A useful robot starts with a defined job
A factory does not buy a human-shaped robot because the shape is impressive. It buys capacity for a particular job. That distinction changes the questions we ask about humanoid robotics: what task is being completed, what happens when the task fails, who intervenes, and what does an accepted unit of work cost? A short demonstration rarely answers all four.
The human form can be useful where the building, tools and material flow already assume human reach and movement. It also creates obligations. The robot must balance, manage contact, carry its own energy supply and coordinate several systems under load. A wheeled mobile manipulator or fixed arm may solve a narrower job with fewer demands. The right comparison is the customer's practical alternative, including leaving the process manual.
This article evaluates deployment evidence and operating economics. The actuation guide examines joints and lifetime component cost. The perception guide examines sensing failures and action latency. Those questions support a deployment decision; repeating the whole category thesis on each page would obscure them.
A customer agreement is stronger evidence than a demonstration
GXO announced a multi-year agreement with Agility Robotics in June 2024 to deploy Digit in logistics operations under a robots-as-a-service arrangement. That is identifiable customer-side evidence of commercial use. The announcement does not disclose enough to calculate the customer's return or establish general-purpose autonomy.
BMW's 2026 Figure 03 announcement describes a new sequencing use case and draws on its earlier Figure 02 production experience. The distinction between generations and tasks matters. Experience on one assembly-line job does not establish that the successor performs every logistics job, or that another site can adopt it without additional engineering.
These sources support a narrower conclusion than a universal labor-replacement forecast: some customers are using or developing humanoid applications in operating facilities. To go further, we would need recurring order evidence and task-level operating data. Neither a company valuation nor a projected addressable market fills that gap.
Reliability lives in the complete work cycle
A work cycle can fail after perception succeeds. The robot may identify the part correctly but grasp it poorly, place it outside tolerance or lose balance during recovery. Increasing the capability of one subsystem may therefore have little effect on accepted throughput when another subsystem controls the outcome. We evaluate the whole sequence, including the return to a safe starting state.
Hands, actuators, power, perception and control are connected constraints. A heavier hand may improve manipulation while increasing joint loads and energy use. More onboard compute may improve decision-making while consuming power and requiring cooling. Stronger actuators can increase peak capability without making sustained operation economical. Engineering tradeoffs should be stated at the system boundary where the customer pays.
Research on a two-speed actuator illustrates a particular torque-speed tradeoff and a proposed architecture. It is useful engineering evidence, not a measured cost curve for commercial humanoid fleets. We do not carry a research prototype's performance directly into a production robot's bill of materials or operating margin.
An illustrative cost per accepted hour
Consider a hypothetical deployment costing $30,000 a year for equipment or service, integration allocation and maintenance, before supervision. Suppose it schedules 2,000 hours and delivers useful work during 60% of them. That produces 1,200 accepted hours, so the stated cost is $25 per accepted hour. At 80% useful utilization, the same annual cost becomes $18.75 per accepted hour.
These are invented scenario inputs, not vendor quotations or measured fleet results. They demonstrate why the denominator matters. Scheduled hours include charging, faults, blocked material flow and waiting for intervention; an economic comparison needs the hours that actually deliver the required output. Task speed also matters: one accepted robot hour is not automatically equivalent to one paid human hour.
Supervision can change the result again. If each robot needs frequent dedicated help, adding robots adds labor as well as capacity. If one person can reliably supervise several robots without delaying recovery, the economics improve. Any deployment model should make that ratio visible and include the cost of the surrounding process changes.
Where value might accrue
Figure AI and Agility Robotics are robot-platform examples in this research because the linked customer announcements identify their applications. They are private companies, and discussion of their progress is not a public-market investment recommendation. A platform can create customer value while still spending heavily on manufacturing, support and field engineering.
Component suppliers become attractive when they solve a difficult, repeatedly purchased problem at an acceptable price. Qualification, replacement difficulty and service requirements can matter as much as a component's share of a prototype bill of materials. A large addressable category does not establish the share available to one supplier, and a partnership announcement does not establish a production order.
The company map is deliberately short. It identifies the platforms supported by the deployment sources rather than treating every motion-control or semiconductor company as confirmed humanoid revenue exposure. Supplier-specific claims belong in sourced company research, with customer relationships and revenue attribution checked separately.
What would change the conclusion
The constructive case strengthens with customer-reported repeat deployment, stable accepted throughput, lower intervention requirements and evidence that equipment and service costs are covered by the work delivered. The strongest comparisons follow the same task over time and explain what changed in hardware, software and operating conditions.
The case weakens when expansion repeatedly requires bespoke engineering, when service costs rise with fleet size, or when a simpler automation alternative provides the same outcome at lower total cost. A robot can improve technically while becoming a less attractive purchase if the surrounding process changes or the customer's alternatives improve.
The current conclusion is therefore conditional: humanoids have credible deployment evidence, but broad commercial readiness must be demonstrated task by task. We would prioritize an understandable customer workflow and a transparent cost model over a larger collection of demonstrations. No site visit, independent fleet trial or undisclosed customer interview is claimed here.