Bei den chinesischen humanoiden Robotern, welches Prozessor wird verwendet?

In the complex physical world, for humanoid robots to achieve autonomous navigation, precise operation, and environmental interaction, they rely on powerful AI computing power, and at the core of all this is the need for a powerful processor. As the cornerstone of computing power in the robot industry chain, the performance of the processor directly determines the intelligence level and application potential of humanoid robots.

 01 The humanoid robot industry is on the verge of explosion, and chips have become a key variable

The global humanoid robot market is on the eve of an explosion, showing amazing growth potential. Data shows that the global humanoid robot market size will be approximately 9 billion yuan in 2025, and it is expected to soar to 150 billion yuan by 2029, with a compound annual growth rate (CAGR) of over 75%. Among them, industrial handling and medical scenarios will become the core engines driving market growth.

With the continuous iteration of software and hardware technologies for humanoid robots, the expansion of application scenarios has become the focus of industry attention. In the latest paper published by the International Federation of Robotics (IFR) in the second half of 2025, it is pointed out that although the development paths of humanoid robots in different countries vary due to different technological foundations and application purposes, the overall trend is clear: in the short term, it is mainly about pilot substitution; in the medium term, it will gradually enter the stage of large - scale application in the manufacturing and service sectors; in the long term, it is expected to be popularized in daily household scenarios. In this process, high - order system - on - a - chip (SoC) will play an increasingly crucial role and become the core component supporting the complex functions of robots.

From a technical principle perspective, the "intelligent operation" of humanoid robots relies on a complete "brain - cerebellum - limb" collaborative system: the "brain" is responsible for high - level cognitive functions such as voice recognition and environmental perception, and after receiving instructions, it disassembles and plans tasks; the "cerebellum" undertakes motion control tasks such as optimal path planning; finally, the servo system is driven to control the movement of the "limbs" to complete the instructed tasks. In this process, the main chips represented by CPU, GPU, and NPU are the core foundation for humanoid robots to achieve complex algorithm operations and intelligent decision - making, and can be regarded as the real "intelligent core" of robots.

 02 Which processor is used in humanoid robots?

Currently, the global humanoid robot market's processor supply is mainly dominated by two giants, NVIDIA and Intel. Domestic chips are still in the catching - up stage. It is worth noting that among many humanoid robot manufacturers at home and abroad, only Tesla has the ability to independently develop chips. Its Dojo chip is used for AI model training, and the FSD chip is deployed on the robot's edge side for real - time computing and control. Most of the other manufacturers rely on purchasing Intel and NVIDIA chips to build their computing power systems. For example, Ubtech's Walker X uses Intel i7 - 8665U (dual - channel, frequency 1.9Ghz) and NVIDIA GT1030 graphics card (384 cores), and Unitree's H1 - 2 is standard with Intel Core i5 (platform functions) or Intel Core i7 (for user development), with an optional Intel Core i7 or Nvidia Jetson Orin NX (up to three pieces).

The specific application situation is shown in the following table:

From the perspective of functional division of labor, the "brain and cerebellum" of humanoid robots are usually borne by different types of chips:

"Cerebellum" (motion control): Generally, Intel CPUs are used, which are responsible for low - level motion control tasks such as maintaining the robot's balance, trajectory planning, and force control adjustment to ensure precise and stable movements.

"Brain" (cognitive decision - making): Mainly, NVIDIA GPUs are used, which undertake high - level cognitive functions such as environmental perception, voice understanding, and task planning. However, due to the high computing power and relatively high price of NVIDIA chips, they are usually only available as an optional configuration in flagship or high - end humanoid robot products.

As the standard - equipped electronic control and platform function chips, Intel Core i5/i7 have multi - core processing capabilities. Usually, i7 is superior to i5 in terms of core frequency and number of threads, which can meet the needs of the robot's basic control, data processing, and the establishment of the user development environment, and support the operation of algorithms and system management that do not require extreme computing power.

Among NVIDIA's chip products, the Jetson Orin series and Jetson Orin NX are the most widely used. The Jetson Orin series includes 7 modules with the same architecture, which can provide a maximum computing power of 275 trillion operations per second (TOPS). Its performance is 8 times that of the previous - generation multi - modal AI inference chip, and it also supports high - speed interfaces. Its supporting software stack includes pre - trained AI models, reference AI workflows, and vertical application frameworks, which can significantly accelerate the end - to - end development of generative AI, edge AI, and robot applications. Jetson Orin NX features high cost - effectiveness, providing a maximum of 100 TOPS of computing power, and can process complex AI tasks such as visual perception and path planning in parallel, making it a popular choice for mid - to - high - end robots.

On August 25, 2025, NVIDIA further released a computing platform designed specifically for robots - the Jetson AGX Thor developer kit and mass - production module, which is now officially on the market globally (including in China). The starting price of the developer kit is $3499. NVIDIA CEO Jensen Huang called it "the ultimate supercomputer to drive the era of physical AI and general - purpose robots." Industry leaders such as Wang Xingxing, the founder of Unitree, and Wang He, the founder of Galaxy Universal, also recognized its practicality in the robot field.

According to the latest research by TrendForce, the Jetson Thor is centered around a Blackwell GPU and 128GB of memory, providing 2070 FP4 TFLOPS of AI computing power, which is 7.5 times that of the previous - generation Jetson Orin. At the just - concluded WRC 2025 conference, Galaxy Universal's humanoid robot Galbot became one of the first products globally to be equipped with the Jetson Thor chip, demonstrating excellent autonomous box - moving capabilities on - site. Wang He, the founder and CTO of Galaxy Universal, said: "All robot companies, including NVIDIA and Galaxy Universal, now have a common goal of building general - purpose robots."

 03 Domestic breakthrough: Multiple manufacturers are making efforts in chip R & D, with higher cost - effectiveness and customization as advantages

Facing the market dominance of foreign chips, domestic manufacturers have begun to accelerate the independent R & D of humanoid robot chips, trying to achieve a breakthrough in this field. The industry generally believes that for humanoid robots to achieve large - scale application, they must deeply integrate general intelligence with actual scenario requirements, and the realization of this goal depends on four core technologies: algorithms, data, computing power, and hardware. China already has a strong leading advantage in the hardware supply chain field, and the next focus is to develop the "brain and cerebellum" of humanoid robot processors.

Rockchip's RK3588 and RK3588S chips have been adopted by humanoid robots such as ZHIYUAN Lingxi X2, Zhujidongli LimX Oli, and Gaoqing Pi/Pi +. The core architectures and computing power of the two chips are exactly the same, and the main differences lie in interface expandability, package size, and power consumption. RK3588 supports more interfaces and is suitable for scenarios with high requirements for external device connection; RK3588S has a smaller package size and lower power consumption, making it more suitable for robot products that are sensitive to space and energy consumption.

As Rockchip's 8K flagship SoC chip, RK3588 uses the ARM architecture and was originally mainly targeted at PCs, edge computing devices, personal mobile Internet devices, and digital multimedia applications. Now, it shows strong potential in the robot field. It integrates a quad - core Cortex - A76 and a quad - core Cortex - A55 processor, and is paired with a separate NEON coprocessor, supporting 8K video encoding and decoding. At the same time, it has a variety of high - performance embedded hardware engines built - in, which can optimize performance for high - end applications. In terms of AI computing power, the NPU of RK3588 supports INT4/INT8/INT16/FP16 mixed operations, with a computing power of up to 6 TOPS, and has extremely strong compatibility. Network models based on mainstream frameworks such as TensorFlow, MXNet, PyTorch, and Caffe can be easily converted and adapted.

The newly launched RDK S100 development kit of Diguajiqiren under Horizon Robotics innovatively integrates the robot's "brain" (computing function) and "cerebellum" (control function) on a single SoC chip, greatly simplifying the robot's hardware architecture. This development kit is in the form of a board, providing a rich set of peripheral interfaces, which can be directly connected to components such as cameras, sensors, and actuators, facilitating embedding in various robot systems.

From a technical architecture perspective, the RDK S100 has an on - board CPU + BPU + MCU heterogeneous computing architecture, which can simultaneously undertake two core tasks: high - performance AI computing and real - time motion control, achieving the full - closed - loop function of "environmental perception - decision - making and planning - low - level servo control." This means that a single RDK S100 development board can replace the traditional combination of "edge AI board + independent controller" and become the "intelligent center" of the robot, significantly reducing system complexity and development costs.

Specifically for the computing unit, the single SoC chip of RDK S100 integrates three types of cores that work together:

"Brain" part: It consists of a 6 - core CPU and a high - computing - power BPU (Brain Processing Unit). The 6 - core general - purpose processor is responsible for complex logical operations and task scheduling. The BPU based on Horizon's new - generation self - developed "Nash" architecture is optimized specifically for deep neural networks (CNN/Transformer), which can provide 80 TOPS (RDK S100) or 128 TOPS (RDK S100P) of AI inference computing power to meet cognitive needs such as environmental perception and voice understanding.

"Cerebellum" part: It consists of 4 Cortex - R52 + cores forming an independent MCU, which operates in the Lock - Step mode to ensure high reliability and functional safety of motion control and can precisely coordinate joint motors and maintain the robot's balance.

Black Sesame Intelligence is cooperating with multiple domestic humanoid robot enterprises to develop embodied intelligence technology. The most representative one is the strategic cooperation with the team led by Academician Liu Sheng, the executive dean of the School of Industrial Science at Wuhan University. The two sides use Wuhan University's independently developed first humanoid robot "Tianwen" as the core carrier, and Black Sesame Intelligence provides it with a dual - chip solution of "Huashan A2000" ("brain") and "Wudang C1236" ("cerebellum"). The computing power of the A2000 chip is comparable to that of 4 NVIDIA OrinX chips, supporting embodied intelligence algorithms and capable of processing multi - modal environmental information and making intelligent decisions. The C1236 chip enables parallel processing of AI computing and control tasks, ensuring stability in complex environments.

CloudWalk Technology also stated on the investor relations platform that the company is developing a new - generation "brain" chip, the DeepXBot series, to accelerate the inference tasks of perception, cognition, decision - making, and control in humanoid robots.

From the perspective of competitive advantages, the core highlights of domestic chips are higher cost - effectiveness and more market - oriented customized services. Take Diguajiqiren's RDK S100 as an example. Its price is only 2799 yuan, almost half of the price of NVIDIA's solution with the same computing power, which significantly reduces the R & D and production costs of mid - to - low - end humanoid robots. At the same time, domestic manufacturers can adjust the chip functions and interfaces according to the specific scenario requirements of robot manufacturers (such as industrial handling, household services, education and scientific research) and provide more flexible solutions.

 04 Future trend: "Integration of brain and cerebellum" becomes the breakthrough direction

Similar to the functional division of the human brain, the current controllers of humanoid robots generally adopt a "brain - cerebellum" separated architecture: the "brain" is responsible for perceiving the environment, planning routes, and making intelligent decisions (such as recognizing gestures, understanding voices, and autonomously learning new skills); the "cerebellum" is like a "sports expert," coordinating joint motors at a frequency of thousands of times per second to ensure that the robot does not fall when dancing and does not shake when carrying things.

The "integration of brain and cerebellum" architecture refers to the deep collaboration between the cognitive decision - making system (brain) and the motion control system (cerebellum), achieving seamless connection of "perception - decision - making - execution" through the integration of software and hardware design. The proposal and evolution of this architecture are the core context of the development of embodied intelligence. Its concept originates from the cross - integration of brain science and AI, aiming to simulate the division of labor and cooperation mechanism between high - level cognition and motor coordination in the human nervous system, making the robot's "thinking" and "actions" more synchronized and efficient.

The current mainstream "separated brain and cerebellum" solution has gradually shown obvious bottlenecks:

Surge in computing power demand: The robot needs to process real - time control (cerebellum) and complex decision - making (brain) tasks simultaneously, which significantly increases the demand for heterogeneous computing power and leads to an increase in hardware costs.

Obvious communication delay: The "brain" and "cerebellum" belong to different hardware systems, and there is a delay in data transmission, which may cause the robot's actions and decisions to be out of sync and affect the operation accuracy.

High development cost: Developers need to maintain two independent code systems. The control code may run on an Arm CPU or x86 CPU, while the AI algorithm needs to run on a GPU or other dedicated modules, making code adaptation and debugging difficult.

Difficulty in sensor fusion: The separation of hardware makes it difficult to efficiently integrate the data of various sensors (such as cameras, force sensors, and gyroscopes), which affects the robot's comprehensive judgment of the environment.

In contrast, the "integration of brain and cerebellum" architecture can solve the above problems through a single - chip or integrated hardware design and will become the mainstream development direction of future humanoid robot controllers.

Recently, NVIDIA and Intel announced a partnership. Public information shows that in the data center field, Intel will customize x86 CPUs for NVIDIA, which will be integrated into the artificial intelligence infrastructure platform by NVIDIA and put on the market. In the personal computing field, Intel will produce and supply x86 system - on - a - chip (SOC) integrated with NVIDIA RTX GPU chips to the market. NVIDIA will invest $5 billion in Intel's common stock at a price of $23.28 per share.

It is worth noting that in the field of humanoid robots, most current solutions adopt the "separated brain and cerebellum" architecture of Intel CPU + NVIDIA GPU. However, with this cooperation, in the future, an SoC with the "integration of brain and cerebellum" architecture may be launched. The integrated - architecture SoC can better integrate into the X86 and CUDA ecosystems and provide developers with a more powerful intelligent core.

Although the humanoid robot market has broad prospects, there are still many challenges to be solved before it can be truly mass - produced and commercially used on a large scale:

Insufficient data accumulation: Embodied intelligence requires a large amount of real - world scenario data to train models. Currently, the application scenarios of humanoid robots are limited, and the quantity and diversity of data are difficult to meet the needs of general intelligence.

Hardware architecture needs optimization: In addition to the "integration of brain and cerebellum," the computing power density, power consumption control, and heat dissipation performance of chips still