Actuators, Reducers and Vision: The Component Supply Chain Behind the Humanoid Boom

The forecast is really a supply-chain forecast

When an investment bank raises a humanoid unit forecast, what it is really forecasting is component volume: harmonic and cycloidal reducers, frameless torque motors, encoders, force-torque sensors, battery packs, and depth cameras.

Goldman Sachs Research has explicitly framed South Korea's position this way — the country's auto-industry expertise in power steering and autonomous driving translates directly into a leading position in actuators, the single most cost-critical humanoid subsystem (Goldman Sachs Research, June 2026).

Meanwhile China remains the largest robot market on earth. The International Federation of Robotics reported 276,288 new industrial robot installations in China in 2023, roughly 51% of global installations, with a worldwide operational stock of 4.28 million units (IFR World Robotics).

The five components that set your price

| Component | Why it dominates cost | Direction in 2026 | |---|---|---| | Precision reducer (harmonic/cycloidal) | Highest single-part cost per joint | Falling; Chinese and Korean second sources maturing | | Frameless torque motor | Determines payload and thermal limits | Falling with EV motor volume | | Encoder (dual, per joint) | Sets repeatability and safe-motion capability | Stable; integration improving | | Force-torque sensing | Required for collaborative operation | Falling; increasingly built into the joint | | Depth camera / vision head | Determines what tasks the arm can do at all | Falling fast; software is the differentiator |

The practical consequence: the mechanical arm is commoditising, and the perception and safety software is not. A buyer paying only for the arm is buying the part that gets cheaper. A buyer paying for the perception stack, the safety validation and the service contract is buying the part that holds its value.

What "AR-marker 2D" versus true 3D actually means

A lot of integrated cameras shipped on arms today are 2D with AR-marker tracking. They locate a fiducial and infer pose. That works for fixtured, repeatable parts and fails on anything variable.

True 3D perception — stereo depth or structured light, fused with a learned segmentation model — is what lets a cell handle mixed SKUs, variable presentation and human-shared workspace. It is a separate module on a separate compute box, and it is where deployment risk lives. Ask any vendor two questions:

1. Is the depth sensor integrated in the arm or a separate module on edge compute?
2. Who owns the perception model, and can it be retrained on my parts?

Why we integrate in Hong Kong

SMERT is a systems integrator headquartered in Hong Kong, with a Delaware branch (smert.ai Inc.). We integrate cobot arms from established manufacturers with globally sourced components and our own AI and computer-vision stack. That structure exists for three reasons:

• Component access. Hong Kong sits inside the Pearl River Delta supply chain with an independent trade and customs regime.
• Certification separation. Hardware and integration are invoiced and certified as distinct scopes, which is how EU and US installations stay clean on product-liability.
• Speed. Our R&D lab in Tsim Sha Tsui builds and validates every cell configuration before it ships, so what arrives on site is a tested cell, not a box of parts.

How SMERT fits

Every SMERT Cobot ships with a normalised gripper interface, our own vision stack (2D plus a true 3D module on edge compute), and ISO 10218-compliant safe-motion validated per cell. Because we control the software layer, component cost reductions in the supply chain reach our customers instead of being absorbed as margin.

Technology stack: /technology · manufacturing applications: /robotics/manufacturing · computer vision: /ai-services/computer-vision · full lineup: /robotics.

Book a technical walkthrough.