A demo shows the action and hides the reset queue
Actuators and sensors create a capability envelope; manipulation policies turn contact into action; recovery protects uptime; technicians, spares, teleoperation, and workflow integration turn machines into dependable capacity.
Body and actuation
Structure, joints, transmissions, power, mobility, and interchangeable end effectors define force, reach, speed, and service life.
- Measure
- Payload · lifetime cycles · Wh/h
- Failure boundary
- Wear, heat, and maintenance shorten the usable life behind every dollar of capital.
Where the frontier moves
Cheaper, more durable actuators and modular hands that survive duty cycles without a full teardown.
Perception and contact
Vision, force, tactile sensing, and proprioception estimate objects, constraints, slip, and hidden contact state.
- Measure
- Uncertainty · sensing bandwidth
- Failure boundary
- Occlusion and ambiguous contact turn a confident policy into a silent failure.
Where the frontier moves
Tactile skins and force-aware policies that observe what cameras alone cannot.
Policy, dexterity, and recovery
Planning and learned behavior sequence compliant actions, detect failure, retry, regrip, or safely hand off.
- Measure
- Accepted sequences per intervention
- Failure boundary
- A high single-step success rate still compounds into frequent multi-step failure.
Where the frontier moves
Autonomous recovery that resolves the long tail without paging a human.
Safety and fleet operations
Validation, scheduling, teleoperation, spares, software updates, technicians, and workflow integration create delivered labor.
- Measure
- Availability · accepted hours per human-hour
- Failure boundary
- Supervision load, not hardware, is what keeps a fleet from scaling.
Where the frontier moves
Pooled remote experts and predictive maintenance that raise the robots-per-supervisor ratio.
A robot must move itself before it moves the work
Force, speed, reach, stiffness, battery energy, and contact information create unavoidable trade-offs. The useful floor is the least machine and supervision needed to complete a duty cycle safely for enough lifetime hours, not the lowest hardware bill.
A $100,000 machine delivering 20,000 accepted operating hours starts at $5 per hour before financing, integration, maintenance, energy, downtime, or supervision. That floor moves fast in both directions: fewer accepted hours from downtime or task mismatch raise it quickly, while pooled fleets that push hardware toward its full service life push it down just as quickly. Capital is the easy line item to quote and, once a fleet is running, rarely the largest one.
What actually sets the cost of an accepted hour?
The model isolates capital recovery, maintenance and integration, and human supervision per accepted hour. It is not a total labor-substitution cost and excludes energy, financing, and taxes.
Fully loaded, before energy, financing, and taxes
- Capital recovery
- $15.00 · 26%
- Maintenance and integration
- $3.75 · 6%
- Human supervision
- $40.00 · 68%
The reference case reproduces the article's own worked example: a $100,000 robot delivering 20,000 accepted hours is a $5/hour capital floor before anything else is added. Moving the supervision ratio and wage sliders shows why that floor is rarely the number that decides fleet economics.
Calculation and boundaries
Cost per accepted hour = (capital ÷ accepted lifetime hours) + (capital × maintenance/integration share ÷ accepted lifetime hours) + (supervisor wage ÷ robots per supervisor). Excluded: energy, financing cost, insurance, facility overhead, tooling changeovers, and taxes. Supervision is modeled as one loaded wage divided across the robots one person can competently oversee, the single input most responsible for why early pilots and mature pooled fleets differ by an order of magnitude.
The falling line is capital, not supervision
Unit manufacturing cost has moved further and faster than most forecasts assumed, but that is the smaller term in the fully loaded equation above.
The 2024–2025 cost drop, and what it does not fix
None of those figures describe accepted hours, supervision ratios, or intervention rates. A cheaper machine still needs the same recovery behavior, the same technicians, and the same safety validation per new task, which is why the calculator above treats hardware cost as one line among three, not the whole answer.
The bottleneck moves from perception to physical reliability
As models broaden scene coverage, manipulation reliability, cable routing, connectors, contamination, impacts, charging, actuator life, autonomous resets, and field repair become the binding constraints.
Publish intervention ledgers
Count human minutes, resets, damaged work, and failure causes alongside task success, instead of reporting autonomy as a binary demo.
Design for recovery
Compliant motion, tactile feedback, safe retreat, retry, and asynchronous escalation protect fleet economics more than one more percentage point of grasp success.
Reuse task packages
Portable behaviors, tools, and safety cases must cut engineering time per new deployment, or every SKU change restarts the cost clock.
Pool fleet service
Standard modules, spares, remote expertise, and predictive maintenance amortize support across many machines instead of one supervisor per robot.
Who is building what
Bipedal humanoids, generalist visual-motor foundation policies, compliance-aware tactile actuators, and fleet management architectures address general-purpose automation. Search the record, or filter by development domain.
Figure AIFigure 01 & Figure 02Bipedal general-purpose humanoids equipped with custom integrated actuators, 16-DOF robotic hands, onboard neural vision, and speech interaction
- Reported evidence
- Deployed in pilot manufacturing trials at BMW Manufacturing in Spartanburg, SC, performing autonomous sheet metal placement into fixtures.
- Announced next step
- Commercial fleet deployment across automotive assembly, logistics sorting, and eventual commercial household labor.
- Unresolved risk
- Hand actuator durability under heavy cyclic loads, battery life per charge (approx. 2–3 hours), and human teleoperation intervention frequency.
Boston DynamicsAll-Electric AtlasFully electric humanoid robot utilizing high-torque custom electric actuators with 360-degree rotational joint ranges, moving past legacy hydraulics
- Reported evidence
- Demonstrated autonomous parts sorting and automotive strut manipulation; validated in Hyundai manufacturing facilities.
- Announced next step
- Commercializing electric Atlas for harsh automotive and manufacturing environments with high duty cycle availability.
- Unresolved risk
- High manufacturing bill of materials (BOM), thermal cooling of dense high-power rotary actuators, and commercial software API integration.
TeslaOptimus (Gen 2)Humanoid robot utilizing proprietary integrated actuators, 22-DOF tactile-sensing hands, custom battery packs, and end-to-end neural network visual policies
- Reported evidence
- Internal factory testing at Fremont and Giga Texas sorting battery cells and moving parts along factory transport paths.
- Announced next step
- Sub-$30k volume manufacturing cost, internal deployment of thousands of units across Tesla factories, followed by external customer sales.
- Unresolved risk
- Reliability of pure end-to-end vision policies in novel unmodeled corner cases without deterministic safety fallback controllers.
Agility RoboticsDigitBipedal logistics robot with specialized tote-gripping end effectors designed specifically for bulk container moving in existing warehouse aisles
- Reported evidence
- Commercial deployments with GXO Logistics and Amazon distribution centers moving totes between conveyors and autonomous mobile robots (AMRs).
- Announced next step
- Opening RoboFab manufacturing facility in Oregon capable of producing thousands of Digit robots annually under Robotics-as-a-Service (RaaS).
- Unresolved risk
- Limited manual dexterity for small-item picking (designed for standardized bulk totes) and fall recovery in crowded warehouse aisles.
Sanctuary AIPhoenix HumanoidUpper-body and bipedal humanoid driven by Carbon cognitive architecture, utilizing fast hydraulic actuators in hands to achieve near-human dexterity
- Reported evidence
- Commercial pilot deployments at Canadian Tire retail distribution centers performing package sorting and shelf stocking.
- Announced next step
- Creating an artificial general intelligence (AGI) brain for physical work, pairing high-speed haptic teleoperation with autonomous policy training.
- Unresolved risk
- Hydraulic fluid maintenance, pump noise and weight, and bridging teleoperation demonstration data to robust autonomous execution.
ApptronikApolloModular bipedal humanoid with swappable batteries, quick-disconnect modular torso options for stationary mounting, and force-controlled compliant actuators
- Reported evidence
- Commercial agreement and pilot deployment at Mercedes-Benz manufacturing plants delivering assembly kits to human workers.
- Announced next step
- Industrial ergonomics: collaborative operation alongside human line workers with standardized 55-pound payload capacity.
- Unresolved risk
- Footprint and balance recovery when operating at maximum payload extensions in narrow factory walkways.
Unitree RoboticsG1 & H1 HumanoidsLow-cost bipedal humanoid platforms utilizing proprietary high-torque density joint motors, targeting sub-$16,000 retail price points
- Reported evidence
- High-volume hardware production and global shipping of G1 robots to academic labs, corporate developers, and hobbyists worldwide.
- Announced next step
- Commoditizing humanoid robotics hardware, allowing the global software community to build policies on affordable physical platforms.
- Unresolved risk
- Joint gear backlash, lighter structural stiffness under heavy loads, and baseline autonomy software maturity.
Physical Intelligence (π)π0 Foundation PolicyUniversal visual-motor foundation model trained across massive multi-robot, multi-task datasets, executing complex manipulation across diverse robot bodies
- Reported evidence
- Demonstrated zero-shot and few-shot dexterous manipulation (folding laundry, busing tables, box assembly) across humanoids, dual-arm, and mobile manipulators.
- Announced next step
- The 'Android of robotics': a single physical intelligence foundation model powering all commercial robot hardware vendors.
- Unresolved risk
- Execution latency on edge compute devices and sample efficiency collecting physical real-world failure trajectory data.
NIST / IFRRobotic Systems Performance FrameworkStandardized benchmarking protocols evaluating robot grasp strength, cycle times, safety collision thresholds, and human intervention taxonomies
- Reported evidence
- IFR World Robotics annual census tracking global robot installations; NIST manufacturing testbeds quantifying collaborative robotic safety.
- Announced next step
- Establishing open international benchmark standards for evaluating general-purpose robot productivity and autonomy reliability.
- Unresolved risk
- Industry reluctance to report standardized intervention ledgers and failure modes on public benchmarks.
Demonstration videos showing continuous autonomous manipulation often cut out manual resets and teleoperation intervention. Commercial viability is defined by mean time between interventions (MTBI) during multi-hour duty cycles.
The optimistic view, with conditions
General-purpose robotics becomes a service-level agreement
The credible product is a managed fleet quoting accepted throughput, availability, task-change cost, intervention burden, and repair response, not a humanoid silhouette with a catalog of demonstrations.
Measure interventions, not demos
Publish human minutes per accepted hour and classify failure causes before claiming autonomy.
Let capital become the small term
As unit prices keep falling toward Goldman's projected path, capital recovery per hour stops being the constraint.
Raise the supervision ratio
Remote experts and predictive maintenance only create leverage once exceptions are rare enough to handle asynchronously.
What a general-purpose robot-hour actually needs
- Falling, financeable capitalUnit cost continuing to track toward Goldman's projected decline, not just an early-pilot price.
- Contact-rich reliabilityManipulation policies that recover from ambiguous contact without paging a technician.
- Pooled supervisionA ratio of robots to humans that rises well past 1:1 as exceptions become rare and asynchronous.
- Portable task packagesReusable behaviors, tools, and safety cases that avoid a new engineering project per SKU.
- Published intervention ledgersOperators willing to report accepted hours, resets, and failure causes, not just task-success clips.
The adoption curve belongs to useful tasks
IFR counted 542,076 industrial robot installations in 2024, more than twice the count a decade earlier. The manufacturing stock reached 177 robots per 10,000 workers, from 163 the year before. Those are mainly task-specific machines. The separate IFR service-robot survey records more than 199,000 professional service units sold in 2024, including about 102,900 for transport and logistics. Neither series is a verified count of general-purpose humanoids.
For a general-purpose robot, the adoption metric should be unassisted productive hours per human supervisor-hour, measured over new sites and tasks, alongside safety interventions and uptime. A vendor video or a pilot shipment does not establish that denominator. Until operators publish comparable logs, no credible global 2025–26 humanoid deployment curve or supervision ratio can be inferred from industrial-arm installations. The economic threshold is task-specific: all-in robot cost per accepted job must beat wages, safety, rework and supervision for that job.
Factory diffusion has a fresh denominator
The IFR's September 2026 release reports 603,000 industrial robots installed in 2025 and 5.079 million operating, up 11% and 9% respectively; China installed 354,000. These are structured industrial robots, not a humanoid shipment count. A 2025 worldwide humanoid total assembled from vendor announcements would mix shipments, pilots, preorders and demonstrations. The missing evidence remains accepted jobs per unsupervised hour across several tasks, interventions per shift, mean time between failures and independently paid deployments.
For a home robot the comparison is more demanding than for a fenced workcell: people, pets, clutter and changing tasks increase exception and safety costs. A transparent commercial threshold would report annual all-in robot cost divided by accepted productive hours, then add human intervention time at the local wage. Until vendors publish those inputs, a low advertised hardware price cannot establish a cheaper useful hour.
Sources, method, and boundaries
IFR deployment figures establish the industrial baseline and are measured 2024 totals. NIST performance and manipulation frameworks define comparable performance dimensions. Goldman Sachs Research and Bank of America Global Research manufacturing-cost figures are analyst estimates, not audited industry costs, and Goldman's shipment path is an explicit projection. The interactive model is a derived, illustrative accounting identity excluding energy, financing, and taxes, not a fleet-price quote. Vendor demonstrations are treated throughout as capability evidence, never as fleet economics.
- Accepted hour
- An hour of robot operation whose output passes quality and cycle-time requirements without human correction.
- Intervention
- A human action required to resolve a robot exception: a reset, a regrip, a manual override, or a hand-off.
- Supervision ratio
- The number of robots one human can competently oversee at an acceptable exception rate.
