Apple Research Team Discloses Progress on Robot 'Brain' System

Cailian Press November 7 news (Editor Wei Qi) Apple research team discloses research progress on robot “brains.” On November 4 local time, the Apple research team submitted the paper “Perceptive Humanoid Motion from Simulation” to the preprint platform arXiv, disclosing the autonomous control system for humanoid robots. HALO.The research team will deploy the system on Unitree's G1 humanoid robot for verification, allowing the robot to bring the actions learned in the simulation environment to be executed in unfamiliar real rooms.

The robot's brain is the high-level decision-making center of the intelligent robot system, responsible for processing complex cognitive tasks and intelligent decisions. It consists of various processors, memories, and algorithms, primarily used to process and analyze data from sensors and make corresponding decisions based on this.

Zhongtai Securities research report pointed out that the brain is Embodied Intelligence one of the core variables from demonstration to practical use. Multiple robot manufacturers have listed insufficient generalization capability as the biggest constraint in public, and technical routes also revolve around VLA and world model fusion and expansion of competition. Capital, policies, and industrial strategies are concentrating toward the brain in the same time window, reflecting the industry's consensus on the judgment that “the physical threshold is lowering, and the brain determines the upper limit.” On the capital side, funds are moving toward, visual–language–action models, and world–Concentration of direction. In the first half of 2026, domestic embodied intelligence track financing was approximately 43,800,000,000 yuan, of which companies with embodied brains as core businesses absorbed 22,253,000,000 yuan, accounting for 50.8%, while robot body complete machine companies' financing during the same period was only 5,600,000,000 yuan, accounting for 12.8%.

On the international side, PhysicalIntelligence has obtained a valuation of 20,000,000,000 USD. Physical Intelligence (abbreviated as PI or π) was founded in March 2024 by research and technology giants including a professor team from UC Berkeley, Stanford University, and Google scientists. The company is committed to creating software that can add advanced intelligence to various robots and machines, with the ultimate goal of creating an AI model as a general-purpose robot system (universal robot brain). The flagship model pi-zero launched by Physical Intelligence is precisely an attempt to solve this difficult problem. This model directly binds images, text, and robot actions, using data collected by real robots in physical environments for training.

On the domestic side, AGIBOT GO-3Preview connects the full chain of pre-training, post-training, and post-deployment learning, and the mobile packaging success rate rose from 50.8% to 98.3%, verified for the first time in real business scenarios PhysicalRSI. XPeng IRON completed the second-generation VLA side deployment, with 0.2-second re-inference and action pre-judgment within 6 seconds, with 3 self-developed Turing chips providing 2250TOPS. Tuoyuan Wisdom open-sourced the 4D world action model E0.5, with a real machine success rate of 85.0%/88.3%, higher than π0.5 by 13.3/26.7 percentage points.

Leading robot manufacturers have already placed the robot brain as a research and development priority. Tesla CEO Musk has clarified multiple times that for Optimus to truly enter home scenarios, it must solve three core problems: hands, brain, and mass production, and an “artificial intelligence brain that understands the physical world” is listed as one of the core challenges. Unitree founder Wang Xingxing defined “insufficient generalization ability” as the biggest bottleneck of global embodied intelligence at the 2026 World Robot Conference, and revealed that the company is pre-researching a self-evolution system for physical AI robots, allowing AI large models to autonomously retrieve academic papers, automatically generate control code, and be deployed to real machines after verification in simulation.

According to an incomplete count by Cailian Press, the main companies among A-share listed enterprises that are positioning in the robot brain field include:

Zhiwei Intelligence: The Thor robot brain controller of the Zhiqing series has added leading humanoid robot collaborative customers such as AGIBOT, becoming one of AGIBOT's first "A-Chain" suppliers; meanwhile, the company has already entered into joint development cooperation with UBTECH, and this technical solution helps to reduce the overall machine cost and shrink the controller volume.

Tianqi Shares: Participated in Shanghai Junao Panshi Technology Co., Ltd., the latter focuses on the research and development of brain-like cognitive world models and embodied intelligent robot brains, focusing on the fusion route of brain science and artificial intelligence, and building high-reliability embodied brain capabilities for real-world scenarios.

iFlytek: Provides robot brains and human-machine interaction solutions for robot manufacturers, core profit models include: robot AI technology licensing, integrated software and hardware module sales, and joint ecosystem partners creating and selling robot scenario-based solutions.

Tianzhun Technology: In 2025, the company's output will increase significantly, among which is the robot brain in the embodied intelligence solution field Domain Controller Commercial vehicle domain controller business is increasing relatively quickly, and PCB-related LDI products in the visual process equipment field and intelligent equipment in the new automotive field are increasing relatively quickly.

Hown Automotive Electronics: The domain controller developed by the company in cooperation with NVIDIA is the brain for robots, which can be applied in multiple fields such as humanoid robots, industrial robots, and logistics robots.

Zhongke Chuangda: The company mainly engages in the R&D of intelligent robot brain modules (chip + operating system + core algorithms) products, providing computing and processing capabilities for intelligent robots.

Nengke Technology: The company's “Lingzhi” embodied intelligence AI training and inference platform is oriented towards embodied agents in industrial scenarios. Relying on technologies such as digital twins, physical simulation, and vertical domain large model training and inference, it provides training and inference support for the perception, decision-making, and execution of embodied intelligence in the physical world, developing and accompanying its “brain system.”

Dongtu Technology: Company Hongdao Operating System Relying on a microkernel virtualization architecture, it achieves the fusion of the brain and cerebellum of embodied robots, unifying multi-device and multi-node tasks into AI systems and motion systems under the architecture, and achieves low-cost, low-power, and small-volume deployment goals through integrated software and hardware technology.