Former Huawei Embodied Brain Leader Found Startup with Cognitive Science World Model, Secures Hundreds of Millions in Funding

In 2026, world models have become the hottest currency in AI. Yann LeCun's newly founded AMI Labs after leaving Meta secured $1.03 billion in funding at a $3.5 billion pre-money valuation. Fei-Fei Li's World Labs continues to bet on spatial intelligence, becoming one of the most sought-after targets in the primary market. NVIDIA's GTC unveiled Physical AI, claiming the potential market for industrial and robotics could reach $50 trillion. Capital, talent, and top minds are all converging on one goal: making AI truly understand the physical world.

Amid this wave, a new startup called Junaopanshi has emerged with a fresh approach. Founded by Zhu Senhua, former head of Huawei Cloud AI Algorithms Innovation Lab, the company is developing a "Cognitive World Model" rooted in cognitive neuroscience. Zhu, often referred to as "Huawei's No.1 in Embodied Brain," led the AI brain science cloud platform, the PanGu embodied large model, and the Global Embodied Intelligence Industry Innovation Center at Huawei. Now, he is applying that expertise to build a robot brain that can reason, remember, and self-update.

The dominant technical paradigm over the past two years has been VLA (Vision-Language-Action models), which unifies visual, linguistic, and motor capabilities into an end-to-end model. While VLA has proven effective in controlled environments like sorting, loading, and simple assembly, its limitations are becoming apparent. It requires vast amounts of data, struggles with generalization, and lacks the ability to accumulate experience. As NVIDIA's robotics lead Jim Fan stated at the Sequoia AI Summit, the industry is shifting from VLA to World Action Models.

Junaopanshi's Cognitive World Model is built on a five-layer framework. The first layer is visual reality, moving from 2D images to 3D spatial understanding. The second is physical reality, modeling gravity, friction, and collision. The third is interactive reality, emphasizing simulation and reinforcement learning. The fourth is abstract representation, which Yann LeCun's JEPA (Joint Embedding Predictive Architecture) exemplifies by learning state changes in latent space without pixel-level prediction. The fifth and most critical layer is active inference, rooted in cognitive neuroscience theories like free energy minimization, predictive coding, and Bayesian brain. This layer enables the robot to form hypotheses, predict outcomes, execute actions, and update its internal model based on feedback.

Junaopanshi's technical roadmap mirrors JEPA but extends it for embodied deployment. While JEPA focuses on representation and prediction, Junaopanshi's Cognitive World Model aims to create a complete loop: perception, cognition, prediction, planning, action, feedback, and learning. The company has broken down its research into four areas: brain-like perception encoding for multimodal fusion, cognitive dynamic prediction for understanding physical laws, lifelong learning memory for experience accumulation, and low-power sparse computing for edge deployment.

The startup recently completed a new funding round of hundreds of millions of yuan, led by top-tier industrial capital with follow-on investments from existing funds and other leading investment institutions. A subsequent round is also underway. The funds will be used for core technology R&D, team expansion, and global market entry.

Junaopanshi's team is uniquely positioned. In addition to Zhu Senhua's deep background in AI and brain science, co-founder Liu Jinyu brings productization and global business experience, having led multiple AI robotics product lines from zero to one. The core team includes graduates from Tsinghua, Peking University, Fudan, and the Chinese Academy of Sciences, with experience at Huawei, Lenovo, Megvii, and Geek+.

On the commercialization front, the company is taking a pragmatic approach. Rather than waiting for a fully mature general-purpose robot brain, it is focusing on reusable modules such as embodied skill learning, cognitive mapping for navigation, and multi-robot collaboration. These modules can generate revenue and data in real-world scenarios, funding further R&D. Junaopanshi is already conducting PoCs with multiple industry customers both domestically and internationally.

The ultimate goal is to achieve "one brain, multiple robots, multiple forms" — a single cognitive world model adaptable to different robot embodiments. In the short term, this means multi-robot coordination; in the medium term, a unified model for various hardware; and in the long term, an open general-purpose embodied brain model for the industry. The key test for Junaopanshi now is to make the Cognitive World Model work in real robotic tasks, a challenge that could define the next era of embodied AI.