WorldIndustrialWeekly General Industry Review: Humanoid Robot Demand Leaves STMicroelectronics Choosing Between Bigger MCUs and MPUs

Industrial automation teams evaluating humanoid robots should not begin with a single processor choice. The architecture is still unsettled, shipment volumes are low, and the decision to concentrate intelligence in stronger microcontrollers or richer microprocessor platforms is being deferred until the market gives clearer signals. For readers tracking how semiconductor vendors respond to factory, logistics and service robotics, the most useful signal is not a finished product announcement, but a staged reading of demand, latency and safety.

Allan Lagasca

Edge latency is the design constraint

STMicroelectronics’ operating principle for onboard intelligence is simple: quick local reactions should stay near the actuator, while broad coordination belongs to a central GPU. The round trip to the GPU creates delay that a moving machine cannot absorb. As Allan Lagasca, who leads STMicroelectronics’ robotics segment strategic program, told EE Times, Every millisecond counts, and he added: If you want a fast reaction, you should move the processing to the edge. If you want the entire system to work together, that is where the GPU comes in.

For two hands to transfer an object, for example, a delayed command can make a joint miss its target. Even a few millimeters per joint can distort the combined movement. Lagasca’s summary was direct: If inference is performed at the edge, the system can send the necessary command locally and respond faster, he said. The point is important for production managers because the failure mode is physical, not merely digital.

The central GPU, he noted, is a larger brain, but it still needs to synchronize head cameras, machine vision and navigation while coordinating joints whose sensors are not connected to one another. Without that coordination, he said, the robot could move both feet forward at the wrong time. STMicroelectronics does not make GPUs.

Case evidence: how a robotic hand reads a gesture

The demonstration at electronica India gives industrial readers a compact reference design. A camera captured an image of a human hand; an AI model identified 21 points to recognize the gesture; the result was converted into commands for a robotic hand with six motors. The recognition ran on the STM32N6 32-bit microcontroller.

It is not just a vision task. The mapping is not simple, Lagasca said. It is not even a one-to-one or one-to-two relationship, and it must run continuously as the hand moves.

The perception chain begins with a camera module that includes an image sensor, an inertial measurement unit to establish position and orientation, and a time-of-flight sensor to measure distance. Capturing an image is not enough; you also need to measure the distance, Lagasca said. The time-of-flight reading tells the hand when to close its fingers.

He also said inference can run mainly onboard the camera, partly for privacy. As long as an image is used for inference without being stored or sent elsewhere, the system can meet privacy and cybersecurity requirements. For the demonstration, a standard PLC bridged AI outputs and robotic control. The difficulty lies in translating the data generated by AI into a format that can be transferred to a standard PLC to control the robot, he said.

The edge architecture can also calculate force. A person holding a water bottle adjusts instinctively; a robot must compute it. A hand that identifies a bottle, measures distance and adjusts grip to avoid crushing or slipping uses local processing. The MCU, in this case the STM32, performs this inference, Lagasca said.

Distributed components replace a single-processor fantasy

Lagasca said most sensors have their own MCUs, as do motion and motor-control functions, while AI inference has its own NPU. This creates a distributed architecture rather than putting everything into a single processor, he said. His current AI inference products are based on STMicroelectronics’ neural processing unit and microprocessor unit. Whether the company develops more powerful chips depends on how the robotics market develops.

Module integration turns third-party supply chains into engineering constraints

Because humanoid robot manufacturers may source hands, joints and cameras from different suppliers, integration creates problems with camera placement, board space and motor control. A camera mounted on top of a hand may require the robot to reposition its hand before lifting an object; a camera mounted underneath may strike the table during pickup.

STMicroelectronics said it is exploring ways to integrate camera technology and other functions into the humanoid hand itself. The company does not manufacture camera modules, but it can combine its technologies with partner technologies. Earlier this year, STMicroelectronics and camera-module maker Leopard Imaging introduced a multimodal vision module for humanoid robots that combines STMicroelectronics’ imaging technology, 3D depth-mapping and motion sensors. The module connects over Ethernet through Nvidia’s Holoscan Sensor Bridge to Nvidia’s Jetson computers and Isaac robot development platform, designed to fit size, weight and power limits.

The broader approach is to divide the robot into modules such as hands, joints, feet and eyes, then adapt those modules to different applications and form factors. Packaging is difficult because designers must fit processing, motor control and connectivity into available board space. A customer asked for a 6 × 6-centimeter board. Lagasca said choosing MCUs, motor drivers and connectivity parts separately could produce a design measuring 140 × 330 mm, far larger than the board. To reduce that gap, ST integrated six MOSFETs and three drivers into one package, so a manufacturer can mount one part instead of nine.

Functional safety does not follow automatically from certificates

Distributed intelligence also spreads safety responsibility. Lagasca said existing frameworks include Safety Integrity Level requirements for industrial applications and Automotive Safety Integrity Level requirements in automotive. These requirements may matter as companies from automotive, industrial automation, smartphone and home appliance markets enter robotics. Compliance with a standard does not guarantee that the entire robot is safe.

If a sensor fails, the robot could still cause an injury, even if it meets the relevant SIL or ASIL standard, he said. AI can introduce additional risks if a system fails to make a decision or makes the wrong one. All elements must be considered together to assess functional safety.

Lagasca added that many existing standards were developed for fixed equipment rather than moving robots, and the industry still needs to determine which standards and regulatory requirements are appropriate for humanoid systems, especially those that interact directly with people.

Data citation: reading a humanoid robot bill of materials

Lagasca said STMicroelectronics can address approximately 600 positions in a humanoid robot’s bill of materials, covering more than 500 components. He added that the share of total robot cost represented by those parts depends on design, including how many joints, sensors and AI functions the robot has.

For illustration, he said that if a robot’s BOM were worth $600 and STMicroelectronics supplied all of it, the chipmaker’s share would be 100%. If the BOM grew to $1,000 while the chipmaker’s parts remained valued at $600, its share would fall to 60%.

Entry-level robots may cost less and have fewer degrees of freedom, while mid-range and high-end systems may need more sophisticated capabilities. Lagasca said that, depending on design, STMicroelectronics could potentially address 800 or more positions as the number of joints, sensors, cameras and AI-processing requirements increases.

From proof of concept to industrial deployment

Many applications remain at proof-of-concept stage despite growing demonstrations. Automotive manufacturing and logistics are early areas because manufacturers can justify automation cost and adjust production lines to meet market demand.

He cited links between humanoid robotics and automotive: Boston Dynamics’ association with Hyundai and Figure AI’s relationship with BMW. Hyundai, which owns Boston Dynamics, plans to start deploying Atlas robots at its Georgia plant in 2028, while BMW has conducted trials for Figure robots in its Spartanburg, S.C., body shop.

Moving beyond demonstrations requires several practical fixes. Training data is one. Lagasca said activities such as drilling, soldering, cooking and sewing can be recorded and uploaded to create datasets. Battery technology does not yet support 24-hour operation, making downtime a practical consideration. Applications also depend on safety and regulatory requirements, especially where direct physical interaction occurs in healthcare.

Cybersecurity is another concern. Chipmakers and device manufacturers must account for the EU Cyber Resilience Act and other applicable requirements at product and application levels. You do not want 10,000 robots operating in the real world only for someone to hack into them, Lagasca said.

Common mistakes

• Assuming a central GPU should handle fast reactions; local edge processing can avoid the latency of sending data out and returning commands.
• Believing image capture alone gives sufficient robot perception; distance measurement is needed to close fingers at the right moment.
• Planning a robot as if all functions can be forced into one processor; sensors, motion control and AI inference are distributed across MCUs and an NPU.
• Treating SIL or ASIL compliance as proof of whole-robot safety, especially when sensors fail or AI decisions are wrong.
• Choosing between bigger MCUs or MPUs before the humanoid market volume and architecture have clarified; the answer depends on how robotics develops.
• Expecting 24-hour operation without downtime; current battery technology does not support it.

Action recommendations

1. Map each robot function by latency class: immediate reaction at the actuator or camera, whole-body coordination at central processing.
2. Validate the perception chain with image sensor, IMU and time-of-flight data, not just gesture recognition.
3. Test data translation from AI models to standard PLCs early, because the format bridge is often the hard integration problem.
4. Score third-party modules for placement, board area and connector count, then consider packaged integration that reduces mounting items.
5. Build a bill-of-materials baseline using component counts and cost-share examples, then revisit the share when joints, sensors and AI functions increase.
6. Document functional safety as a system property that includes sensors, AI failure modes and human interaction.
7. Check EU Cyber Resilience Act exposure at product and application levels before pilot scale-up.

FAQ

Why has STMicroelectronics not selected a bigger MCU or MPU path?

Robotics remains low-volume while STMicroelectronics targets high-volume production as an integrated device manufacturer. By following this market, we will better understand which technologies are needed, Lagasca said.

What does the company think about the final humanoid robot shape?

The system architecture has not stabilized, Lagasca said. Some developers ask why a robot should have only two hands rather than six, or only two legs. We do not yet know what the final configuration will be, he said.

Where does the current inference product line stand?

Lagasca said current AI inference products are based on STMicroelectronics’ NPU and MPU; more powerful chips will depend on market evolution. The STM32N6 example shows the company can run gesture inference on a 32-bit microcontroller, while the broader architecture remains edge-centric.

For WorldIndustrialWeekly readers, the conclusion is procedural rather than speculative: track where local inference removes latency, where distributed components create integration work, and which safety and cost questions remain unresolved before a humanoid robot moves from a demonstration into an industrial workload.