Ask four research houses where humanoid robots are headed and you will get four different answers. ABI Research expects a $6.5 billion market by 2030. Goldman Sachs Research projects $38 billion by 2035. Morgan Stanley models a humanoid economy approaching $5 trillion by 2050. And Gartner predicts that fewer than 20 companies will have humanoid robots in production by 2028.
The forecasts differ because they measure different things over different horizons. The direction, however, is not in dispute. In our Top Tech Trends for 2026, we predicted that a new generation of large action models would drive “deployments of humanoid robots” alongside autonomous industrial systems. That prediction is now playing out, and ST is involved on both sides of it: as a supplier of the silicon humanoids are built from, and as a manufacturer putting humanoid robots to work in its own factories.
What is physical AI?
Physical AI is artificial intelligence embedded directly into machines that sense, decide, and act in the real world. Unlike large language models running in data centers, physical AI operates under hard real-time constraints: a robot must perceive its environment, plan a motion, and execute it safely within milliseconds, on a limited power budget. We explored it in our article on physical AI and the sim-to-real gap.
A humanoid robot is the most demanding expression of this idea. It must balance on two legs, manipulate objects designed for human hands, and share space with people. Allan Lagasca, who leads our Smart Industrials segment worldwide, puts the design challenge simply: “as knowledge and wisdom are different things, robots need to both understand the physical environment and apply this knowledge by acting safely and effectively.”
Why build robots in human form?
Humanoid robots make economic sense for two practical reasons. First, the world is already built for humans, so a machine with human proportions can work in existing facilities without redesigns. Second, the labor gap is a gap in human-shaped work: in the United States alone, manufacturers may need 3.8 million new workers by 2033, and about half of those openings could go unfilled, according to Deloitte and The Manufacturing Institute, and the unfilled roles, and the workstations built around them, are designed for people. A robot that shares the human form can step into that gap without redesigning the task.
There is a third reason: a humanoid can switch tasks the way a person does. A traditional robot is bought for one job; a humanoid is bought for a job description.
Where humanoid robots work today
Humanoid robots left the laboratory in 2024 and 2025, and the first industrial deployments are now documented. Agility Robotics’ Digit moves totes at a GXO logistics center in Georgia under the industry’s first multi-year robots-as-a-service contract. FigureAI’s robots completed a ten-month pilot at BMW’s Spartanburg plant, where the next-generation Figure 03 is now taking on logistics sequencing, while BMW brings Hexagon’s AEON humanoid to its Leipzig plant in Germany. At CES 2026, Hyundai Motor Group presented the production version of Boston Dynamics’ Atlas and a plan to bring it into its own factories. In China, Unitree cleared approval for a Shanghai IPO in June 2026 after reporting 5,500 humanoids sold in 2025. And NVIDIA released the first open humanoid reference design at GTC Taipei in June 2026.
Still, Gartner’s caution about stalling pilots is a useful counterweight. Our own reading is pragmatic: deployment will scale task by task and factory by factory, and the winners of the next five years will be decided by engineering economics rather than by forecasts.
Anatomy of a humanoid: how we see the machine
The public conversation focuses on AI models and processing power. The engineering reality is broader. When our teams work with humanoid developers, we decompose the robot into its functional blocks: a head unit for vision and spatial awareness, a body sensing unit for balance and motion, hand units for fine manipulation, articulated legs and arms with a motor in every joint, a main processing unit orchestrating it all, and a power management system that sets its runtime.
Perception comes first. A humanoid needs a three-dimensional understanding of a changing environment, built from image sensors, time-of-flight ranging, and inertial measurement units, fused in real time.
Balance is a sensing problem before it is a control problem. A robot on two legs is never passively stable. It stays upright the way people do: a continuous loop between inner ear and muscles. In a robot, the inner ear is an inertial sensor, and the loop must run fast enough to catch a stumble in progress. This is why the industry is putting machine-learning cores directly inside the sensors themselves: the first layer of motion recognition happens at the sensing point, in microseconds and at milliwatts, before the main processor is even involved. Balance recovery, fall detection and step planning built on high-g inertial sensing were among the demonstrations at Embedded World 2026.
Motion is multiplied by forty. A humanoid has dozens of joints, from shoulders to fingers, each needing its own driver, precise position feedback, and smooth torque control, the same discipline that runs industrial motor drives, where each joint drive now carries its own controller and runs its own control loop.
Energy is the binding constraint. A battery-powered humanoid must run perception, computing, and dozens of actuators within a power budget a fraction the size of an electric vehicle’s. Wide-bandgap semiconductors such as gallium nitride make power stages smaller, cooler, and quieter, which matters inside a machine that shares a room with people, and every watt saved becomes runtime.
Trust is built in, not bolted on
A machine that walks among people earns its place only if people trust it, and trust in a humanoid cannot be added at the end of the design. It has to be built in from the system perspective, safety and security together: a robot acts safely only if it is also secure, because compromised software or communications turn a physical machine into a physical risk. The real engineering challenge is balance, keeping the speed and agility of humanoid development, and respecting the hard constraints these systems live with in operation, energy above all, while still guaranteeing safe and secure behavior.
Regulation for humanoid safety is only starting to take shape, and the journey will be long. Our footprint in automotive and industrial, where functional safety and security are certified engineering disciplines rather than aspirations, is the experience we bring to that journey.
Our place in the humanoid stack
Seen from our side of the industry, a humanoid is a semiconductor system. We estimate around $600 of addressable semiconductor content in a typical humanoid robot, spread across sensing, control, actuation, connectivity, and power. Few suppliers ship products in every one of those domains, and that breadth, more than any single device, is what humanoid developers are assembling today.
Our strategy is to build everything below the brain. Through our collaboration with NVIDIA, announced in March 2026, our sensors, microcontrollers, and motor control integrate with the leading robotics AI platform. Our high-fidelity sensor models in NVIDIA Isaac Sim let developers train and validate robots in simulation with realistic sensor behavior before any hardware exists, and with Leopard Imaging we have packed 2D imaging, 3D depth, and motion sensing into a humanoid-ready vision module.
The same pattern continues down into the hand. Working with Ruiyan Intelligent Control, the entire control system of a dexterous robotic hand fits on a board small enough to sit in the palm: a microcontroller coordinating five joints across six degrees of freedom, high-precision analog conversion reading finger position and grip force, and motor drivers efficient enough to deliver torque without turning a closed hand into a heat problem. Ecosystem work like this is how the pieces come together, one partner and one subsystem at a time.
And we act on our own forecasts. In December 2025, Oversonic Robotics signed an agreement with ST to bring RoBee cognitive humanoid robots into production and logistics at several of our plants.
What happens next
The obstacles between today’s pilots and a real humanoid workforce are practical: cost at scale, software maturity, safety standards, data security, and human acceptance. They will be solved the way industrial problems always are: by iteration across thousands of engineering teams.
Companies bringing humanoids to scale will need partners across the full system, from sensing and motion to power and connectivity. It is a system that comes together the way it will run: sensor by sensor, joint by joint.
As humanoid robots approach cost parity with human labor, and hardware and sensor advances are improving their abilities to respond and react to evolving environments, we believe they are poised to join the workforce in significant numbers over the coming years.
— Allan Lagasca, Smart Industrials, Robotics Worldwide Leader, STMicroelectronics
Forecasts will keep disagreeing about how many humanoids arrive, and when. What is already decided is where the race will be won: below the brain, in the sensing, motion control and power efficiency that let a machine work a full shift beside people. In that shift, the winners will be the companies that master those layers at production scale. To see how we are approaching it with NVIDIA, read how ST and NVIDIA are accelerating physical AI. And for the other half of this story, our companion ST Perspectives article looks at collaborative robots, where physical AI is already earning its industrial credentials today.