podcast and Nature Videos, reflecting its international academic influence.

In a series of studies, the company team proposed and theoretically demonstrated the "Curse of Rarity" in intelligent learning: as the model is continuously trained, the learning value of data may decay exponentially, and high-value data is often sparsely distributed in the massive data pool. A large amount of low-value data is not only difficult to continuously improve the model's ability, but may also submerge the truly valuable learning signals, seriously restricting the efficiency and effect of model training.

To solve this fundamental problem, the team creatively proposed the Dense Learning theory and method, which automatically and intelligently identifies and screens high-value data by using the model's own internal learning signals, and combines corresponding model architecture innovation to improve the information density and learning value of data, so as to realize more efficient intelligent learning.

02

Rethinking the World Model Scaling Law: Prioritize Large-scale Data First

When more and more companies lay out world models around video generation, large model parameters and robot ontology, DenseAI puts forward a more fundamental question: What is the ultimate source of physical intelligence?

From DenseAI's perspective, large-scale data is the source of intelligence. The key bottleneck restricting the large-scale development of physical AI is not only model architecture and computing power, but also scalable real physical data. Compressing the massive, already recorded, most authentic data in the physical world into intelligence efficiently through Dense Learning is how physical intelligence is born.

The company believes that high-value data sources for physical world models need to meet three conditions: sufficiently low marginal cost, sufficient real physical information, and the ability to expand on a large scale from day one. Focusing on this judgment, DenseAI focuses on exploring world-centric real interactive video data, using physical activities naturally generated by humans in real scenarios such as industrial production and daily life to build a low-cost, sustainably scalable pre-training data system, reducing dependence on specific robot bodies and expensive action labeling.

The company hopes to form a continuous growth flywheel of "scenarios generate data, data trains models, and models in turn empower scenarios" through real scenarios, to establish a truly scalable training foundation for physical world models. It is reported that the company's world model has provided underlying technical support in a global robot competition, and has begun commercial delivery in real scenarios, completing initial product market verification.

03

 Dense Learning Drives Autonomous Evolution, Enabling AI to Learn to Create Challenges on Its Own

If scalable data solves how world models acquire extensive physical knowledge, what Dense Learning needs to solve is how agents continuously acquire new capabilities from limited experience.

DenseAI believes that a large number of interactions in the real physical world are highly repetitive, and the key to determining generalization ability is often hidden in unknown environments, complex interactions and relatively long-tail scenarios. Based on the Dense Learning theory accumulated by the team for a long time, the company further explores the Self-Evolution training system, enabling agents to actively discover their own capability boundaries, dynamically create more challenging tasks, and continuously improve their capabilities through real interaction and feedback.

Enable AI not only to solve problems, but also to actively create new problems to train itself. This technical route has achieved phased experimental progress. According to the company's disclosure, in multiple real machine tasks, the failure rate has decreased by orders of magnitude using the dense reinforcement learning method, which verifies the potential of high-value experience selection in improving robot training efficiency, and also provides practical basis for autonomous evolution driven by Dense Learning.

04

Build the Next-Generation Physical AI Foundation Model

DenseAI believes that the next-generation general physical intelligence should not be limited to a certain type of robot or specific task, but should be built on a world foundation model that can transfer across scenarios, learn continuously and evolve autonomously.

In the future, the company will continue to promote the integrated innovation of world models, Dense Learning and autonomous evolution technologies, and build a technical system in which real physical data, foundation model pre-training and agent autonomous evolution promote each other.

World models enable AI to understand the world, and Dense Learning enables AI to evolve efficiently.

DenseAI is exploring the realization of general physical intelligence along this core technical path.