Unitree Invests 2 Billion in AI Brain Development for Robots as Shanghai Advances Embodied Intelligence Ecosystem

Unitree Technology is making a significant investment of 2 billion yuan in the development of intelligent robots, aiming to equip them with a “Chinese brain.” Recent advancements in robotics have enabled machines to perform impressive feats such as flying kicks, aerial walking, and ground spins. However, industry experts recognize that only a few out of the numerous types of existing robots can achieve stable movement when faced with challenging terrain. This discrepancy highlights the current state of embodied intelligence: while the “little brain”—the motion system powered by reinforcement learning—has progressed remarkably, the true challenge lies in developing the “big brain.”

Countries around the world are striving to make breakthroughs in robotic “brains.” The ability to overcome core challenges in perception, decision-making, and evolution will determine who gains a competitive edge in the next technological race. Recently, Unitree Technology raised 4.202 billion yuan through its IPO, with nearly half (2.022 billion yuan) allocated to a core project focused on developing intelligent robot models, emphasizing the significance of this initiative.

At the inaugural Embodied Intelligence Brain Technology Ecosystem Conference, Shanghai launched a series of initiatives aimed at establishing standards for humanoid robot AI software and hardware, as well as a global co-construction plan for embodied intelligence ecosystems. The unveiling of the Embodied Intelligence Brain Research Institute signals Shanghai’s ambition to make strides in this field.

Major tech companies are vying to develop advanced “brains” for robots. Elon Musk has adapted the FSD (Full Self-Driving) chip used in Tesla vehicles for the humanoid robot, Optimus. This custom SoC chip can perform hundreds of trillions of operations per second, enabling the execution of complex neural network models. The multimodal fusion network integrates visual, tactile, audio, and motion data, with a latency of less than 50 milliseconds, allowing robots to perform complex tasks such as “pouring coffee while listening to commands.”

Boston Dynamics and Toyota Research Institute have jointly released the AtlasMTS model, which features 450 million parameters. This model acts like a conductor with a “global eye,” overseeing the relationship between overall structure and detailed movements. Meanwhile, Google is investing in VLA (Vision-Language-Action) models and embodied reasoning. In September of last year, DeepMind introduced the Gemini Robotics 1.5 series, focusing on enabling robots to “think before acting,” integrating language, images, and physical reasoning.

Different technological pathways reflect varying interpretations of what constitutes a “brain.” Some equate it with GPU chips, while others simplify it to large models. Yan Weixin, chief scientist at the Shanghai Artificial Intelligence Research Institute, believes that the human brain represents an optimal structure formed through millennia of evolution. He suggests that studying the human brain is the correct approach to developing robotic brains, proposing that the functions of carbon-based brains should be mapped to silicon-based systems. Thus, an embodied intelligence brain must encompass four essential elements: appropriate chips, a dedicated operating system, a complete brain-like framework, and a hardware system.

Lingjing Zhiyuan, a key incubator at the Shanghai Artificial Intelligence Research Institute, showcases another approach to “building brains” through four fully localized technological achievements that span the entire spectrum of computing architecture, perception processing, motion control, and brain base. The domestically produced embodied “little brain” T40 enhances traditional robot movement by integrating visual perception and motion control through deep human-like imitation learning, allowing robots to navigate obstacles effectively.

The Si-Cortex S100, a product of Lingjing Zhiyuan’s embodied intelligence brain, serves as the “perception hub” for robots. This industry-first real-time multimodal fusion system achieves millisecond-level perception processing, handling signals from 16 high-definition visual and radar sources with latencies below 1 millisecond, nearly matching biological neural response times. The domestically produced embodied brain models T200 and T80 successfully overcome the challenges of size, computing power, and low energy consumption, providing a computational foundation for real-time intelligent decision-making in robots.

Lingjing Zhiyuan’s innovative Dvořák Super-Heterogeneous Architecture represents a significant technological ambition. According to founder Sun Bo, traditional robotic systems have independent modules for “brain” (decision-making), “little brain” (motion control), and “perception system” (environment sensing). Lingjing Zhiyuan integrates these components at the hardware level, reducing the closed-loop delay between perception and action to under 5 milliseconds, enabling robotic brains to think, accumulate experiences, learn from mistakes, and evolve continuously.

To date, Lingjing Zhiyuan has connected with over 250 industry clients, collaborating with more than 100 and achieving a coverage rate of 72% among the top 150 clients in the sector.

In Shanghai, the focus is shifting from “single-point breakthroughs” to “ecosystem co-construction.” The “Global Embodied Intelligence Brain Ecosystem Co-construction Plan” emerged as a highlight of the conference, with over ten leading industry companies, including Unitree Shanghai, Lingjing Zhiyuan, Fourier Intelligence, and Qingtian Intelligence, signing a collaborative agreement. Shanghai aims to unify the disparate forces across various stages, including chips, algorithms, manufacturing, and application scenarios.

The establishment of the Embodied Intelligence Brain Research Institute marks a systematic upgrade in China’s embodied intelligence sector, transitioning from technology development to industrial implementation and ecosystem construction. This institute will serve as the technical source for China’s embodied intelligence solutions.

As early as November 2024, under the guidance of the Ministry of Industry and Information Technology and the China Computer Industry Association, the Shanghai Artificial Intelligence Research Institute initiated a working group for AI software and hardware integration in humanoid robots, collaborating with leading domestic chip manufacturers. This group aims to create a public platform integrating technological collaboration, standard co-construction, and ecosystem integration.

In terms of standard evaluation and promotion, the working group relies on resources from various national committees and aims to research standards for humanoid robots and embodied intelligence. It is expected to complete over five national and industry standards and introduce five replicable and implementable application scenarios.

According to the Shanghai Implementation Plan for Embodied Intelligence Industry Development, by 2027, Shanghai aims to achieve breakthroughs in at least 20 core algorithms and technologies, gather hundreds of industry-leading enterprises, and promote hundreds of internationally leading products, with the scale of the core embodied intelligence industry exceeding 50 billion yuan.

In February of this year, Minhang District also released measures to promote innovation and development in the embodied intelligence industry, allocating substantial funds to encourage companies to invest in this future-oriented sector.