How AI is expanding the tasks that fixed robots can perform—and the supply chain that turns an arm into a working cell.
AS OF 2026-09-02DATED · NON-EXHAUSTIVE
Supply badges: “not assessed” means this report has not established a node-wide supply rating. It does not mean supply is plentiful; product maturity and local constraints are discussed separately.
Family 1
Industrial manipulators & robot cells
Industrial arms and cobots become robot cells only after task tooling, guarding, controls and workflow engineering adapt them to a production application. The arm can cross the market as equipment, or an integrator can carry application risk through acceptance of a commissioned cell.
A familiar machine, a widening set of tasks
A robot arm welds a car body, loads a machine tool, picks a part from a bin or places an item into a box. This family covers the arms and the working cells built around them: a bounded workstation with the tools, controls and surrounding equipment needed to do a job. Articulated arms, fast pick-and-place robots and collaborative robots belong together here because their enabling chains overlap. A warehouse picking arm fits this chain; an autonomous vehicle carrying its bins belongs in the mobile-robot family.
The machinery is already widely deployed. The International Federation of Robotics counted roughly 4.66 million industrial robots operating worldwide at the end of 2024. That is the installed industrial base, not a count of AI-powered robots. Much of it performs carefully programmed tasks in deliberately arranged environments. IFR, World Robotics 2025 →
The opportunity is to make more variation economical. An arm can repeat a movement very well; choosing what to pick, locating a displaced part and recovering from a failed grasp are different problems. Better perception and adaptive software can reduce how much an engineer must specify afresh for each part or job. That could extend automation into smaller batches, changing product mixes and tasks that are difficult to fixture consistently. It is the central expectation behind this family's place in Physical AI.
Following the supply chain means distinguishing the robot from its application. A manufacturer sells an arm and controller. A machine builder, integrator or customer then combines it with tooling, sensing, process equipment and safeguards to create the working cell. ISO's separate standards for the robot and the integrated application reflect that distinction. Robot product → · Application and cell →
What AI changes—and what still needs engineering
A useful first step is often a narrower skill: recognize a part under changing lighting, estimate its position, choose a grasp or plan a collision-free path. A full general-purpose robot policy is not required for every improvement. These skills can be combined with conventional motion control and the machine information a factory already has.
A concrete example · sheet-metal sorting
How adaptive sorting works
In TRUMPF's SortMaster system, a mechanical station first separates laser-cut parts from the surrounding sheet. Vision then locates the separated pieces, while software uses the cutting program to identify them and calculate how to grip and move them. The robot, gripper, camera, software and handling equipment all contribute. System description, October 2024 →
TRUMPF offers the system for sorting changing part mixes without manually programming each sorting sequence. This is a concrete way AI can reduce repeat engineering. The product page establishes what is offered; it does not supply independent fleet-wide uptime or customer payback measurements. Current product offering →
Someone must choose tooling, calibrate the camera and robot, connect machine controls, handle failed picks, define safe stops and show that output meets the customer's requirements. NIST's workcell study documents the interacting robot controllers, industrial controls and measurement systems; Pilz's application-safety services illustrate the continuing safety-engineering role.
These examples ground the opportunity in bounded skills and applications. Transferable manipulation across unfamiliar jobs is the larger possibility. Its economic test is whether the next deployment needs less engineering and fewer interventions while preserving quality and throughput.
The enabling supply chain
From a robot to a working cell
Follow the arrows into the two combinations: building the robot, then integrating its application. AI-related inputs enter where the task uses them. Circles open the industrial-specific explanation below.
Product or componentEngineering or productionSupplied outputSupply / processApplication-dependentDevelopment supportMaintenance
Each bubble opens its focused read
A causal map, not a mandatory bill of materials. Software and sensing can be bundled, and the same company can perform several activities. The blue rings distinguish arm equipment from the installed application; they do not require a sale between every step. Product qualification is included in manufacture and test. Contract and ownership choices are discussed below rather than drawn as extra physical stages.
Key players are selected examples, linked to their relevant products or work—not a market-share ranking. Some activities are performed inside a manufacturer or operator rather than sold separately. “(private)” marks unlisted businesses; listed parent groups are identified separately.
Building the robot
The arm's reach, payload, precision and duty cycle emerge from its mechanics and control. These inputs predate today's AI systems and remain necessary when the application becomes more adaptive.
Economic nodes
Precision motion
Supply not assessed
Motors and drives create movement; reduction gears trade motor speed for useful joint torque; bearings, encoders and brakes support precision and control. They can arrive as separate components or integrated actuators. Different joints need different torque, stiffness and weight, so a humanoid hand and a large welding arm do not imply the same component market.
Links, housings, base, cable routing and mounting interfaces carry the loads and connect the joints. Geometry and stiffness constrain where the tool can reach and how it behaves under load. These are robot-specific manufactured parts; the map stops here rather than extending into all metals and general machining.
Key players
FANUC — arm bodies engineered as part of its robot products.
Yaskawa — Motoman arm structures across payload and reach classes.
Worth understanding
The manufacturer's mechanical design combines these parts with its selected motion components. The physical architecture is visible in ABB's account of its industrial robot designs.
The controller coordinates joint motion, executes trajectories and exposes the interfaces through which the application gives commands. Robot-level safety functions supervise limits and stopping. A higher-level AI system may select an action without replacing this controller.
ABB OmniCore illustrates the supplied controller platform. Cell-level safeguards are mapped separately because the complete application introduces hazards the bare arm cannot resolve.
Mechanical, electrical and control design turn the components into an arm/controller combination. Engineers select the joint arrangement, model loads, manage cables and thermal limits, and define operating specifications. This remains a distinct value-adding activity when the robot manufacturer performs it internally.
Key players
FANUC — internal arm, drive and controller engineering.
Yaskawa — internal robot and controller integration.
Worth understanding
ISO 10218-1:2025 addresses the industrial robot product. It helps distinguish product-level engineering from the downstream integration of a specific job.
Assembly converts the design into equipment. Calibration relates commanded positions to the physical robot; performance tests and product-safety work establish what can be delivered. Qualification also influences design—it is not a standalone material through which every finished arm passes.
Yaskawa — production of Motoman industrial robots.
Worth understanding
NIST's robot-accuracy work shows why calibration and later degradation measurements matter. Product requirements sit with ISO 10218-1; neither substitutes for testing the installed application.
This is what the robot manufacturer can sell before a particular production cell exists. An integrator selects equipment for payload, reach, speed, precision, interfaces and environment. A collaborative arm changes the integration possibilities; its name does not make every attached tool or application safe.
Universal Robots (private) · Teradyne — cobot arms and their control platform.
Worth understanding
Universal Robots' integration interfaces show how equipment connects to end effectors, vision, industrial controls and safety systems. The arm is a commercial product and also a downstream input.
These are capabilities for the applications that use them, not a compulsory AI stack for every arm. A supplier may bundle camera, compute and software; separating their roles helps explain what must work.
Economic nodes
Machine vision
Supply not assessed
Cameras, lighting, depth sensing and image-processing software locate parts and provide information about the scene. Rule-based vision and learned perception coexist. Camera-to-robot calibration and the actual objects, glare and occlusion determine whether the information is useful for a grasp.
Universal Robots (private) · Teradyne — AI Accelerator vision-development kit.
Worth understanding
The UR AI Accelerator includes an Orbbec 3D camera; ABB's Item Picker bundles vision with a robot and suction grippers. Product packaging can cross several map nodes.
Industrial computers or accelerator modules run perception and other AI workloads close to the application. Their useful contribution depends on latency, software compatibility, thermal design and reliable interfaces to the robot. A camera may include its own inference hardware; not every cell needs a separate compute box.
Universal Robots (private) · Teradyne — AI Accelerator computing kit for robot applications.
Worth understanding
UR's toolkit includes a Jetson Orin box. Siemens' Robot Pick AI specification also supports industrial PCs without a dedicated GPU. Compute is sized to the workload; a separate AI accelerator is not universal.
Perception models, grasp selection, motion planning and task logic turn observations into useful actions. Some functions are learned; others use geometry, constraints or conventional programming. They may be licensed as a platform, embedded in a machine product or developed internally.
Key players
Intrinsic · Google / Alphabet — Flowstate application and robot software.
Skild AI (private) — robot intelligence applied to industrial work.
Worth understanding
Intrinsic's FANUC integration shows a software platform working with industrial hardware. Its value depends on the skills and interfaces it makes reusable, and on how failures are detected and handled.
Broader policies are also being pursued. Skild's March 2026 announcement described robot-manufacturer partnerships and planned Foxconn assembly deployment. It shows the intended direction; the announcement's future-tense shipping plan is not evidence of sustained production performance.
Images, object geometry, demonstrations and success/failure records help develop and evaluate learned capabilities. The relevant examples must cover the intended parts and conditions. This can be work inside a software supplier or manufacturer; an independent merchant data market is not required for the activity to matter.
A digital representation of the robot and cell helps engineers check reach, interference, sequencing and programs before occupying real equipment. Simulation can also generate training scenes. The model's fidelity limits what the result establishes; contact, calibration and real production variation still need physical checks.
Key players
ABB — RobotStudio cell simulation and offline programming.
Intrinsic · Google / Alphabet — application development and simulation with integrators.
Worth understanding
ABB RobotStudio is a supplied simulation/programming suite. Intrinsic and Trinity's machine-tending work uses a digital-twin development environment. The dashed connections show tools used in development, not another mandatory runtime stage.
A customer needs reliable work on its parts, at its required rate. This is where the general equipment and software meet the process, the site and the people operating it.
Economic nodes
Tooling & fixtures
Supply not assessed
Grippers, suction tools, weld guns, tool changers and fixtures determine how the robot engages the work. Part presentation and fixturing can simplify a difficult perception or manipulation problem. Conversely, better perception can sometimes reduce the need to hold every part in an identical pose.
ATI (private) · Novanta subsidiary — robotic tool changers and tool interfaces.
Worth understanding
Robotiq's grippers illustrate the separate tool market. The right tool depends on the part, force, finish and cycle—not on the robot being labelled AI-powered.
Industrial controls coordinate the robot with surrounding machines. Safety controllers, interlocks, protective sensors and guards implement the application's protective measures. Learned object recognition is not automatically a safety-rated protective function.
Key players
Pilz (private) — robot-cell safety controls and safeguarding.
KUKA (private) · Midea subsidiary — cell safety planning and integration services.
Worth understanding
Pilz supplies safety sensing, controls and application services. The correct configuration depends on the whole task, including the tool and workpiece; this is distinct from the robot controller's own motion functions.
Integration combines the arm, tool, work handling, sensing, software and controls. It includes layout, interfaces, programming, calibration and exception handling. Existing process equipment enters here; the map does not attempt to redraw every machine tool or conveyor's upstream chain.
Key players
KUKA Systems (private) · KUKA / Midea — production-system engineering and integration.
Engineers install the cell, test its actual tasks, validate protective measures, train users and resolve acceptance issues. A successful pick in a demonstration does not establish sustained output across the customer's part mix. Revisions to the application can require renewed engineering and validation.
Key players
KUKA (private) · Midea subsidiary — installation, programming and functional start-up.
ISO 10218-2 includes application integration and commissioning. It is a safety reference, not a promise of production performance; commercial acceptance also depends on the agreed job and contract.
The useful output is an integrated application that performs an agreed task: a tending cell, sorting station, welding cell or assembly system. It can contain one arm or several. Its quality, cycle time, availability and need for intervention determine the productive capacity delivered.
Key players
TRUMPF (private) — integrated sorting and material-handling systems.
KUKA Systems (private) · KUKA / Midea — integrated production systems.
Worth understanding
SortMaster Vision is one concrete application offering. Its task-specific design explains why a working cell is economically different from the robot arm alone.
Maintenance, spares, repair, recalibration, software support and upgrades preserve useful operation. Mechanical wear and changing applications create work after installation. Some service is performed by the owner; some is supplied by the manufacturer, integrator or specialist.
Key players
ABB — robot maintenance, upgrades and refurbishment.
KUKA (private) · Midea subsidiary — on-site repair and servicing.
Worth understanding
ABB's service offering spans support for installed robots. This is a continuing input to uptime, not a manufacturing step after the cell is finished.
The economic question is repeatability across jobs
The first cell pays for engineering. The next part, job or site can require new tooling, programming and validation. If perception and planning become more reusable, some of that repeated work can fall. More applications may then justify the equipment and integration cost, including smaller production runs.
That is a causal expectation, not a measured industry-wide cost decline. The useful evidence would be less engineering per changeover, more parts handled by the same configuration, fewer interventions and acceptable output over sustained operation. A fast demonstration or a large robot installed base cannot settle those questions.
Reuse already has specific limits. Siemens sells Robot Pick AI as a per-instance software licence, with specified cameras and a single-suction-cup gripper configuration. Reuse across a range of items is valuable, but it is not the same as transferring the skill to any tool or task.
A reported deployment also shows why recovery matters. Inbolt's Stellantis Trenton case study, checked September 2026, reports bolt-tool non-engagement falling from 2.55% to 0.24% with vision guidance. The same account says an operator remains at the end of the line. It is a vendor-reported improvement in a defined failure mode, not elimination of operating labour.
Longer operating hours help only when they produce usable output. Replenishment, tool changes, maintenance, quality failures and recovery still matter. NIST's assembly-performance study explains why measuring the complete task is necessary rather than relying on the arm's specifications.
Who sells what?
The component maker supplies motion or sensing. The robot manufacturer supplies equipment. The software vendor supplies capabilities and tools. The machine builder or integrator supplies an application. These roles may sit inside one company, and ownership can remain with a provider rather than pass to the user. A rental or managed-use contract changes who finances and supports the cell; it does not eliminate the engineering chain.
For now, the strongest grounding is the growing availability of bounded adaptive products and developer tools. The larger prospect is a wider range of profitable robot tasks. The unresolved issue is how consistently those capabilities transfer into customer operations without rebuilding the solution each time.