Machines that move and manipulate: how bodies, learned skills and production engineering become useful physical work.
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 3 · Humanoids & mobile manipulators
Humanoids & mobile manipulators
Robots that move through a workplace and handle objects with arms or hands. Their promise is to use spaces and tools already built for people; their distinctive supply chain combines many joints, whole-body control, dexterity and task learning.
A robot that can move to the work
An industrial arm usually stays at its workstation. A mobile manipulator carries its arms to the task: lift a tote from a cart, reach a shelf, load a machine or bring a part to a line. A humanoid is one way to package that capability around a human-like body. Legs may help in spaces designed for people, but wheels can be a better choice on a factory floor. This family follows the combination of mobility and manipulation, not a requirement to reproduce every feature of a person.
The attraction is reuse. A machine that can use existing aisles, containers and tools could perform several jobs without rebuilding the workplace around each one. That is a larger ambition than moving reliably along a route. The robot must coordinate its body while making contact with objects whose position, weight or behaviour can vary.
There is now evidence beyond demonstrations, although the jobs remain bounded. In June 2024, GXO announced a multi-year agreement to use Agility's Digit at a SPANX warehouse after a 2023 pilot. Digit transfers totes from mobile robots onto conveyors; the accompanying Arc software connects it to the workflow. The commercial deployment establishes a useful task, not a general replacement for warehouse labour. GXO's deployment announcement →
BMW's March 2026 account describes both its earlier Figure work in Spartanburg and a pilot with the wheeled AEON in Leipzig. Those different bodies belong in the same initial family because both must combine manipulation, movement and plant integration. BMW's account →
More reusable skills, with a harder physical integration problem
Language and vision models can help interpret instructions and unfamiliar objects. Robot policies translate observations into actions; training uses demonstrations, simulated experience and trials on the machine. NVIDIA's GR00T work makes pretrained models and development tools available beyond one robot producer. Its research results are evidence of a development capability, not a measured reduction in every customer's deployment cost. GR00T development workflow →
Why the body matters
Picking something up changes the rest of the robot
Reach for a load and the robot's balance changes. Grip too weakly and the load slips; grip too hard and it is damaged. An arm, hand and mobile base cannot always be controlled as independent machines. Figure's January 2026 Helix 02 demonstration combines whole-body control with vision and tactile sensing in a continuous kitchen task. It illustrates the technical direction, while remaining a vendor demonstration rather than evidence of unattended household operation. Figure's system description →
The same coupling extends into manufacturing. Joint torque, transmission losses, sensor calibration, hand durability and battery heat affect what a learned policy can safely ask the machine to do. Better software increases the value of a capable body; it does not make the body interchangeable.
Our expectation is that early useful work will remain narrower than the advertised generality of the platform. Reuse becomes economically meaningful when a second task or site needs less engineering, and when failures can be handled without consuming the labour the robot was meant to save.
The enabling supply chain
A coordinated body, then a useful job
Mechanical and electronic inputs meet in the robot design. Training and simulation support its control software. Production makes repeatable equipment; task integration turns that equipment into working capacity.
Product or componentEngineering or productionSupplied outputSupply / processAlternative routeApplication-dependentDevelopment supportMaintenance / support
Each bubble opens its explanation below
The map separates the platform from its deployed task. A supplier can perform several activities internally, and a platform can be sold or supplied under a service agreement. The rings mark concrete outputs, not mandatory sales. The particular body, hand and learning architecture vary by product.
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.
The physical body
These inputs determine what the robot can lift, feel and sustain. Their specifications are coupled: saving weight in one place can change the demands elsewhere.
Economic nodes
Motion components
Supply not assessed
Motors, transmissions, bearings, encoders and drive electronics convert electrical commands into controlled force. Their size and loading differ between a hip, wrist and finger. Precision, efficiency and shock tolerance matter because losses and added mass propagate through a mobile body.
Key players
Schaeffler — bearings, motors and actuator components.
maxon (private) — compact motors and drive components.
Worth understanding
Schaeffler's humanoid presentation describes several actuator architectures rather than one universal joint. Component and module supply are distinct opportunities, but a product portfolio does not establish a current shortage. Schaeffler's December 2025 presentation →
A joint module packages motion components into a mechanical and electrical interface the robot builder can use. It must deliver the required torque, motion range and feedback while surviving repeated contacts and impacts. Rotary and linear designs solve different problems; one supplier appearing in several designs does not make the products identical.
Key players
Schaeffler — strain-wave actuator development and supply agreement with Humanoid.
Unitree — internally developed high-torque joint motors.
Schaeffler's January 2026 agreement with Humanoid identifies development and supply of strain-wave actuators. That supports a separable supplier role, while the announced partnership alone does not disclose shipped volume or customer economics. Actuator partnership →
The end effector makes useful contact: grasp a tote, hold a part or use a tool. A task-specific gripper can be sufficient; a multi-fingered hand seeks broader dexterity at the cost of more small actuators, sensing and failure points. The right hand is determined by the task, not by how human it looks.
Key players
Figure (private) — internally developed hands with tactile and visual feedback.
Agility Robotics (private) — Digit’s task-specific tote-handling end effectors.
Robotiq (private) — merchant grippers for suitable manipulation tasks.
Cameras describe the scene; joint-position, force and inertial measurements describe the robot itself. Touch can reveal contact that vision cannot see. These signals must be synchronized and calibrated before a controller can use them to coordinate a moving body.
Key players
ATI (private) · Novanta subsidiary — force/torque sensors for robotic contact measurement.
Figure (private) — integrated head, palm, fingertip and joint sensing.
Unitree — integrated depth sensing and lidar on H1.
Worth understanding
Figure describes head and palm cameras, fingertip sensing and joint state as inputs to Helix 02. This is one concrete architecture, not a universal sensor bill of materials. Sensor inputs →
The frame, links, covers, wiring and mechanical interfaces carry loads and protect the internals. Their arrangement fixes reach, payload, accessibility for repair and how damaging a fall or collision can be. A wheeled body and a biped can therefore share a family without sharing a production design.
Key players
Figure (private) — body and component integration for its own robot.
Boston Dynamics (private) · Hyundai Motor Group-controlled — Atlas humanoid body architecture.
BMW's AEON pilot is a practical reminder that human-like manipulation does not require a walking base. Body choice follows the workplace and job. Wheeled deployment example →
Cells become a usable robot power system through pack design, battery management, protection, power conversion and thermal engineering. The pack must support peak joint demands as well as average runtime; charging or swapping has to fit the operation.
Agility Robotics (private) — Digit charging and power-management integration.
Worth understanding
Agility's product updates describe autonomous charging and battery/status integration. Those features support availability, but a charging capability is not proof of continuous productive operation. Digit's operating features →
Training builds capability; deployment compute and control make it usable on a particular body. These activities can remain inside a platform producer.
Economic nodes
Training examples
Supply not assessed
Demonstrations, robot trials and motion recordings provide examples of what successful behaviour looks like. Collecting them requires a task, a body, instrumentation and a way to distinguish success from failure. More recordings are useful only if their actions and observations can be transferred to the target robot.
Key players
Scale AI (private) — robot-data collection and annotation.
NVIDIA — GR00T data-generation and training workflows.
Figure (private) — human-motion and task data for Helix.
Worth understanding
GR00T combines different data sources; Figure describes retargeted human motion for whole-body control. This work may be internal to the developer even when the data collection and processing are economically distinct activities. GR00T research → · Whole-body training →
Simulation lets developers test contacts, vary objects and practice behaviour without wearing out a physical robot on every trial. Models of the body and environment, training compute and evaluation tools contribute to this process. Incorrect friction, actuator or sensor models can still make a simulated success fail on hardware.
Key players
NVIDIA — Isaac simulation and GR00T training tools.
Figure (private) — internal simulation training for whole-body control.
Worth understanding
NVIDIA supplies Isaac simulation and training tools alongside its robot-model work. They support development rather than replacing physical validation. Isaac and GR00T tools →
Policy engineering turns training examples into software that chooses actions from observations and goals. The model must be adapted, evaluated and connected to the particular body. Pretrained models can reduce a starting burden while leaving task data, tuning and deployment work.
Figure (private) — Helix policies integrated into its own robot.
Worth understanding
GR00T N1.6 describes improvements on robot tasks and transfer experiments. The tests establish capability under their reported conditions; they do not establish general reliability across customer workplaces. Model evaluation →
Onboard processors run perception, task reasoning and learned policies within limits on power, heat and response time. Training can happen elsewhere; time-critical control needs a predictable response even when a distant connection degrades.
The GR00T development workflow distinguishes training and deployment on the robot. Compute can be a purchased module while the surrounding electrical and thermal integration remains the platform builder's work. Training and deployment workflow →
Control software coordinates joints, balance, contact and task execution at their different response times. Some implementations combine learned controllers with conventional control; others learn more of the low-level behaviour. Safety-related stopping and limits still need a defined implementation and validation.
Boston Dynamics (private) · Hyundai Motor Group-controlled — Atlas locomotion and manipulation control.
Worth understanding
The distinction from policy development is the delivered control system: software, sensor interfaces and actuator commands that work together on the physical robot. Figure's layered control description provides one example of that division. Control architecture →
The remaining work reproduces the machine and makes it dependable in a customer task. A platform demonstration does not establish these steps.
Economic nodes
Robot integration
Supply not assessed
Platform engineering chooses the body, joints, hands, sensors, power and control architecture as a coupled system. It resolves interfaces, timing, wiring, thermal limits and failure handling. This is where apparently interchangeable component choices become a specific robot design.
Key players
Figure (private) — internal integration of body, electronics and learned control.
Apptronik (private) — Apollo platform design and integration.
Worth understanding
Integration is often performed by the robot producer rather than a merchant integrator. Keeping it visible explains why an attractive actuator or policy does not, by itself, produce a dependable platform. The different implementations described by Agility → and Figure → illustrate the dependency.
Assembly, calibration, end-of-line checks and reliability testing must reproduce the design across units. Small errors in joint zero points, cameras or force readings change the behaviour the control software sees. Production also has to make worn parts replaceable and repeatable.
Key players
Figure (private) — BotQ assembly and production testing.
Jabil — Apollo manufacturing and validation collaboration with Apptronik.
Worth understanding
Figure's account of its BMW work includes operating experience with its second-generation platform and a subsequent platform transition. It is evidence for the importance of lifecycle learning; it does not turn announced factory capacity into completed robot shipments. Figure's operating account →
The platform is the assembled and tested mobile manipulator: body, controls and supported capabilities. It can be delivered as equipment or remain owned by a provider. Its specification does not promise that an unfamiliar customer task is already solved.
Agility's Digit is a concrete supplied platform. Its commercial deployment at GXO pairs the robot with software and integration, which is why the map continues beyond the equipment. Robot and software agreement →
Deployment specifies the job, connects work orders and surrounding machines, defines the operating area and validates acceptable behaviour. It includes task tuning, staff procedures, exception handling and acceptance. A new site can require this work even when the robot itself is unchanged.
Key players
Agility Robotics (private) — Digit and Arc integration into GXO’s workflow.
At GXO, Arc coordinates Digit with mobile robots and conveyors. BMW describes integration through standardized plant interfaces. Those are enabling activities, not background details outside the economic chain. Warehouse integration → · Plant integration →
The economic result is a specified job being performed in its working environment: totes transferred, parts handled or a machine tended. Throughput, damage, intervention needs and availability matter more than the range of motions shown in a demonstration.
Key players
Agility Robotics (private) — Digit supplied for tote transfer under the GXO agreement.
Figure (private) — deployment of Figure robots in BMW production work.
Worth understanding
This output does not prescribe the contract. The customer may own equipment, or pay a provider for its use. GXO's agreement demonstrates the service route without disclosing a universal price per task or profitable unit economics. Commercial route →
Spare joints and hands, calibration after repair, software updates, diagnostics and field support preserve the ability to do the job. A low purchase price can be outweighed by downtime or frequent specialist visits.
Key players
Agility Robotics (private) — continuing Digit fleet support through its service offering.
Worth understanding
Agility's Arc and operating-interface updates include monitoring and troubleshooting. Service demand grows with use, but its commercial attractiveness depends on reliability and how much support can be provided remotely. Operations and support features →
The most important uncertainty is how much deployment work becomes reusable. If each new job requires extensive demonstrations, custom handling logic and frequent intervention, a versatile body may still scale like bespoke automation. The economics improve when the same platform can enter another task with less engineering and maintain output over long operating periods.
Component demand can grow before general-purpose work is established. More prototypes and deployments require joints, hands, sensors, compute and test equipment. Sustained supplier value will depend on product differentiation, manufacturing yield and whether robot producers buy the module or bring it inside. Schaeffler's actuator partnership demonstrates one merchant route; it does not establish industry-wide outsourcing. Supplied actuator example →
A service agreement changes who pays for the robot and who carries downtime. GXO's RaaS deployment shows that this route exists. It does not remove charging, maintenance, human exceptions or integration from the system cost. The more useful comparison is cost per accepted task against the existing workflow, not a robot's nominal hourly wage equivalent. Commercial deployment →