First Roadshow of Zhiyuan's Source Creation Program: Five AI Hard-Tech Projects Unveiled

From underlying chips and neuromorphic computing to autonomous robot deployment, physical intelligence in extreme environments, and modeling of the Earth's physical environment, frontier AI technologies are advancing industrial application along multiple distinct paths. How to bridge the gap between original achievements in the laboratory and engineering validation and commercialization has become an important topic as scientific research innovation moves toward industrial practice.

On October 26, the first roadshow of the "Source Innovation Program" (Yuanchuang Plan) released by the Beijing Academy of Artificial Intelligence was held at the Zhongguancun International Technology Trading Center. Five projects—Critical Core Technology, Lingshu Qizhi, Tensor Field Technology, Cangqiong Jizhi, and Latent Vector Technology—were presented together, focusing respectively on 3D self-selecting V-RRAM chips, neuromorphic embodied intelligence, Hamiltonian world models, reliable physical intelligence in extreme environments, and unified world models of the Earth's physical environment. The event also included a project-fund negotiation session to promote exchanges and connections between startup teams and investment institutions.

From "source innovation" to "industrial validation," what the Source Innovation Program aims to open up is not just funding

At the event, Yang Yang, vice president of the Beijing Academy of Artificial Intelligence, introduced the positioning and support mechanism of the Source Innovation Program. As an original technology startup incubation program aimed primarily at the community of BAAI scholars, the Source Innovation Program focuses on original scientific research topics in the field of artificial intelligence, follows the principles of "source innovation, rapid decision-making, small-step trials, and full-process compliant transformation," and integrates technology, talent, capital, and industrial resources to support research teams in moving from the laboratory to industry. The program focuses not only on funding connections but also on building a collaborative mechanism linking original research, startup teams, and industrial markets, promoting the transition of scientific research achievements from technological discovery to engineering validation and commercial application.

The launch of the Source Innovation Program also responds to the practical needs of AI industry development and the transformation of scientific and technological achievements. At present, AI is at a critical stage of moving from technological breakthroughs to industrial application, and the engineering validation, scenario adaptation, and commercialization of original technologies are attracting increasing attention. At the same time, policies and mechanisms for the transformation of scientific and technological achievements are continuing to improve, providing new conditions for research institutions to explore the transformation of original achievements. The Source Innovation Program hopes to seize this window of opportunity, further open up the connection between scientific research innovation and industrial needs, and support more promising original technologies in reaching the market.

1. Critical Core Technology: From original devices to software and hardware platforms

Critical Core Technology focuses on 3D V-RRAM chips, with a layout covering underlying self-selecting RRAM devices, chips, the FlagOS software ecosystem, and system applications, and is committed to building a full-stack product system with coordinated software and hardware. The project takes storage-class memory and compute-in-memory as its entry point, explores a new path for the coordinated development of new memory and AI computing power, and aims to break through the memory and computing power bottlenecks faced by existing computer architectures. The team is led by a national-level young talent from Tsinghua University and has won a provincial or ministerial first prize in natural science (2025), combining advantages in both basic research and engineering technology development.

2. Lingshu Qizhi: Reconstructing the embodied intelligence perception architecture with neuromorphic computing

Lingshu Qizhi focuses on cross-scale multimodal perception and neuromorphic embodied intelligence, and is committed to building physical perception and brain-inspired computing infrastructure for embodied intelligence. The project extends from underlying optical devices and sensing-computing integrated chips to model algorithms, exploring a full-stack technical architecture from perception to decision-making.

Its core goal is not only to enable intelligent agents to perceive the external environment, but also to understand material properties and human intentions, and on that basis achieve millisecond-level autonomous decision-making. Through the coordinated design of devices, chips, and algorithms, the project explores a new path to improve embodied intelligence's environmental understanding and real-time decision-making capabilities.

3. Tensor Field Technology: Advancing autonomous robot deployment with Hamiltonian world models

Tensor Field Technology targets high-value scenarios with high human-factor risks where engineers find it difficult to repeatedly debug, and builds a robot deployment engine driven by Hamiltonian world models, providing full-chain services such as environment modeling, policy generation, edge deployment, and continuous evolution.

The team proposes a Hamiltonian world model architecture that improves the physical consistency of model outputs through physical prior constraints. According to the project, its pretraining data efficiency is 3-4 orders of magnitude higher than mainstream solutions. The model can adapt to a variety of robot bodies, and the project also proposes a "deployment-as-a-service" model, planning to develop in coordination with robot body manufacturers, hardware suppliers, and cloud computing platforms. The team previously won third place in the Tsinghua University President's Cup and has completed an angel round of financing; the company was registered and established in Haidian, Beijing, in August 2026.

4. Cangqiong Jizhi: Reliable physical intelligence in extreme environments

Targeting extreme environments such as deep space and deep earth, Cangqiong Jizhi focuses on building physical intelligence infrastructure and attempts to solve problems such as the difficulty of obtaining data, training models, and deploying intelligent systems in extreme scenarios.

The team proposes the AI4π (AI for Physical AI) technology paradigm, using a physics-embedded simulation engine, a self-evolving hierarchical brain training and inference system, and a pan-embodiment operating system to build a technical loop covering intelligent agent design, training, testing, and improvement, and to explore breakthroughs in the data and engineering bottlenecks of physical intelligence R&D.

The team is led by a national-level leading talent and relies on the scientific research resources of Beihang University and BAAI. It has won more than ten international awards, including outstanding paper awards at top conferences and a global third-place finish in the World Humanoid Robot Games. At present, the project is exploring commercialization through simulation data subscriptions and system licensing, and has obtained tens-of-millions-yuan-level orders in fields such as park maintenance and commercial aerospace.

5. Latent Vector Technology: Connecting the virtual and the real with a unified world model of the Earth's physical environment

Latent Vector Technology focuses on a unified world model of the Earth's surface physical environment. Its team independently developed the EWM (Earth World Model), which incorporates external drivers such as atmospheric forcing and solar radiation, surface processes such as soil, vegetation, water bodies, mountains, and the carbon cycle, and human systems such as power stations, farmland, industry, and infrastructure into a unified latent space for representation, exploring the construction of a general-purpose physical foundation model that is intervenable, invertible, and falsifiable.

The project adopts a token-based model service model, and its new energy scenarios have completed empirical validation at 12 sites in 5 countries. In photovoltaic scenarios, the team uses its self-developed Boson series cleaning and O&M robots as execution terminals, jointly building a "perception-decision-execution-feedback" loop with cloud-based models, exploring the path for collaborative implementation of surface physical environment modeling and embodied intelligent systems.

Connecting original technology with capital to drive scientific research outcomes toward industry

From original chips to world models, from brain-inspired embodied intelligence to autonomous robot deployment and physical intelligence in extreme environments, the five projects present diverse explorations of original AI technologies in directions such as underlying hardware, computing architecture, model R&D, and embodied applications. Their common feature is the attempt to transform cutting-edge scientific research capabilities into verifiable, deployable, and sustainably developing technologies and products.

For hard-tech companies founded by scientific research teams, technological breakthroughs in the laboratory are only the starting point. Subsequent engineering validation, product development, scenario adaptation, and business model exploration all require sustained resource investment and industrial collaboration. How to effectively connect scientific research innovation, entrepreneurial teams, capital, and the market is an important link in the industrial transformation of original technologies.

This Yuanchuang Program roadshow, through project presentations and fund negotiations, built a platform for exchange and matchmaking between entrepreneurial teams and investment institutions, promoting further connections between original technologies and capital and industrial resources. The event showcased BAAI scholars' original explorations in different technical directions and also presented diverse paths from mining scientific research outcomes to cultivating entrepreneurial projects.

From the laboratory to the industrial frontline, the transformation of original technologies requires not only sustained scientific research breakthroughs but also the joint support of engineering capabilities, application scenarios, and industrial resources. The Yuanchuang Program will continue to focus on cutting-edge original achievements in artificial intelligence, promote effective alignment between scientific research innovation and industrial needs, and support more promising original technologies in moving toward industrial practice.

BAAI Scholars: From a "Dream Team" to a "New Force in Industry"

The BAAI Scholars Program is a leading initiative by the Beijing Academy of Artificial Intelligence (BAAI) to build a high-level talent team for frontier basic research. It aims to select and cultivate internationally influential academic leaders and young academic talents with development potential, establish collaborative innovation teams oriented toward major research tasks and future frontiers, lead the direction of future artificial intelligence technology development, and build an AI industry innovation ecosystem. Over the past seven years, the BAAI Scholars Program has undergone three stages of iteration—free exploration, task-driven, and innovation-led—and has become an important vehicle for attracting and cultivating top talent in the AI field. The number of BAAI scholars has exceeded 180, forming a global AI "Dream Team."

More importantly, this "Dream Team" is becoming an important source of AI entrepreneurship. In recent years, BAAI has incubated more than 20 innovative companies with core technical capabilities, including Zhipu AI, ModelBest, Galbot, Xingyuanzhi, Inverse Matrix, and Meridian Lightyear. From large models to embodied intelligence, from underlying vision to security governance, the figures of BAAI scholars and the BAAI ecosystem can be seen behind a group of representative AI startups. The launch of the first Yuanchuang Program roadshow is the latest footnote to this transformation of talent potential into industrial momentum.