Robotics · Pillar guide · 2026-08-22

The robotics stack: what has to scale before physical AI can leave the lab

A useful robot is not a model with legs. It is a six-layer industrial system that has to secure materials, make precise motion, perceive contact, control the body, pass production test and earn its keep in the field.

Dylan Austin Bristot, Founder and publisher, AI Bottlenecks

A useful robot is not a language model with legs. The model matters, but it sits on top of a physical system that has to move weight, survive contact, reject heat, measure the world, pass a safety case and earn back its installed cost.

That system is easy to miss because demos compress the stack into one visible machine. The robot walks, so the control problem looks solved. The hand lifts an object, so dexterity looks solved. A polished shell hides the suppliers, test benches, calibration routines and field service that made the moment possible.

The better way to read physical AI is layer by layer. Six layers capture most of the industrial work between raw material and productive fleet.

Layer 1: materials and power

Every robot begins with mass and energy. Permanent magnets turn current into compact torque. Copper fills motor windings and carries power. Electrical steel shapes magnetic flux. Aluminum, steel and engineered polymers form the structure and housings. Batteries, connectors and power electronics decide how much work can happen away from a cable.

The bottleneck is not simply access to material. It is material with the right magnetic, thermal, structural and manufacturing properties at a cost that survives volume production.

High-performance permanent magnets are the obvious strategic input. They support strong torque density in compact motors, but expose the stack to rare-earth processing concentration and price volatility. Copper is less exotic but hard to avoid. More torque, higher current and more motors pull it into windings, busbars, cables and chargers. Thermal materials matter because every watt lost inside a compact joint has to leave through a small surface.

What proves the layer is scaling

• Qualified second sources for magnets and critical alloys.
• Motor torque density improves without an unacceptable rise in temperature.
• Battery runtime holds under real duty cycles rather than scripted demos.
• Material cost falls while scrap and magnetic performance remain controlled.

What breaks the thesis

A materials basket can look attractive while the robot architecture moves in a different direction. Smaller robots may require less absolute material. Tethered or fixed systems reduce battery content. Motor designs can trade magnet content for copper, control complexity or lower peak torque. The exposure has to be traced through a real design.

Layer 2: precision motion

This is the layer that turns stored electrical energy into controlled force. It includes frameless motors, reducers, screws, bearings, encoders, brakes, servo drives, housings and the integrated actuator.

Precision motion is likely to attract attention first because it is tactile. A good joint makes a robot look competent. A poor joint makes it look nervous, weak or unsafe.

The technical problem is a bundle of conflicts: torque against weight, stiffness against compliance, precision against cost, power against heat, and service life against compact packaging. The production problem is harder. The factory has to make every module behave close enough to the last one that control software does not become a bespoke rescue operation.

The actuator flagship takes this layer apart and shows why a motor is only one part of the joint.

Public-company map

Harmonic Drive Systems and Nabtesco sit in different precision-reducer niches. Kollmorgen supplies frameless motors for integrated joints. Schaeffler is assembling a broader motion stack and has begun disclosing early humanoid orders.

This is not one basket with one sensitivity. Industrial-robot recovery can help reducer and motion demand before humanoids matter. An OEM decision to integrate actuators can help one supplier and commoditize another.

What proves the layer is scaling

• Production awards replace development agreements.
• Qualified actuator output grows faster than test and rework hours.
• Continuous torque and thermal performance hold in long shifts.
• Field failures, backlash drift and lubricant issues stay inside warranty assumptions.

Layer 3: sensing and dexterity

A robot that cannot measure contact is a moving liability. This layer includes cameras, depth sensing, lidar, encoders, force-torque sensors, tactile arrays, proximity sensing and the mechanical hand that turns perception into useful contact.

Vision answers where an object appears to be. Force and tactile sensing answer what is happening now that the robot has touched it. The distinction matters. A rigid factory cell can rely on fixtures and repeatability. A mobile robot in a changing workplace has to detect slippage, unexpected weight, soft surfaces and human contact.

Dexterity is not a hand-shaped shell. It is a control loop connecting sensors, mechanics and software at useful speed. More sensors can improve observability but add wiring, calibration, compute load and new failure modes. The valuable architecture may be the one that reaches sufficient manipulation performance with fewer fragile inputs.

What proves the layer is scaling

• Manipulation succeeds on unfamiliar objects, not only a fixed demo set.
• Tactile and force sensors retain calibration after repeated impacts.
• Sensor cost, wiring burden and failure rate fall together.
• The hand can be serviced without replacing an uneconomic share of the arm.

Layer 4: compute and control

The robot has several clocks. A perception model may run at video rate. Motion control loops run much faster. Safety responses cannot wait for a cloud round trip. This forces compute to split across the body, an edge controller and, in some architectures, a remote training or planning system.

The stack includes accelerators, microcontrollers, networking, real-time operating software, servo control, power management and the higher-level models that plan actions. The constraint is not raw model capability alone. It is deterministic behavior under latency, power and safety limits.

The hardware winners may differ by layer. A large accelerator can handle perception and planning while dedicated controllers close joint loops. More capable models can increase compute demand, but better policies may reduce wasted motion and lower energy use. The economic question is how much silicon and bandwidth are needed per productive hour.

What proves the layer is scaling

• On-device latency stays within control and safety budgets.
• Compute power does not erase battery or thermal targets.
• Software can transfer across hardware revisions without a full requalification.
• Fleet learning improves performance without creating uncontrolled behavior changes.

Layer 5: test and integration

This is the least glamorous layer and one of the most important. Components become a robot through assembly, calibration, end-of-line test, traceability and system-level validation.

The first machine can be tuned by the engineering team. The thousandth has to pass through a factory. Every actuator needs characterized torque and friction. Every encoder needs an alignment reference. Every camera needs extrinsic calibration to the body. The completed robot needs joint zeroing, safety checks, burn-in and a digital record of the parts and software inside it.

The calibration deep dive explains why accepted output can lag component capacity.

The public-market exposure is less clean than the motion layer. Machine-vision vendors, metrology suppliers, automation firms and test-equipment makers can benefit, but the most valuable integration knowledge may remain inside OEMs and contract manufacturers. That makes factory evidence more important than thematic labels.

What proves the layer is scaling

• First-pass yield rises as output grows.
• Calibration and end-of-line test takt fall without hiding defects.
• Traceability connects field failures to a specific part, process and software version.
• Contract manufacturers or integrators win repeat programs, not one-off pilot work.

Layer 6: deployment economics

A robot is not commercially useful because it can perform a task. It is useful when the task can be performed safely, often enough and cheaply enough to beat the alternative.

Deployment economics include installation, workflow redesign, training, charging, supervision, maintenance, spares, insurance, safety validation and the cost of idle time. A machine priced attractively at the factory can still fail the customer if it needs constant human rescue.

This is where the humanoid story meets the much older lessons of industrial automation. Structured environments, clear tasks and expensive labor create the easiest early markets. Unstructured consumer environments may be technically exciting but economically unforgiving. The first large fleets are likely to appear where the operator can measure utilization and redesign the work around the machine.

What proves the layer is scaling

• Customers expand after a paid pilot.
• Productive utilization rises without a matching rise in remote supervision.
• Mean time between service events supports the promised payback period.
• Installation and workflow-integration time fall with each deployment.
• The buyer reports an operating metric, not only a press release.

The constraint migrates

The six layers will not tighten at the same time.

Early in a product cycle, motion hardware and sensing can dominate because the robot simply has to work. During the factory ramp, calibration, end-of-line test and supplier yield can become binding. In deployment, uptime, supervision and task economics can overwhelm every component improvement that came before.

This migration is the core of the Robotics Bottlenecks framework. A part can remain strategically important while no longer being the marginal constraint. A supplier can grow units while losing economic leverage. A software improvement can relax one hardware requirement and tighten another.

The map has to move with the evidence.

How to read the public basket

The Robotics universe is best treated as a set of layer exposures rather than a claim that every covered company is a pure humanoid winner.

Some names already sell into industrial robots, machine tools, warehouses or automotive automation. That base business matters because it funds capacity and establishes qualification history. Some names are coverage positions whose role is to reveal demand or architecture, not to carry portfolio weight. Some are optionality on a production system that may take years to become material.

Five questions keep the map honest:

1. Which physical or operational constraint does the company own?
2. Is demand visible in orders and revenue, or only in prototypes and partnerships?
3. Does the robot architecture expand or remove its content?
4. Can production scale at stable yield and price?
5. What specific evidence would make the thesis weaker?

That is a better starting point than asking which stock is the next robotics winner.

The status of the stack

Robotics in 2026 is moving from technical possibility toward industrialization, but the receipts are uneven.

Motion incumbents are disclosing sample orders, supply agreements and, in rare cases, a serial order. Robot OEMs are announcing factory deployments and development partnerships. Industrial-robot demand provides real revenue underneath the newer humanoid narrative. What remains scarce is broad evidence of profitable, high-utilization fleets and repeatable high-volume production.

The next phase will be less cinematic. Watch production awards, factory yield, calibration time, field uptime, service cost and customer expansion. Those measures tell us whether physical AI has left the lab.

Sources and receipts

• Harmonic Drive Systems product list, accessed August 22, 2026. Primary source.
• Nabtesco FY2026 financial reports, accessed August 22, 2026. Primary source.
• Kollmorgen frameless motors, accessed August 22, 2026. Primary source.
• Humanoids at Schaeffler, February 5, 2026. Primary source.
• Schaeffler and Humanoid technology partnership, January 13, 2026. Primary source.

This research is educational and is not investment advice.