Robotics · Annual review · 2026-08-30
Robots Became Operations
A few humanoids crossed into measured work. The harder story was everything around them: qualified joints, hands, safety, field failures, integration and the economics of deployed labor.
Dylan Austin Bristot, Founder and publisher, AI Bottlenecks
TL;DR
- The year did not prove that humanoids had arrived. It proved that a few robots could do real work, then exposed the manufacturing, supervision, repair and integration behind that work.
- Figure reported more than 90,000 BMW part loads across 1,250 hours. Agility's Digit passed 100,000 tote moves at GXO. These were the best receipts because they measured tasks rather than demos.
- UBTECH supplied the clearest audited humanoid P&L: 1,079 full-size units, RMB820.6 million of related revenue and a RMB789.8 million annual net loss. Volume arrived before attractive economics.
- Capacity ran ahead of demand proof. Figure demonstrated one robot per hour, UBTECH reported annualized capacity above 6,000 units, Agility designed for 10,000 and Hyundai targeted 30,000 Atlas units by 2028.
- The next useful KPI is successful autonomous work per human intervention, after charging, maintenance, teleoperation and integration labor.
The tape paid the factory before the robot
The current Robotics Bottlenecks map contains 45 report-date names across six themes. The model projects today's universe backward, equal-weights themes daily and equally weights names inside each theme. That makes this a look-ahead research test, not a live portfolio.
| Theme | One-year return | Drawdown from peak | | --- | ---: | ---: | | Automation, Test & Integration | +81.5% | -13.9% | | Magnets, Materials & Power | +55.0% | -13.9% | | Precision Motion & Actuation | +52.3% | -19.4% | | Robot Compute & Control | +37.2% | -27.0% | | Sensing, Vision & Dexterity | +30.7% | -8.7% | | Robot Platforms & Demand Signals | +23.0% | -25.1% | | BOTZ | +7.1% | n/a | | S&P 500 | +19.4% | n/a |
The full retrospective book returned 82.5% and peaked at 111.7% on 26 May. These companies are diversified and several listings are thin. The ordering is still worth noticing: the market paid most for the machinery that tests, integrates and moves robots, not for the public platform basket closest to the demo feed.
!The robotics layers reached the finish line by different paths
Seven rungs between a demo and deployed labor
“Deployment” came to mean everything from a staged video to paid factory work. The language needs a harder ladder.
!A humanoid changes meaning seven times before it becomes deployed labor
- Demo: one task in a controlled scene.
- Sample: parts or robots reach an OEM program.
- Qualified: the system clears tolerance, lifetime and safety work.
- Pilot: the robot operates inside a real customer workflow.
- Production order: firm volume and delivery dates exist.
- Fleet uptime: successful cycles survive charging, failure and recovery.
- Unit economics: the task pays back after service and human support.
The distance between rung one and rung seven is the robotics bottleneck.
The year in five moves
1. Compute became available
NVIDIA made Jetson Thor generally available in August 2025 with 128 GB of memory and a 130 W power envelope. Qualcomm introduced Dragonwing IQ10 in January. Google moved through Gemini Robotics 1.5 and 2. NVIDIA expanded GR00T, Isaac and its simulation stack. Jetson Thor | Qualcomm IQ10 | Gemini Robotics 2
Raw TOPS mattered less by the end of the year because several credible platforms could supply them. The hard problem was mixed-criticality control. A neural policy can plan a motion. The robot still needs deterministic loops, motor control, thermal limits, safety logic and graceful recovery when the world refuses the plan.
Compute availability did not remove the data bottleneck either. Human video does not contain measured torque, joint temperature, slip, collision force, intervention reason or the part that failed after the 9,000th cycle.
The valuable dataset moved toward reality disagreeing with the model.
2. A few robots crossed into measured work
Figure reported more than 90,000 sheet-metal parts loaded at BMW, more than 1,250 operating hours, contribution to more than 30,000 vehicles and greater than 99% placement success per shift. That was real production work. It was also one constrained task, without disclosed fleet size, intervention rate, maintenance labor or customer payback. Figure said the forearm was the leading hardware-failure area. Figure at BMW
Agility's Digit passed 100,000 tote moves at GXO. By August, Agility reported more than 65,000 aggregate operating hours and more than $300 million of multi-year orders subject to milestones. Digit moved totes between existing automation assets. The humanoid acted as an exception-handling bridge inside a designed system, not a replacement for the warehouse. Agility at GXO
Both companies proved something more useful than general labor: a narrow workflow can create enough value to justify learning in production.
3. China put audited volume beside the story
UBTECH's 2025 filing was the clearest public humanoid receipt. Group revenue reached RMB2.00 billion, up 53.3%. Full-size embodied humanoid product and service revenue reached RMB820.6 million, up 2,203.7%. The company reported 1,079 units sold, more than 1,000 delivered and revenue-recognized, and annualized capacity above 6,000 units. UBTECH annual results
Then came the number that keeps the story honest: the group still lost RMB789.8 million.
The category included products and services, and UBTECH did not disclose autonomous productive hours, intervention rate or unit-level gross margin. This was the bull and bear case in one filing. Units and revenue were real. Platform economics were not solved.
China's advantage extends past one OEM. It installed 295,000 industrial robots in 2024, 54% of the global total, and operated more than two million. Local suppliers captured 57% of domestic installations. That density creates a learning loop across motors, reducers, controls, magnets, factories and customers. IFR World Robotics
It also threatens simple component-scarcity trades. The lasting rent has to survive localization, redesign and brutal price competition.
4. Manufacturing stopped being an aspiration
Figure's BotQ update published the useful factory metrics. It had delivered more than 350 F.03 robots, moved from one unit per day to one per hour in fewer than 120 days, achieved above 80% end-of-line first-pass yield and built more than 9,000 actuators across more than ten SKUs. “Delivered” included internal research and data-collection robots, not 350 customer sales. First-pass yield was still a better signal than a capacity target. Figure BotQ
The capacity targets ran much further ahead:
- Figure demonstrated one robot per hour.
- UBTECH reported annualized capacity above 6,000.
- Agility designed RoboFab for as many as 10,000 per year.
- Hyundai targeted a 30,000-unit annual Atlas system by 2028. Hyundai
Investors had annual capacity. Customers needed qualified output matched to a task. Those are very different factories.
5. The installed base remained the control group
The world installed 542,000 industrial robots in 2024. U.S. installations rose 11% to 38,000 in 2025. The listed receipts were cleaner than most humanoid disclosures:
- FANUC's March-quarter robot orders reached JPY100.6 billion, up 10.6%.
- Teradyne's Q2 Robotics revenue reached $99.9 million, up 33.4%.
- Symbotic reported $721 million of quarterly revenue, $55 million of net income, $95 million of adjusted EBITDA and 77 systems in deployment.
- SoftBank agreed to pay $5.375 billion for ABB Robotics, a mature operation with $2.279 billion of 2024 revenue, service history and an installed base.
FANUC | Teradyne | Symbotic | ABB transaction
The control group shows what customers already buy: a robot, integration, safety, service, spares and a workflow that survives a production calendar. Humanoids do not need to replace industrial robots. They need to fill enough expensive gaps around installed automation to make the line more flexible.
The robot is a chain of error budgets
!Every physical layer adds another error budget
The scarce object is a qualified joint
Investors often stop at precision reducers. The robot buys a joint envelope: motor, reducer or screw, encoder, torque sensor, drive, brake, cabling, thermal path, electronics and structural interfaces. Each part can meet its catalog specification and the joint can still fail after integration.
Harmonic Drive Systems showed the cyclical recovery. Its following first-quarter orders rose 55.7%, sales rose 23.6% and full-year guidance increased. Nabtesco reported JPY22.1 billion of first-quarter precision-gear sales and an 8.8% Component Solutions margin. HDS | Nabtesco
The competitive response was integration. Schaeffler introduced a planetary actuator combining gearbox, motor, encoder and controller, then claimed a formed strain-wave gear could cut cost by more than 25%. The cost claim remains a company claim. The direction matters: a reducer can be scarce until a customer redesigns the joint. Qualification data, calibration yield and field life are harder to substitute. Schaeffler
Hands are where generality hits the bill of materials
A useful hand meets objects with unknown shape, friction and fragility. It needs tactile sensing, compact motors, transmissions, cabling, heat removal and grasp recovery inside a small envelope.
Figure redesigned its hand around palm cameras and in-house tactile sensors after commercial options failed its durability requirements. The last centimeter is expensive because a model error turns directly into wear or damage. Hands will remain where impressive videos encounter service inventory.
Failure data became more valuable than generic video
Human video can teach task structure. Simulation can generate variation. Neither automatically captures the state around a near miss, a slipping object or a hot actuator.
The valuable field record includes sensor state, contact force, current, intervention reason, recovery action, environmental variation and maintenance outcome. That data comes from consequential work. The integrator, fleet operator or internal factory customer may own more of it than the model company.
Safety became product architecture
ISO 10218:2025 modernized industrial-robot safety and UL 3300 created a clearer U.S. conformity route for service robots. Certification matters, but a badge does not make a stochastic robot safe beside people. The system still needs bounded force, safe stop, fault containment, verified tools, site rules and incident recovery. ISO 10218 | UL 3300
Agility said current Digit deployments remained separated from people. Mixed-workspace operation is not merely a model update. It is a safety case and a customer-liability decision.
Magnets became strategic before they became the first bottleneck
China's April 2025 controls covered heavy rare earths and permanent-magnet products. U.S. policy answered with a ten-year $110/kg NdPr price floor, purchase support and financing for MP Materials. By Q2 2026, MP had $16.5 million of Magnetics revenue and $7.5 million of adjusted EBITDA. DoD partnership | MP Q2 2026
Policy created optionality. Qualification still controls the clock, and humanoid demand remains too small to explain most magnet economics today.
Six calls for the next 12 to 24 months
1. Paid humanoid work concentrates in controlled workflows
Material handling, sequencing and repetitive manipulation should produce the next repeat orders. Homes remain further away.
Watch: customer expansions after six months of operating history. Wrong if: household deployments produce better disclosed uptime and payback first.
2. The humanoid becomes an exception handler
The strongest economics come from filling gaps between fixed automation, AMRs and people, not replacing the whole line.
Watch: task boundaries, handoffs and productive hours outside the headline task. Wrong if: general humanoids operate full shifts across many tasks without engineered boundaries.
3. Repeat orders replace videos as the scarce signal
A customer that has measured interventions, maintenance and payback knows more than a new pilot logo.
Watch: paid fleet expansions and gross profit after service. Wrong if: first pilots keep receiving full credit after mature programs publish weak economics.
4. Integrated actuator quality matters more than reducer capacity
The supplier has to clear calibration, thermal, cabling, drive and lifetime requirements as one joint.
Watch: named awards, joint yield, warranty data and field replacement. Wrong if: catalog reducers remain the binding constraint after integrated modules scale.
5. Hands remain the main physical reliability bottleneck
Hands absorb contact uncertainty and pack too many wear items into a small envelope.
Watch: fingertip life, cable replacement, grasp recovery and cost per hand. Wrong if: several vendors disclose thousands of mixed-object hours with low service cost.
6. Traditional automation captures more revenue than humanoids
The installed base, service network and integration channel remain larger through 2027.
Watch: FANUC, Yaskawa, Teradyne, Cognex, Symbotic and ABB receipts against disclosed humanoid revenue. Wrong if: global humanoid product and service revenue overtakes incremental traditional-automation growth before 2028.
The monitoring board
| Signal | Strong evidence | Warning | | --- | --- | --- | | Paid active robots | Growth at named customer sites | Capacity rises faster than paid units | | Autonomous hours | More hours with stable task success | More supervision as fleets grow | | Interventions | Declining interventions per 100 hours | Teleoperation remains hidden | | First-pass yield | Higher yield and lower rework | Output targets rise while yield stalls | | Joint and hand life | Longer life and lower service cost | Forearm, cable or fingertip failures persist | | Repeat orders | Expansion after measured operating history | Endless first pilots and frameworks | | Gross profit per system | Positive after integration and service | Revenue grows with support losses |
The read
The year began with the model and ended at the loading dock.
The useful systems were robots inside controlled factories and warehouses, moving parts between known stations. The best disclosures counted parts, totes, runtime and yield. Then the failures pointed to the next constraint: forearms, hands, charging, safety, maintenance and the operator recovering an edge case after it became physical.
The scarce object is no longer the ability to make a humanoid move. It is a qualified robot plus a designed workflow plus enough field data and service capacity to keep the task cheaper than the alternatives.
> Deployed labor = task success x cycle rate x uptime x useful life, net of human intervention
Robot count is inventory. Deployed labor is the product. That distinction is what Robotics Bottlenecks is built to track.
Methodology
Market figures cover 28 August 2025 through 28 August 2026 and project the report-date universe backward. Prices are Yahoo adjusted closes converted to U.S. dollars. The model equal-weights themes, then names within them, rebalanced daily. It excludes fees and factor adjustment and is descriptive, not a live performance record.