“Physical AI” is trending yet again. From Jensen Huang’s assertion earlier this year that physical AI’s “ChatGPT moment” has arrived, to his recent decision to frame an entire keynote around the declaration that “the era of physical AI has arrived,” and then to Wang Xingxing’s remark in Hangzhou in October that “we can already see signs of physical AI breaking through”—a consensus is rapidly taking shape: the next frontier for AI is no longer confined to text and images on a screen, but rather extends into the physical world to perceive, reason, and act.
The rise of robotics has paved the way for physical AI, and the technology is expected to subsequently expand into other terminal systems such as automobiles and drones. Combined revenue across the sensing, imaging, and RF markets approached USD$100 billion in 2025 and is projected to grow at a compound annual growth rate (CAGR) of 4.4% to reach USD$128.8 billion by 2031.
Sensing, Imaging, and RF Market Revenue
As the sensing, imaging, and RF industries mature, companies across these sectors are actively building and integrating specialized capabilities to better address the challenges of product development, large-scale manufacturing, and profitability in markets such as automotive, consumer electronics, artificial intelligence (AI), and robotics.
A New AI Perception Stack
For decades, machine perception typically meant mounting a camera on a robotic arm and using a limited set of rules to read barcodes or inspect welds. Today, the perception stack—encompassing imaging, sensing, and RF technology—has become the “sensory nervous system” of a new generation of machines, including humanoid robots, autonomous mobile platforms, drones, and self-driving cars. What has changed is not only the hardware but the capabilities behind it. The rise of generative AI and multimodal foundation models has transformed perception from a narrow, task-specific process into a general-purpose understanding layer that can not only detect scenes but also reason about them. This is the essence of physical AI: intelligence that goes beyond generating text or images to perceive, reason, and act in the physical world through sensor fusion, connectivity, and powerful computing.
The Foundation of Physical AI: MEMS Sensors
According to the Status of the MEMS Industry 2026 Report, the MEMS market reached USD$17.1 billion in 2025, with shipments of MEMS sensors and actuators exceeding 33 billion units. Yole Group expects the market to reach USD$24.4 billion by 2031, reflecting a CAGR of more than 6%. Led by companies such as Bosch, STMicroelectronics, and TDK, MEMS devices are now widely deployed across devices ranging from smartphones and automobiles to data centers, supporting hundreds of applications including motion tracking, audio interaction, magnetic field sensing, and light projection.
Physical AI has redefined the value of MEMS devices. For models to reason about the physical world, they must first have the means to measure it. MEMS sensors are the cornerstone of this stack, converting motion, sound, pressure, and light into the data required for model training and operation. The emergence of humanoid robots underscores this dependency. To navigate, manipulate objects, and interact, they require high-density, low-power, and cost-effective sensing deployed at scale—precisely the domain in which MEMS devices have already achieved billions of units in shipments.
For decades, the MEMS industry has focused on better converting physical quantities into digital values, and this has enabled MEMS sensors to rapidly advance humanoid robotics. The deeper shift lies in a change of direction: in the future, physical AI will need the ability to “act back”—that is, through actuator components such as MEMS micromirrors, micro-MEMS speakers, and ultrasonic transducers, models will not only observe the environment but also exert influence on it.
CMOS Image Sensors: Another Cornerstone of Physical AI
According to the Status of the CMOS Image Sensor Industry 2026 Report, the CMOS image sensor market reached USD$25.3 billion in 2025 and is expected to grow to USD$30.3 billion by 2031, reflecting a CAGR of 3.1%. Behind this figure is a trend: cameras based on CMOS image sensors are gradually transforming from mere recording tools into the primary means by which machines understand their surroundings.
Humanoid robots are a prime example of this trend and the field in which CMOS image sensors are most closely integrated with physical AI. Real-time object recognition, distance judgment, and reacting to people cannot rely on fixed layouts or preset programs. For humanoid robots to recognize objects, track hand movements, or perceive indoor environments, they must first capture scene information. CMOS image sensors serve as the “eyes” of this system, converting light signals into the data required for model reasoning.
Sony holds an absolutely leading position in this market, followed by Samsung, Omnivision, and STMicroelectronics. In practice, this typically involves multiple cameras distributed across the head, hands, and torso working in coordination, each with a distinct role—for example, wide-angle lenses for indoor navigation and close-range lenses for object manipulation—enabling the robot to construct a coherent overall view rather than being limited to a single narrow perspective. Humanoid robots highlight this dependency: to achieve safe grasping, navigation, and interaction, they need not only 2D RGB image information but also distance information, which opens the door to combining SPAD-based depth sensing with traditional CMOS imaging.
RF Sensing: Perception Beyond Line of Sight
In addition to connectivity for communication and cloud data offloading, RF technology provides a perception dimension that cameras and LiDAR cannot match. Radar and millimeter-wave sensing can penetrate obstacles, fog, dust, or darkness and measure object speed. This capability makes radar the workhorse technology for object tracking in industrial robots and self-driving cars, independent of lighting conditions. Ultra-wideband (UWB) technology complements this by providing positioning in indoor environments where GPS cannot reach. Even the connectivity layer itself—including Wi-Fi, Bluetooth, and 5G/6G—is increasingly being endowed with “passive sensing” capabilities, using RF links originally intended for data transmission to extract information about the presence and movement of people. The innovation lies in fusing these RF data with generative and predictive models that process visual and physical data, building a unified data representation rather than relying on independent sensor data streams.
For humanoid robots operating around humans, this fusion is particularly important. As these robots increasingly enter homes, hospitals, and workplaces, RF sensing can help them track the dynamics of people and obstacles beyond the camera’s field of view in real time, creating a critical safety layer. This technological convergence is unfolding in a rapidly growing RF market that reached USD$553 billion in 2025.
Main Functional Modules of Humanoid Robots
(Source: Status of the Humanoid Robot Industry 2026 Report)
From Sensors to Systems: The Future of Physical AI
Humanoid robots are the most visible embodiment of sensor fusion technology, but they represent only the leading edge of the field. The same underlying technology elements—MEMS sensing, CMOS imaging, and RF sensing—are being widely applied to a broader range of physical AI systems, including automated warehousing fleets, agricultural robots, surgical robots, smart infrastructure, and even consumer wearables. These end markets share a common fundamental shift: perception is no longer merely a limited detection task but is evolving into a rich, multimodal reasoning layer built on generative AI. As sensor costs continue to fall and multimodal models continue to improve, whoever masters the complete chain from perception to action will lead the next stage of physical AI development.
* Source: Compiled and translated from MEMS WeChat Account and Internet