NXP Proposes Neural Axis Architecture: Giving Robots a "Spinal Cord"
During the ICCHINA exhibition, NXP Semiconductors proposed the "Neural Axis Architecture," which breaks down robot computing power into three layers: brain reasoning, cerebellum coordination, and spinal cord reflex. Local processing at the endpoints can achieve response times at the microsecond level. This architecture is built on three technical pillars: TSN time-sensitive networking, a three-in-one cerebellum chip, and intelligent neural endpoints. The same architecture can be tailored to different forms such as AMRs, wheeled humanoids, and bipedal humanoids. NXP cited a UBS forecast that the market size of robot-related semiconductors will be approximately US$6 billion in 2030 and approach US$100 billion in 2040.
Body · AI translation
To give robots a reflex arc, NXP decided to add a spinal cord. Original source: Semiconductor Industry Insight. Published: 2026-11-13 18:46. Author: Zihao. AI humanoid robot poster: accelerating physical AI with a neural axis architecture.
The robotics sector has continued to gain momentum recently. Many companies focused on automotive in the year before last and on data centers last year, and this year they have listed robotics as a key area of investment. During the ICCHINA exhibition, Yuan Wenbo, Vice President of NXP Semiconductors, gave an interview to Semiconductor Industry Insight and put forward a rather novel idea: in addition to a "brain" with powerful computing capability and a "cerebellum" for coordinated movement, robots should also be equipped with a "spinal cord" that responds more quickly. Around this idea, NXP launched the "Neural Axis" architecture, whose core concept is to break up centralized computing power and distribute it to the various sensing and actuation endpoints of the system.
01 Four types of problems facing robots on the path to mass production
Yuan Wenbo grouped the problems that robots will encounter at scale into four aspects: latency, mechanics, power consumption, and perception.
On latency, he said that the latency requirements of robotic systems from perception to actuation are distributed by layer between 300 milliseconds and 5 milliseconds. This metric is directly related to a robot's braking distance in crowded environments and also affects stability during high-speed motion. Networks are one of the main sources of latency problems. At present, multiple buses coexist inside robots, including CAN, Ethernet, I3C, and various vendor-proprietary buses, each handling part of the functions. Data must be converted between different protocols, and real-time performance and determinism are affected during conversion. As the number of cameras and sensors increases, the bandwidth demand for multimodal data also rises.
On mechanics, the constraints mainly come from the robot's own structure. Space inside joints is limited, and heat dissipation and wiring layout compete with each other for it. Larger electronic components reduce the room available for degrees-of-freedom design. Yuan Wenbo also mentioned joint durability and wiring management, and he believes these problems will be more prominent in the mass production stage than in the prototype stage.
On power consumption, battery management and computing power per watt determine battery life, and energy efficiency directly affects the device's usage time and operating costs.
On perception, cameras have insufficient coverage at corners, in backlight, and under conditions such as dust and rain or fog. NXP's proposed supplementary solution is millimeter-wave radar and UWB, using wireless methods to cover areas that cameras cannot.
Yuan Wenbo gave an overall judgment: "We do not want to use a single chip to solve all problems; rather, we want to treat it as an entire system."
02 Why add a spinal cord
Why, when major vendors are all working hard on the robot brain and cerebellum, did NXP propose adding a spinal cord? Yuan Wenbo used the human body as an analogy: when a person's hand touches a hot object, the signal travels to the spinal cord and directly directs the hand to pull back. This action does not require the brain to make a decision; only afterward is the pain signal sent up to the brain, and that is when we feel pain. Robots should also have this capability, so that localized processing at endpoint nodes can respond before central scheduling in emergency situations, with response times reaching the microsecond level, while reasoning and decision-making on the brain side usually occur at the millisecond level. Placing instinctive reactions in the place closest to action means the shorter the distance, the faster the response, and it also helps reduce energy consumption. For sudden events, hesitation can lead to bigger problems. To this end, NXP proposed the "Neural Axis Architecture," which views the robot as a biomimetic nervous system: the brain handles reasoning, the cerebellum handles coordination, and the spinal cord handles reflexes. Each of the three layers has its own response time, power budget, and safety boundary, and they work together through a unified network and a unified safety foundation.
In the interview, Yuan Wenbo also distinguished two concepts that are often confused. Embodied intelligence describes the physical entity in which intelligence is carried; robots, quadruped devices, and drones all fall into this category. Physical AI describes a technical paradigm in which perception, decision-making, and actuation form a closed loop on the device side. The Neural Axis architecture serves the latter, so the same architecture can correspond to different robot forms. In addition to robots, the architecture is also applicable to other physical AI carriers. Although automobiles, AMRs, and drones differ in structure, they can likewise adopt a layered design. According to him, this architecture is scalable and can serve customers in different regions.
03 Three technical pillars of the platform
At the specific product level, NXP summarizes the platform's capabilities into three directions.
The first is Time-Sensitive Networking (TSN). Yuan Wenbo believes that TSN is the main direction for the next generation of real-time networks for robots. At present, bus standards inside robots are fragmented, and TSN can converge these standards onto a schedulable and predictable deterministic network while also providing greater bandwidth to carry multimodal sensor data. Most of NXP's processors and switches natively support TSN, and the company is also promoting the deployment of commercial protocol stacks. He said that scenarios with higher requirements for bandwidth and scalability will increasingly adopt TSN, and NXP's solution supports both TSN and EtherCAT, which can be combined according to different real-time level requirements.
The second is the functional reconstruction of the cerebellum. In existing robot solutions, the cerebellum is mostly used for motion control. NXP has expanded it into a three-in-one combination of motion control, functional safety, and network communication integrated into the same chip. After the adjustment, the brain handles millisecond-level reasoning and decision-making, while the cerebellum handles real-time deterministic control and coordinated scheduling, and is also responsible for safety monitoring and network data distribution. The cerebellum thus changes from a pure execution unit into the real-time coordination node of the entire robot.
The third is the intelligentization of nerve endings. Dexterous hands and grippers are beginning to have local computing power, enabling on-device environment detection, on-the-fly adjustment and collision response, before sending the processed data back to the central unit. On the hardware side, NXP has introduced small-package and SIP solutions to fit narrow spaces such as finger joints. In addition, the company offers a 60GHz wireless connectivity solution that replaces some of the wiring harnesses inside joints with wireless links, improving joint durability and reducing assembly complexity. In the perception stage, visual-tactile sensors, multimodal data, models and chips need to form a closed loop—this is NXP's vision for the next generation of dexterous hands.
Another role of the architecture is reflected in form-factor adaptation. Robot forms have not yet converged: AMRs, AMRs with humanoid torsos, wheeled humanoids, bipedal humanoids and stair-climbing humanoids differ markedly in workspace, degrees of freedom and cost structure. NXP adopts a combination of a centralized cerebellum and a distributed spinal cord, placing coordination and safety functions at the central node and pushing reflexes and perception down to the endpoints. The same architecture can be tailored to different form factors, with function combinations and upgrades realized through software definition and supporting OTA updates. For customers, changing form factors no longer requires redesigning the entire electrical/electronic architecture.
04 The robot market has considerable prospects
NXP's 2025 revenue was US$12.27 billion, and robotics and physical AI are listed as one of the key directions of its industrial and IoT business. NXP cites a UBS forecast that the robot-related semiconductor market will reach about US$6 billion in 2030, between US$20 billion and US$50 billion in 2035, and close to US$100 billion in 2040. By comparison, the automotive semiconductor market will be about US$80 billion in 2035.
English text is being prepared.
The position of the Chinese market is an important variable in this set of forecasts. Statistics from Deutsche Bank in June 2026 show that more than 85% of global humanoid robot shipments come from China. At the policy level, robotics and embodied AI are listed among the national priority development directions for 2026 to 2030. Data at the enterprise level is similarly concentrated. Unitree Technology shipped more than 5,500 units in 2025, ranking first globally; Zhiyuan Robotics' 2025 revenue reached RMB 1 billion; UBTECH's humanoid robot-related revenue grew 23-fold year on year. In 2026, manufacturers such as Yunshenchu Technology, Leju Robotics, Galaxy General and Booster Robotics continue to build out their positions.
Yuan Wenbo also mentioned changes in the demand structure. Robots' requirements for wide temperature ranges, vibration and long-term reliability are converging toward automotive-grade standards, and some automotive customers have also begun to enter the robotics field. NXP's next-generation products will consider both types of demand, migrating the company's automotive-grade accumulation in temperature and reliability to robotics, while applying robotics' real-time control experience to automotive products. He believes robotics will become an important growth engine for the company, but at the platform level it will share the same technology foundation with the automotive and industrial businesses.
05 Safety is a precondition for mass production
Yuan Wenbo summarized NXP's strategy as trustworthy physical AI, the core of which is to ensure deterministic control, functional safety and information security. As robots enter homes and production lines, functional safety and information security are becoming preconditions for mass production. The industry still lacks a set of safety standards spanning perception, networking and computing, and API interface specifications and certification methods need to be advanced jointly by the industrial chain.
NXP's safety capabilities mainly come from its accumulation in the automotive and industrial fields. EdgeLock 2GO provides credential management, secure OTA and verified software, enabling devices to maintain cryptographic agility without changing hardware. In response to the risks that quantum computing may bring, the company plans to incorporate post-quantum cryptography (PQC) into its certification system. The reason he gives is the product lifecycle: robot products have long lifecycles, while the timeline of the quantum threat remains unclear, so security capabilities need to be updated as threats change and cannot be fixed at the factory.
In terms of the product portfolio, the brain uses the i.MX series of application processors, the cerebellum uses the S32 series aimed at high functional safety, sensors, motor drives and battery management are handled by MCX and analog devices, and the endpoints extend to joint modules and dexterous hands. Current solutions include joint and multi-motor drive reference designs based on the i.MX RT series, ultra-small-package motor modules for small joints, and a dexterous hand design that completes preliminary data processing in the palm. NXP's goal is to enable the same software stack to be reused at different levels, rather than pursuing single-point performance.
06 The Chinese market is becoming increasingly important
Yuan Wenbo said that China is transforming from a sales and market center into a source of product definition and localized innovation.
In terms of production and marketing, NXP's China business unit incorporates R&D, marketing and product definition into the same closed loop; the Tianjin assembly and test plant has about 4,000 employees and remains NXP's largest manufacturing base in China; the company carries out platform-level cooperation with local wafer fabs such as SMIC, integrating partners' processes into its own technology platform; this year it established a China Robotics Innovation Center; in the assembly and test segment, in addition to the Tianjin plant, the company is also working with a number of domestic assembly and test plants, and some overseas product lines also use this capacity. The completion of the 12-inch wafer fab in Singapore is incorporated into the global supply network. Yuan Wenbo said that this fab complements the Chinese plants and local partners, and its geographical location makes it convenient to serve Southeast Asia and the Chinese market, and the company's localization work for China will continue to advance.
In terms of ecosystem building, the TSN commercial protocol stack needs to be promoted jointly with Chinese industry partners; on the brain side, the company maintains cooperation with NVIDIA; on the operating system side, it connects with the ROS community and local software companies. The company lists software partners as a priority for its 2027 ecosystem work, because what customers need are system-level solutions that can run directly, rather than a single chip and a datasheet.
From "making robots smarter" to "making robots more trustworthy." In the next stage of physical AI, what is being compared is no longer the peak computing power of a single chip, but whether the entire system can achieve low latency, predictability and recoverability in real-world scenarios. NXP's answer is to break intelligence back down to the nerves themselves: the brain thinks far ahead, the cerebellum controls steadily, and the spinal cord reacts quickly—each of the three layers performs its own role, and they are then stitched into a network through TSN, functional safety and edge computing.
This content is original to the platform, and the copyright belongs to the platform. Reproduction is not permitted without permission; to reprint, please contact the platform. Comments: No user comments yet. Latest News More Related Recommendations
Planning to raise RMB 8 billion, xFusion is expected to become the largest IPO on the ChiNext this year 2026-11-13 18:40
Chiplet interconnection, driving the future | HIPI CHIPLET FORUM 2026 full agenda grandly announced, you are cordially invited to attend 2026-11-13 18:40
Bypassing HBM and liquid cooling! European AI chip company launches Titania architecture 2026-11-13 18:40