STMicroelectronics is shifting how robots think by distributing artificial intelligence processing across multiple specialized chips in a robot's limbs and body, rather than relying on a single central processor. The semiconductor company argues that this approach, called distributed architecture, enables robots to react instantly like human reflexes, making it possible for a robot hand to grip an object with the right force without waiting for commands from a central brain.

Why Can't Robots Just Use One Central AI Chip?

The challenge with humanoid robots is that they must balance on two legs, manipulate objects designed for human hands, and interact in real time alongside people. This requires split-second decisions that a single central processor cannot make fast enough. When a robot hand detects an object slipping, it needs to grip harder immediately, not send that information to a main processor, wait for a decision, and then execute a command. That delay creates complexity and reduces precision.

"The biggest advantage of a distributed structure is that it can react instantly like the human autonomic nervous system, and it can also distribute AI compute loads across multiple points. When picking grapes, squeezing too hard bursts them and squeezing too lightly drops them; having the MCU in the hand make the force-judgment immediately is far more efficient than sending it to a central processor or the cloud," explained Park Jun-sik, head of STMicroelectronics Korea.

Park Jun-sik, Head of STMicroelectronics Korea

STMicroelectronics is positioning itself as a strategic partner for robot developers by offering a broad portfolio of components beyond just the AI processor. The company provides microcontroller units (MCUs), which are smaller processors that handle specific tasks; power devices that drive motors and actuators; sensors that detect the environment; and connectivity solutions that allow different parts of the robot to communicate.

What Components Make Up a Robot's Nervous System?

Rather than thinking of a robot as having a single brain, STMicroelectronics divides robots into zones, with each zone handling its own sensory data and decision-making. This zonal architecture mirrors how the human body works, with local reflexes happening at the spinal cord level rather than waiting for the brain to process every sensation.

• Microcontroller Units (MCUs): Smaller processors placed in each zone of the robot that handle local sensor data and make immediate decisions without waiting for a central processor.
• Power Devices and Drivers: Components that deliver precise electrical power to motors and actuators, enabling smooth and controlled movement of joints and limbs.
• Sensors: Devices like time-of-flight (ToF) sensors and lidar that complement cameras by detecting properties cameras alone cannot determine, such as whether an object is glass or plastic.
• Connectivity Solutions: Systems that allow different zones of the robot to communicate with each other and share information when needed.

STMicroelectronics has introduced the STM32N6 microcontroller, which integrates the company's own Neural Processing Unit (NPU) called the Neural-ART Accelerator. This chip delivers up to 600 giga-operations per second (GOPS) of artificial intelligence performance, meaning it can execute 600 billion mathematical operations per second, enough to handle vision and gesture recognition directly on the chip without sending data to the cloud.

How Does On-Device AI Processing Benefit Robot Developers?

Processing artificial intelligence workloads locally on the robot, rather than in the cloud, addresses two critical concerns: speed and security. Cloud-based decision-making introduces latency, or delay, which can be problematic for robots that need to react instantly. Additionally, sending sensitive sensor data to cloud servers raises security concerns, as that information could be intercepted or misused.

"I believe most simple execution functions should be handled by local intelligence. Due to security concerns, cloud-based decision-making can have side effects, and even processing in the main AI brain entails latency and high energy consumption. For fast responses, I think distributed intelligence is the system closest to and most efficient like the human nervous system," stated Park Jun-sik.

Park Jun-sik, Head of STMicroelectronics Korea

The company is also collaborating with Nvidia to provide simulation models and demo boards that help robot developers get started quickly. Rather than sourcing individual components from multiple vendors, robot makers can use STMicroelectronics' turnkey solutions, which bundle hardware, software, and design guidelines together. This approach accelerates development speed and simplifies system integration.

When Will Robots Become More Affordable?

STMicroelectronics acknowledges that its broad portfolio approach currently prioritizes speed and integration over price. However, the company expects price competitiveness to improve naturally within one to two years as the humanoid robot market expands and production volumes increase. At this early stage of the market, robot developers value faster development cycles and easier system integration more than the lowest possible component cost.

The company is particularly focused on partnerships with Korean corporations, which have developed strong expertise in core robot modules such as actuators and grippers. STMicroelectronics is discussing technical cooperation, joint design, and joint marketing arrangements with these Korean firms to accelerate the development of humanoid robots tailored to the Korean market.

Steps to Building a Robot With Distributed Neural Processing

• Divide the Robot Into Zones: Segment the robot into functional areas such as head, arms, and legs, with each zone assigned its own microcontroller unit to process local sensor data independently.
• Integrate Specialized Sensors: Combine cameras with complementary sensors like time-of-flight or lidar to give each zone a complete understanding of its environment without relying solely on vision.
• Use Reference Designs and Guidelines: Leverage STMicroelectronics' demo boards and reference designs to accelerate development and ensure compatibility across different zones of the robot.
• Implement On-Device AI Processing: Deploy neural processing units in each zone to handle vision, gesture recognition, and force control locally, eliminating the need to send data to the cloud.
• Establish Inter-Zone Communication: Set up connectivity solutions that allow different zones to share information when necessary, creating a cohesive system while maintaining local autonomy.

The shift toward distributed neural processing represents a fundamental change in how the robotics industry approaches artificial intelligence. Rather than treating the robot as a single entity controlled by one powerful processor, developers are learning to think of robots as systems with multiple intelligent zones, each capable of making decisions independently. This approach mirrors biological systems more closely and enables the kind of real-time responsiveness that humanoid robots need to operate safely and effectively alongside humans.