The first industrial robot went to work at a General Motors plant in 1961. Unimate weighed about two tonnes, cost $65,000 to make and sold for $18,000. 1 H. In 2026 the Chinese humanoid Unitree R1 sells from $4,900 (excluding taxes and delivery; in China from 29,900 yuan) 2 M, and 5m industrial robots work in the world’s factories 3 H. Over six decades robots have become cheaper, more precise and more numerous. But what remains hardest for them is what seems simplest to people.
This asymmetry was described in 1988 by the roboticist Hans Moravec: it is comparatively easy to teach a computer to play draughts at an adult level, and hard to make it perceive the world and move like a one-year-old 4 M. Thirty-eight years later, the gap looks like this. Language models learn from trillions of words from the internet. Open X-Embodiment, one of the largest open datasets for training robots, contains, by the estimate of a 2025 survey, roughly 200,000 times fewer tokens than the corpora of language models 5,6 L. Laboratory tests of leading models in 2022–2026 show roughly 50–80% success on complex tasks (from 32% to 92% depending on the task). The authors ran each test themselves, on different robots and tasks, so the figures are not comparable 7,8,9,10 L. A factory, by our estimate, needs reliability of 99% and above L.
Robots scale not where they are cleverest but where the world is easier to adapt to the machine: in factories, warehouses, operating theatres and at the front. The home, where every room is different, remains the hardest market.
In brief
- Scale. In 2025 more than 600,000 industrial robots were installed worldwide, 59% of them in China. 5m are in operation 3 H. Almost 34.2m household service robots, mostly vacuum cleaners, were sold 11 M; IDC puts cleaning robots at 32.7m 12 M.
- Money is already being made where the environment is standard: warehouses (Amazon’s network has more than 1m robots), operating theatres (3.15m da Vinci procedures in a year) and the roads of about 15 US cities (Waymo, more than 500,000 rides a week, April 2026).
- Humanoids are only beginning to be counted in tens of thousands a year. More than 60% of those bought in the first half of 2026 went to performances, research and data collection, 13% to factories and 5% to warehouses 13 H.
- The 2023–2026 breakthrough is the vision-language-action (VLA) model: the robot receives a command in words and transfers skills between different bodies. The bottleneck is data: it is collected by human operators, who are paid $25–55 an hour in the US 14 L.
- The labs show self-healing muscles, microrobots that cured sinusitis in rabbits in an experiment, and a hand moved by muscles grown from human cells. They are anywhere from a few years (artificial muscles are already supplied to individual customers) to 15 or more years (the hand on living muscle, not before 2040) from a product L.
- Ukraine is the world’s fastest proving ground for ground robots: according to West Point analysts, about 2,000 were delivered in 2024 and almost 15,000 in 2025 (this matches the Ministry of Defence’s plan for 2025), and as of July 2026 more than 22,000 had been contracted (contracts are not delivered machines) 15,16 M.
- 2035–2040. The banks disagree: Goldman Sachs (February 2024) expected more than 250,000 humanoids a year in 2030 and 1.4m in 2035 17 M, while BofA (March 2026) expects 1.2m as early as 2030 18 M. We found no bank forecasts for 2040. BofA’s own analysts call the home humanoid “the story of the 2040s” 18 M.
01 · OverviewWhere robots work in 2026: 5m in factories, 34m household robots sold in a year, thousands on the roads
The word “robot” covers very different machines. The largest and oldest group is industrial manipulators: “arms” that weld, paint and move parts. According to the International Federation of Robotics (IFR), factories worldwide installed more than 600,000 such robots in 2025, 11% more than a year earlier, and 5m are now in operation 3 H. China installed 354,000, or 59% of all new robots. The US (38,500) moved into second place, overtaking Japan (36,200), followed by South Korea (~30,000) and Germany (under 25,000) 3 H.
The long series shows that robots grew not evenly but in waves. In 2009 the crisis cut installations by 47%, to ~60,000, the lowest since 1994 19 H. Then came ten years of rapid growth, driven mostly by China, and a plateau: 553,000 in 2022, 541,000 in 2023 and 542,000 in 2024. The jump in 2025 beat IFR’s own forecast (575,000), but the counters disagree: Interact Analysis counts 549,600 for the same year, 9% fewer 20 M. For 2029 IFR expects 806,000 installations, that is, growth of 7–8% a year 3 H.
The customer has changed too. In 2013 the car industry accounted for 45% of installations with a known sector, in 2023 for 30%. In 2024 electronics (24% of all installations) again overtook cars (23%): the two sectors swap places every year or two 21 H. Robots got cheaper: by Graetz and Michaels’s calculations from IFR price data for six developed countries, their quality-adjusted price fell between 1990 and 2005 to about a fifth of the 1990 level 22 H. We found no comparable price index for later years, but competition keeps up the pressure. Chinese brands hold 55% of their home market (57% in 2024, about 28% ten years ago), and according to a supplier citing Interact Analysis, their robots are 30–40% cheaper than the global average price 3,23 L.
The second yardstick is density: how many robots there are per 10,000 manufacturing workers. Here the world divides sharply. South Korea has 1,220, Singapore 818, Germany and Japan about 450, the US 307, and the world average is 132 24 H. In April 2026 IFR recalculated density using updated employment data for China. On the new basis China has 166 in 2024 (22nd place); for 2023 it comes to roughly 142 (our calculation) L. On the old basis, the figure for 2023 was 470. There were not fewer robots; the denominator changed. The world average for 2024 was likewise revised from 177 to 132. This is a useful reminder: rankings in robotics can depend on method more than on machines.
Beyond factories, IFR counts service robots. In 2025 almost 250,000 professional ones were sold (+24%), of which 117,500 were for transport and logistics and 8,400 were surgical (+57%) 11 H. Household ones numbered 34.2m, mostly vacuum cleaners; according to IDC, all five of the biggest cleaning-robot brands are Chinese 12 M. And iRobot, which launched the first Roomba in 2002 and sold more than 50m robots, filed for bankruptcy protection in December 2025 and was bought by a Chinese contractor 25 M.
02 · PrehistoryFrom automata to Unimate: culture imagined the robot before engineers built it
Mechanical dolls and automata existed as far back as antiquity: Hero of Alexandria described temple doors that opened by themselves, and in 1206 the engineer al-Jazari described a boat with four mechanical musicians 26 M. But these machines repeated one and the same action. The idea of a programme that could be changed came with Jacquard’s loom of 1804: the pattern of the cloth was set by punched cards 27 M.
The word “robot” appeared in the theatre: in Karel Čapek’s play “R.U.R.”, first staged on 2 January 1921 in Hradec Králové (at the National Theatre in Prague on 25 January). The word comes from the Czech robota, meaning forced labour 28 M. Isaac Asimov formulated the three laws of robotics in 1942, 19 years before the first industrial robot went to work 29 M. George Devol filed a patent for “programmed article transfer” in 1954, and in 1961 Unimate began work at a General Motors plant 30,1 H.
Robotics then split into two branches. The industrial one quickly became a business: Japan bought a Unimation licence in 1968 and by 1990, by our calculation from UNECE and IFR data, accounted for about three-quarters of new installations worldwide 31 M. The research branch learned to see and reason. The robot Shakey (SRI, 1966–1972) planned actions in a room 32 H. The A* path-finding algorithm also came out of its development M. The two-legged WABOT-1 of Waseda University (1973) walked and spoke simple Japanese 33 H.
There was a Kyiv chapter in this story too. In 1972–1975 the biocybernetics department of the Institute of Cybernetics, headed by Mykola Amosov, built an autonomous transport robot, TAIR, with a rangefinder, touch sensors and a neural network assembled from transistors 34 M. Several more machines followed, including MAVR for off-road movement (1984–1986). According to the department’s own history, the line of work faded because industry could not use these developments. We found no independent confirmation of these details, so we present them as the testimony of the school itself.
03 · Milestones 2000–2026A hundred milestones in a quarter-century: from a race nobody finished to a humanoid that runs faster than a human
The twenty-six years can be divided into four eras. The boundaries are arbitrary, but each era solved one big problem and opened the next.
2000–2007: machines learn to navigate
In March 2004 DARPA launched a race for driverless cars across the Mojave desert. None finished: the best covered 11.78km and got stuck. A year and a half later Stanford’s Stanley covered 212km in 6 hours 54 minutes, and five vehicles finished 35 M. Between the two races the software changed too: vehicles began to estimate how sure they were of their own position and to allow for sensor error L. Robots also entered mass-market niches at this time: in July 2000 the FDA cleared the da Vinci system for general laparoscopic surgery 36 H, the Roomba vacuum cleaner appeared in 2002 25 M, and the Opportunity Mars rover, designed to last 90 Martian days, worked for 14 years and drove 45km 37 H.
2008–2015: cobots, warehouses and open source
Robots were let out of the cage. In 2008 Denmark’s Universal Robots showed the UR5, a light arm that works safely alongside people 38 H. In 2010 the first version of ROS appeared, an open software platform that made robot code something laboratories could share 39 M. In 2012 Amazon bought Kiva for $775m: mobile robots began to bring shelves to people rather than the other way round 40 M. In 2013 China became the largest market for industrial robots 41 H. Humanoids meanwhile remained an experiment: at the DARPA Robotics Challenge final in 2015 robots fell over while opening doors and getting out of a car, and production of the “social” Pepper was suspended in 2021 after about 27,000 units had been built 42,43 M.
2016–2022: deep learning reaches hands and wheels
In 2016 Google showed an “arm farm”: six to 14 manipulators made more than 800,000 attempts to grasp an object over two months, and a neural network learned to grasp from this data 44 H. In 2019 an OpenAI hand, trained only in simulation, solved a Rubik’s cube — but in 60% of attempts for simple scrambles and 20% for the hardest 45 M. In October 2020 Waymo opened rides with no driver and no safety operator to all comers in Phoenix M. And in 2021 it emerged that OpenAI had disbanded its robotics team (the company dates the decision to October 2020). According to a co-founder, there was not enough data for further progress 46 M. That phrase describes precisely the main constraint of the years that followed.
2023–2026: language models control the body, China scales
In July 2023 Google DeepMind presented RT-2: robot actions were written as text, and a model trained on internet data began to control an arm. This is how vision-language-action (VLA) models appeared 47 H. Since then π0, Helix, Gemini Robotics, GR00T and other models of this type have followed 48,49,50,8 M. Meanwhile China turned humanoids into a commodity. At the first robot half-marathon in Beijing in April 2025 the winner ran 2 hours 40 minutes, with battery swaps and a fall. A year later a Honor robot covered the same distance in 50 minutes 26 seconds — faster than the human record on the day of the race (57 minutes 20 seconds). The comparison is imperfect: the course and rules for robots were different, and the times of teleoperated robots were multiplied by 1.2 51 M. Running tests walking, not hands. But the pace of change is telling.
04 · MoneyWhere robots already earn their keep: operating theatres, roads, warehouses — wherever the environment can be standardised
Surgery. The most convincing success story outside the factory is the surgical robot. In 2008 about 136,000 operations were performed with the da Vinci system; in 2025, 3.15m (+18% on the year) on 11,106 installed systems; in the first nine months of 2026, already 2.64m (+15%) 36,52 H. The robot does not operate by itself: it transmits the movements of the surgeon’s hands to instruments, filters out tremor and provides a magnified three-dimensional image. At the same time its range of use is widening. In 2017 general surgery accounted for 38% of da Vinci operations in the United States; in 2025, 62% 36 H. Rivals have appeared, but with narrow American clearances: Versius (2024, cholecystectomy only), Medtronic’s Hugo (2025, urology only) and J&J’s Ottava (2026, upper abdomen, phased launch with no date) 53,54 M. In December 2024 a surgeon in Shanghai operated on a patient in Kuwait about 7,000km away. This was an isolated case; the average round-trip signal delay was 181ms, and a local surgeon stood by 55 H.
Robotaxis. In mid-2024 Waymo made about 50,000 paid rides a week; in April 2026, more than 500,000, and it plans more than a million by the end of the year 56 M. Its vehicles had driven 271m miles without a driver by June 2026 57 H. By the company’s own analysis (there is no independent audit), they had 82% fewer injury-causing crashes than human drivers in the same cities on streets excluding motorways 57 L. China’s Apollo Go made 3.2m driverless rides in the first quarter of 2026 58 H; the quarter before it was 3.4m, and in the second quarter, according to a single account of a Baidu call, only about 1m L. The market is not yet profitable: the Alphabet segment of which Waymo is the main part lost $1.8bn in the second quarter of 2026 59 M. It is also vulnerable: on 31 March 2026 at least 100 Apollo Go robotaxis stopped on the roads in Wuhan at the same time, after which, according to a Bloomberg report of 29 April, Chinese regulators suspended the issuing of new permits (one source; status as of October 2026 not confirmed) 60 M.
Warehouses and delivery. Amazon bought Kiva in 2012, and in July 2025 said it had deployed its millionth robot (the counter covers different types of machine); the whole fleet works in more than 300 centres 61 M. According to the company, its DeepFleet model cuts the fleet’s travel time by 10%. According to the companies themselves, Starship’s delivery robots have made more than 10m deliveries in eight countries and Zipline’s drones more than 2m 62,63 M.
Failures mark the limit as clearly as successes do. SoftBank suspended production of the “social” robot Pepper in 2021 at an estimated 27,000 units because of weak demand 43 M. Cruise, General Motors’ robotaxi business, was closed after one of its vehicles dragged a pedestrian 64 M. In February 2026 Amazon stopped using its multi-arm Blue Jay station, which it had shown only in October 2025, and a partner of Britain’s Ocado closed three automated warehouses 65,66 H. A robot pays for itself where a task repeats millions of times in a predictable environment. Once the environment becomes unpredictable, the cost of handling exceptions eats the savings.
05 · HumanoidsHumanoids: from showpieces to shipments — tens of thousands a year, but mostly for the stage and the laboratory
In 2024 only a few thousand humanoids were shipped: by our estimate, based on the growth rates given by Omdia and IDC, about 2,300–3,000 67,68 L. In 2025, depending on the counter, between 13,000 (Omdia) and 18,000 (IDC) were shipped worldwide; IFR, which counts only full-size machines taller than 140cm, counted ~7,000 67,68,11 M. In the first half of 2026 Counterpoint counted more than 22,000, almost four times as many as a year earlier. The five biggest makers are all Chinese and hold 86% of the market: AgiBot shipped about 9,700 robots, Unitree more than 7,000 13 H. For the whole of 2026 Counterpoint expects more than 50,000, BofA 90,000 13,18 M.
What matters more than how many is what for. More than 60% of the humanoids shipped in the first half of 2026 went to entertainment, performances, research and collecting data for training. Manufacturing took 13%, warehouses and logistics 5% 13 H. In other words, humanoids are bought today mainly to be shown and trained, not to work. Unitree, in its stock-exchange prospectus, itself warns that its robots’ hands are not yet precise or durable enough for prolonged work, and that mass commercial use may arrive later than expected 69 M.
Prices are falling faster than expected. The starting price of the Unitree R1 humanoid was 39,900 yuan in July 2025 and 29,900 yuan in June 2026 2 M. The older, larger G1 cost 99,000 yuan at launch in 2024, according to Reuters L; it is a different machine, so the series is not comparable. According to Bank of America, the components of a Chinese humanoid cost ~$35,000 in 2025, and those of a Western pilot-production one $90,000–100,000; by 2030 the component cost of a Chinese humanoid could fall below $17,000 18 M. Goldman Sachs wrote in 2024 that unit cost had fallen by 40% in a year, although the bank had expected 15–20% a year 17 M.
There is more money in the industry than there are machines. According to Dealroom, humanoid start-ups raised $4.8bn in 2025 and $11bn by 6 October 2026 (Dealroom has since revised its estimates upwards: for 2025 the figure was previously $3.2bn), 63% of it Chinese 70 H. Figure AI was valued at $39bn in September 2025; by its own account, it had delivered more than 350 Figure 03 robots by April 2026 71 M. In October 2026 Tesla announced that the first production Optimus robots were coming off the lines in Fremont, but gave neither volume nor price 72 H. China’s National Development and Reform Commission warned as long ago as late 2025 of a bubble risk: more than 150 companies in the country make humanoids, and their products resemble one another 73 M.
There is also a sign of maturing. Hyundai, which owns Boston Dynamics, plans from 2028 to build 30,000 robots a year in the US and to put the production Atlas to work first on simple operations, and by 2030 on component assembly 74 M. Agility, whose Digit robots move containers in GXO warehouses, announced a merger with a SPAC that is meant to take it public (the deal has not yet closed). It showed investors an illustrative model: a $200,000 robot plus ~$20,000 for deployment and ~$36,000 a year for software and maintenance pays back for the customer in about 1.1 years, assuming 20 hours of work a day 75 M. This is the company’s own calculation, not an audit. But it is such calculations, not videos of somersaults, that will decide how many humanoids are working in the 2030s.
A caveat: who is controlling the robot
Many “autonomous” demonstrations are in fact teleoperated. Critics noted that at Tesla’s We, Robot presentation in October 2024 the robots were mostly controlled by people 76 L. According to press reports, 1X’s home robot NEO does not yet perform any task fully on its own: it is driven by the company’s remote operators, with the owner’s consent 77 L. At the half-marathon in Beijing, the times of teleoperated robots were multiplied by 1.2 51 H. Teleoperation is not deception but a way of collecting training data. But it means that behind the robot there is still a human.
06 · The brainThe robot’s brain: why teaching a hand is harder than writing text — by one estimate there is hundreds of thousands of times less data
In 2023 researchers at AI Impacts surveyed 2,778 scientists working on AI. Asked when AI could, without human help, build an online-payments website from scratch, the aggregate answer was: with 50% probability, by 2028. When could a robot wire a new house on its own? About 17 years after 2023 (by our calculation, around 2040) 78 M. The 12-year gap is Moravec’s paradox in numbers. What comes easily to a person (hands, balance, touch) is hardest for a machine.
The reason is not only mechanics but data. A language model is trained on text from the internet. A robot has to be trained on recordings of movement, and there are few of them. One of the largest shared datasets, Open X-Embodiment, was assembled in 2023 by 21 institutions from existing data; it holds more than 1m trajectories from 22 types of robot 6 H. By the estimate of a 2025 survey, measured in tokens this is roughly 200,000 times less than the corpora of large language models 5 L. The leading companies train on thousands of hours of recordings: Physical Intelligence on more than 10,000 hours, of which only 9% are open 48 H.
Data is obtained in three ways. The first is teleoperation: a person in virtual-reality goggles or a motion-capture suit steers the robot, and the recording becomes a lesson. Tesla and Figure advertised $25–48 an hour for such work (intermediaries up to $55) 14 L; according to unverified reports by Business Insider, OpenAI returned to robotics in 2025 and hired about 100 data-collection operators 46 L. The second is simulation: NVIDIA generated 780,000 synthetic trajectories for GR00T N1, the equivalent of ~6,500 hours, in 11 hours of computing, whereas its own teleoperation recordings amounted to only 88 hours 8 H. The third is video of people: in a 2026 preprint the authors report that a robot learned four tasks from 30 minutes of video of each, filmed from the human’s point of view, with no recordings from a robot, and performed them in 92.5% of attempts on average; this is the authors’ own test, with no independent replication 79 L.
Over the years architectures have converged on one type: a “slow brain and fast hands”. A vision-language model of 1–7bn parameters interprets the command and the scene 7–10 times a second, while a small network drives the joints 50–200 times a second; Figure Helix 02 adds a whole-body control layer running at 1,000Hz (by the company’s claim) 49,8,48 M. This yields new skills: the π0.7 model folded laundry on a dual-arm robot on which nobody had recorded that task — the skill was transferred from other robots 80 H.
But reliability is still laboratory-grade. Google’s RT-1 in 2022 handled new combinations of instructions in 76% of cases; GR00T N1 on one real humanoid, after fine-tuning, succeeded in 76.8% of its own tabletop tasks (the authors’ estimate); the π0.6 model at a “robot olympics” of hard tasks such as turning a sock inside out succeeded in 52% (the company’s self-assessment); Gemini Robotics 2, according to Google, between 32% and 92% depending on the task 7,8,9,10 M. The figures are not comparable: the tasks, robots and methods of evaluation differ. For a demonstration this is enough. For a production line it is not.
Investors are betting that data and scale will close the gap. Physical Intelligence was valued at $5.6bn in November 2025, Skild AI at more than $14bn in January 2026 81,82 M. In October 2026 Physical Intelligence opened a beta version of its API (by waiting list): companies fine-tune the model on their own data, and one of the first partners, Actor, showed in Utah a loader that fills a dump truck without an operator 83 M. Part of the industry, it seems, is shifting from selling robots to selling a “brain” for other people’s robots.
07 · NowWhat is being worked on in 2026: eight fronts, and the bottlenecks differ – reliability, data, permits, standards
In 2026 the IFR named five main trends in the industry: AI-based autonomy, the merging of factory and office software, humanoids that must prove their reliability, safety – and a shortage of labour that pushes everything else 84 H. Below are eight areas where money and engineers are now concentrated, and what each has already shown.
08 · LaboratoriesAlready in the labs: self-healing muscles, microrobots in the body and a hand on living muscle – by our estimate, 5 to 15 years or more from a product
To compare things as different as a microrobot and an artificial muscle, engineers use the technology readiness level (TRL) scale from 1 to 9. Levels 1–3 are research: the idea is proven on a laboratory bench. Levels 4–6 are a prototype: from testing components in the lab to operation in conditions close to real ones. Levels 7–9 are a product: from a prototype in real conditions to a system that has proved itself in service. The chart below shows 16 prototypes from 2018–2026. The TRL and the approximate year of the next milestone (the first product or first human trials) are our estimate based on published results, not data from the developers L.
Body: muscles, skin, hands. HASEL electrohydraulic artificial muscles (2018) heal themselves after electrical breakdown 97 H; their spin-off Artimus Robotics, by its own account, already supplies actuators to defence and commercial customers L. A new dielectric elastomer (Science, 2025) drives a robotic arm over a large range of motion 98 H. Electronic skin for wearable devices recovers more than 80% of its function within 10 seconds of damage; it has not yet been tested on robots 99 H. The Digit 360 sensor sees details of 7 µm at the fingertip and senses forces of ~1 mN 88 H; by its stated specifications this is close to the capability of a human finger – that is our comparison, not a measurement in the paper L. Hardware is ceasing to be the bottleneck of demonstrations; the bottlenecks are becoming cost and durability.
Microrobots in the body. A swarm of light-controlled microrobots penetrated pus more than three times better and treated sinusitis in a rabbit model (Science Robotics, 2025) 100 H. ETH Zürich’s magnetic drug capsules were guided through the blood vessels of large animals with a clinical navigation system (Science, 2025) 101 H. Robots 0.2 mm in size, with their own computer, a sensor and a 75 nW solar cell, swim in solution 102 M. All of this is animals or test tubes; the first human trials, by our estimate, should be expected in 2027–2031 L.
Living machines. Xenobots (2020) are “organisms” made of frog cells and designed by an algorithm; a year later they learned to assemble copies of themselves from loose cells 103 H. Anthrobots (2023), made from adult human lung cells, healed scratches in a layer of neurons in a Petri dish 104 H. In Tokyo in 2025 a robotic hand was moved by muscles grown from human cells – about 8 mN each 105 H. The headlines are loud, the force is meagre: a biohybrid swimmer from 2026 which, according to its authors, is faster than earlier robots built on skeletal muscle, moves at ~8 mm per second 106 M.
Flying insects and swarms. Harvard’s RoboBee took off in 2013 on a tether; in 2019 the X-Wing version flew without a tether for the first time – on solar cells under very bright light 107,108 H. In 2025 a 750-milligram MIT robot performed 10 consecutive somersaults in 11 seconds, and carried out some fast turns even in wind gusts of 160 cm/s – but again on a tether 109 H. In 13 years agility and autonomy have not been combined in one robot: there is not enough energy on board.
Brain and thought control. A non-invasive brain–computer interface let 21 healthy users experienced with such interfaces move individual fingers of a robotic hand with 81% accuracy for two fingers, 61% for three and about 46% for four 110 H. In other words, fine control of a hand through the scalp hits a ceiling; brain implants (BrainGate, UCSF) work better but require surgery 111,92 H. A neuromorphic chip navigated a route of up to 8 km, using less than 8% of the energy of conventional methods 112 H.
Robot scientists. A mobile robot chemist at the University of Liverpool carried out 688 experiments in 8 days and found photocatalysts 6 times more active than the starting ones 113 H. The autonomous laboratory A-Lab claimed 41 new materials in 17 days in 2023, but critics concluded in 2024 that none of them was new 114,115 H. A-Lab’s authors replied that the main aim was to demonstrate autonomy, not the most precise refinement of X-ray diffraction patterns (according to Chemistry World); we found no peer-reviewed resolution of the dispute M. The lesson: it is easier to robotise an experiment than to robotise the checking of its result.
09 · UkraineUkraine: ground-robot deliveries rose from 2,000 machines in 2024 to 15,000 in 2025, and the number of their logistics and evacuation missions tripled between January and August 2026
Ukraine is rapidly expanding its fleet of ground robots. According to West Point’s Modern War Institute, in 2024 Ukrainian industry delivered about 2,000 unmanned ground vehicles (UGVs) to the front, and in 2025 about 15,000 15 M. In April 2026 the Ministry of Defence announced a plan to contract 25,000 robots in the first half of the year and set a goal that all front-line logistics should be carried by robots 116 M. That is more than double the contracts of the whole of 2025 (that year’s deliveries are a different measure): there is no official figure, but by our estimate from publications, 11,000–12,500 were contracted then L. By July more than 22,000 had been contracted 16 M. The number of units using ground robots rose from 67 to 167 in four months 117 M.
The main job of these machines is not to shoot but to carry. They deliver water, ammunition and food to the “kill zone”, where vehicles with people are spotted and hit by enemy drones very quickly, and they take out the wounded. In the 3rd Assault Brigade, by its own account, robots do 80% of logistics, and near Pokrovsk and Myrnohrad about 90% 118 M. In one recorded case an evacuation robot, the Maul, picked up a wounded soldier from a depth of 64km in occupied territory 119 M. Demining robots such as the Zmiy clear about 7,000 m² a day, according to the manufacturer 120 M.
There are combat episodes too. In December 2024 the National Guard’s Khartiia Brigade attacked a fortified position near Lyptsi in Kharkiv region. The first wave consisted of ground robots and FPV drones (first-person-view drones); infantry followed 121 M. In April 2026 the president said that seven types of ground robots and drones had taken a Russian position without infantry and without losses among Ukrainian troops (the location was not named, and there is no independent verification) 122 L. An important caveat: “unmanned” does not mean “autonomous”. In March 2025 the Center for Strategic and International Studies (CSIS) found no full autonomy on the Ukrainian front: the robots near Lyptsi were teleoperated, and autonomous navigation of ground robots was still in its infancy 123 H. There is no independent data on the April 2026 episode. According to developers, autonomous guidance could raise the hit rate of FPV drones (not ground robots) from 10–20% to 70–80%; this is potential, not a measured result 123 M.
One reason is low prices and a large number of makers. The average cost of a Ukrainian ground robot, according to Foreign Policy, is under $20,000; the analyst Rob Lee puts some models at $30,000–40,000 124 M. In 2025 the US Army halted development of its Robotic Combat Vehicle programme; an anonymous Army source put the projected cost at nearly $3m a unit if a single supplier were chosen 125 L. This is a different class of machine, and a design price rather than a price paid, so the comparison with $20,000 shows only an order of magnitude: a gap of roughly two orders. More than 200 companies make ground robots in Ukraine; between January and mid-2026 alone, 67 new models were cleared for use 126,16 M. In August 2026 the Defence Procurement Agency announced its first tender under the ministry’s technical requirements (logistics and evacuation, and strike systems); a single supplier’s share is capped at 50%, but it is unclear whether per class or per lot 127 M.
The weak points are obvious too. Near infantry positions a robot may survive only one or two missions, in the estimate of the analyst Rob Lee 124 L. The VAT exemption for logistics robots ended on 1 January 2026; according to the head of the maker Tencore (TerMIT), this forced prices under current contracts to be revised by about 20% 128 M. After the change of defence minister in the summer of 2026, the target of 50,000 robots a year that had been mentioned earlier was no longer repeated 129 L.
The law lags behind practice. The UN Secretary-General called for a binding instrument on autonomous weapons to be adopted by 2026 130 M. In September 2026 the Group of Governmental Experts under the Convention on Certain Conventional Weapons (128 states parties) agreed only a draft report; a decision on negotiations is for the review conference in November 131 H. The technology is already moving in both directions: in the Netherlands a line assembling Estonian THeMIS robots for Ukraine has opened, and US special forces at the Balikatan exercises sank a target with the Ukrainian Magura naval drone 132,133 M. Ukrainian robotics has, for the first time, become mass-scale in war.
10 · Forecast2035–2040: what to expect. Two to three times as many industrial robots, but humanoids are unlikely to arrive in homes in large numbers before the 2040s
Forecasts in robotics go wrong in both directions. In 1999 UNECE and the IFR expected 870,000 industrial robots in operation in 2001, and the following year cut the forecast to 800,000 for 2002 31 M. The IFR’s own forecast for 2025 was 575,000 installations 21 H; reality exceeded it by at least 25,000 3 H. In 1997 the founders of RoboCup set a goal of beating the human football world champions by 2050 134 M. So below we distinguish three things: organisations’ forecasts, our arithmetic and our estimate of probability.
What can be calculated
Industrial robots. The IFR expects 806,000 installations in 2029, which is growth of 7–8% a year 3 H. If the number of robots in operation grows by 6–8% a year (slightly slower than the 9% of 2024 and 2025), then from 5m in 2025 it will reach 9–11m in 2035 and 12–16m in 2040 (our calculation) L. China, as in 2021–2025 (51–59% of new installations), will probably account for more than half of new installations (our estimate) L.
Surgery. If da Vinci procedures keep growing by 12–15% a year (the company forecasts 14–15% for 2026), there will be 10–13m a year by 2035 (our calculation) 52 L. This growth needs no breakthrough: only wider use of what already works.
What banks forecast for humanoids
For the same year, forecasts differ by a factor of two to five, and differ even more in what they measure: annual shipments, installed base or market value. Goldman Sachs in its base case (report of February 2024; according to unverified information, the bank revised its forecast upwards in 2026) expected more than 250,000 humanoids a year in 2030 and 1.4m in 2035, with a market of $38bn 17 M. That is roughly $27,000 per robot (our calculation) L. Bank of America in March 2026: 1.2m a year as early as 2030 and 3bn humanoids in the world by 2060, 62% of them in homes 18 M. Morgan Stanley, according to secondary accounts: about 1bn humanoids by 2050 and a market of ~$5trn including repairs and supply 135 L. Citi: about 650m by 2050 136 M. Barclays in its most optimistic scenario puts the market at $200bn in 2035, against $2–3bn now 137 M; its base case, according to one account, is at least $40bn, close to Goldman 137 L.
Test the forecasts yourself. The calculator below works out how many humanoids will be in service in 2040 under your assumptions. For reference: BofA’s forecast for 2030 of 1.2m shipments, against 90,000 in 2026, implies growth of roughly 91% a year (our calculation; Fortune gives 86%) L. By default the calculator starts from 50,000 in 2026 (Counterpoint), so for 2030 it gives 221,000, not 1.2m.
Calculator: how many humanoids there will be in 2040
Set shipments for 2026, the growth rate and the robot’s service life. The model is simple: each year new robots are added, and those that have reached the end of their service life drop out. The defaults are our assumptions: growth of 45% a year to 2035 gives 1.4m shipments in 2035, as in Goldman’s base case, while growth of 15% after 2035 and a service life of 7 years are arbitrary. “In service” includes demonstration and research robots: more than 60% of shipments in the first half of 2026 were for entertainment, research and data collection 13 H. The chart scale is logarithmic.
When a robot is cheaper than a human
The best clue to where humanoids will appear first is wages. In Agility’s illustrative model (robot $200,000, deployment ~$20,000, software and service ~$36,000 a year, service life 5 years) average costs are ~$80,000 a year, or ~$11 per hour of work at 20 hours a day (our calculation) 75 L. The calculator below adds 90% uptime and productivity of 70% of a human’s (our assumptions), so for this model it shows ~$17 per human-equivalent hour. An hour of labour in US manufacturing costs the employer $48.62 138 H, in Poland ~€19 on average across the economy 139 M, and in Ukrainian industry ~€4.5 by our calculation from manufacturing wages plus the 22% unified social contribution 140 L. So humanoids will pay for themselves first in factories and warehouses where labour is dear or scarce: the US, western Europe, South Korea and Japan (our estimate) L. China is an exception: there the robots themselves are cheap, and the five largest humanoid makers are Chinese, with a share of 86% of shipments in the first half of 2026 13 H.
Calculator: robot versus human
Choose a type of robot and a country. The calculator compares the cost of an hour of human labour with the cost of an hour of the robot’s work (converted into a human-equivalent hour, allowing for productivity) and works out how many years it takes for the purchase to pay back. The parameters for Digit come from Agility’s model, the G1 price is from Unitree, the NEO subscription is from 1X; the rest (the productivity of all robots, the working hours and service life of the G1 and NEO, G1 servicing) and the 90% uptime are our assumptions. For a home robot, compare not with a factory wage but with the cost of domestic labour in your country.
What this means for human work
The best-known estimate is by Daron Acemoglu and Pascual Restrepo for the US: one additional robot per thousand workers lowers the employment rate by 0.18–0.34 percentage points and wages by 0.25–0.5% 141 H. But in Germany, according to a study by Dauth and colleagues, robots reduced manufacturing employment mainly because fewer young people were hired, not because of layoffs, and jobs in services made up for it 142 M. In France, firms that bought robots grew while their competitors shrank, and employment in the industry as a whole fell 143 M. So the effect depends on the country and the industry; in the German data it fell hardest on those just entering the profession (our interpretation).
Our estimate: what is plausible by 2035–2040
| Area | What, by our estimate, is plausible | Confidence | What could change the picture |
|---|---|---|---|
| Factories | 9–11m industrial robots in operation in 2035 and 12–16m in 2040; China accounts for more than half of new installations (extrapolation of the trend; the IFR has no official forecast for these years) | Medium | A global recession, trade wars, a sharp fall in the price of humanoids, which would partly replace conventional manipulators |
| Warehouses | The share of automated operations in large warehouses in rich countries will grow substantially; humans will handle mainly the “long tail” of awkward goods (our estimate, with no quantitative basis) | Medium | The pace at which robots learn to pick up any item, rather than three-quarters of the range |
| Surgery | 10–13m operations a year with da Vinci systems by 2035 (our extrapolation); some stages of operations will probably be performed autonomously under a surgeon’s supervision | Medium | Regulators: a fully autonomous operation on a human before 2040 is unlikely |
| Robotaxis | Robotaxis operate in many cities in the US and China, and in Europe the first commercial launches have begun; according to Goldman, there will be about 6m robotaxis in the world by 2035, 3.1m of them in China 144 M. Their share of all trips worldwide will probably remain small (our estimate) | Medium | Accidents and failures that halt permits (such as Cruise in 2023 or the mass outage of Apollo Go in Wuhan in 2026, after which, according to Bloomberg, China suspended new permits), or, conversely, a breakthrough in vehicle prices |
| Humanoids at work | From hundreds of thousands to 1.4m a year in 2035 (the upper bound is Goldman’s base forecast), mainly in China, in factories and warehouses; price below $20,000 (our estimate) | Low | Reliability of hands and models: success rates must rise from 50–80% to 99% and above |
| Home robots | The first hundreds of thousands of humanoids in wealthy households after 2030, often with remote operators; at scale, in the 2040s | Low | Privacy (a remote operator sees your home), safety around children, price |
| Medical microrobots | First trials in humans in the late 2020s, first approved products in the 2030s | Low | Safety and the long road to regulatory approval |
| War | Front-line logistics and evacuation mainly by robots; the share of autonomous functions is growing | Medium | The end of the war, international limits on autonomous weapons; in 2021 US intelligence described autonomous lethal weapons as a possible scenario for 2040 145 H |
| Labour | Fewer entry-level hires in industry and logistics rather than mass layoffs; the greatest pressure is in countries with expensive labour | Medium | The pace at which humanoids get cheaper, and demography: ageing speeds up automation |
The AI Impacts survey and the estimates of BofA analysts agree on one thing: robots will come to the home later than AI to the office. In the 2023 AI Impacts survey, AI researchers give a 50% chance that a robot will install the electrical wiring in a new home on its own in about 17 years, that is, around 2040 (our calculation from 2023) 78 M. BofA’s analysts, the most bullish on humanoid numbers, themselves call the home humanoid a “story of the 2040s” 18 M. And in August 2026 the journal Science Robotics devoted a special issue to humanoids, in which a commentary by Murphy asks whether the ubiquitous humanoids of 2035 promised by the film “I, Robot” are more fiction than science 146 M.
11 · LimitsHonesty and limitations
The strongest counter-argument
It can be argued that we underestimate the speed. In three years language models went from curiosity to a tool that, by the companies’ own figures, hundreds of millions of people use every week L. Robotics has acquired the same “brain”; funding for humanoids in the first nine months of 2026 was already more than double that of the whole of 2025 ($11bn against $4.8bn) 70 H, and the starting price of Unitree’s cheapest humanoid fell in two years from 99,000 to 29,900 yuan (though these are different models: the new one is smaller and simpler) 2 M. If data from simulation and videos of people really replace teleoperation, reliability could jump as sharply as the quality of machine translation once did. Then BofA’s forecasts would turn out to be cautious, and the home robot would arrive ten years sooner. We consider this scenario possible but not the base case: unlike text, a robot’s mistake costs broken things and injuries, and regulators’ permits do not speed up along with the models.
Limitations of this briefing
- Different counts. The number of humanoids in 2025 ranges from 7,000 (IFR, full-size only) to 18,000 (IDC). Robot density in China “fell” from 470 (2023) to 166 (2024) because the IFR revised the number of people employed in manufacturing, by which the number of robots is divided; the stock of robots itself (about 2m) did not shrink. We show the spread, not a single number.
- The 2000–2009 series for industrial robots was measured off an IFR chart, not taken from a table; the values are approximate.
- Banks’ forecasts are mostly known from press accounts: the pages of Goldman Sachs, Morgan Stanley and Barclays did not open when we checked. Goldman’s forecast is dated February 2024; according to unverified information, the bank raised it in 2026.
- Company data (Figure’s deliveries, Vulcan’s productivity, Waymo’s safety, Agility’s payback model) are the companies’ own claims, without independent audit.
- Ukrainian data on the number of robots come from official statements and analysis, not from delivery tables; some combat episodes are known only from the words of military personnel or the president. The Ministry of Defence counts only logistics and evacuation missions.
- TRL and timelines for laboratory prototypes are our estimate based on published results, mostly paper abstracts.
- Calculators are models with explicit assumptions. No manufacturer publishes a humanoid’s productivity relative to a human, so its default value is our assumption.
- What is missing. We did not examine in detail agricultural robots, underwater and industrial drones, exoskeletons for workers, or the market for components (actuators, reducers, magnets); we did not check China’s share of it.
12 · QuestionsFrequently asked questions
How many robots are there in the world in 2026?
According to the IFR, 5m industrial robots were working in factories at the end of 2025, and more than 600,000 new ones were installed during the year, 59% of them in China. Some 34.2m household robots, mostly vacuum cleaners, were sold in 2025. Humanoids are still counted in tens of thousands: Counterpoint counted more than 22,000 shipped in the first half of 2026.
When will a humanoid robot be in every home?
Not soon. The first home humanoids can already be ordered (1X NEO: $20,000 or $499 a month), but we found no confirmed deliveries to homes, and a remote operator does part of their work. Even Bank of America, the most bullish on humanoid numbers, calls the household humanoid “a 2040s story”, and researchers in the AI Impacts survey give a 50% chance that a robot can wire a new house on its own only around 2040.
Will robots take people’s jobs?
Partly and unevenly. In America, Acemoglu and Restrepo estimate that one more robot per thousand workers lowers the employment-to-population ratio by 0.18–0.34 percentage points and wages by 0.25–0.5%. In Germany robots cut manufacturing jobs mainly through fewer hires of young people rather than lay-offs. Robots pay back first where labour is dear: an hour of manufacturing work costs an American employer $48.62, and, by our estimate, about €4.5 in Ukraine.
Why does China lead in robotics?
China installs 59% of the world’s new industrial robots, and the five biggest humanoid makers, with 86% of the market, are Chinese. The reasons are the scale of its own industry, cheaper components (Bank of America puts the bill of materials of a Chinese humanoid at ~$35,000 against $90,000–100,000 for a Western one) and state policy: the 15th Five-Year Plan (2026–2030) reportedly names humanoids a priority. At the same time a Chinese regulator warns of a bubble: more than 150 companies make humanoids.
What is Ukraine doing in robotics?
Ukraine quickly became a mass proving ground for ground robots. The army received ~2,000 unmanned ground vehicles in 2024 and ~15,000 in 2025, and contracted more than 22,000 in the first seven months of 2026. In August 2026 robots carried out 25,143 logistics and evacuation missions. More than 200 Ukrainian firms build them, at an average price of up to $20,000. And in the 1970s, according to the department’s own history (we found no independent confirmation), Kyiv’s Institute of Cybernetics built TAIR, an autonomous robot with a neural network.
13 · MoreFurther reading
14 · SourcesSources and methodology
A number in the text leads to its source. Confidence labels: H — primary source (statistics, scientific paper, a company’s financial reporting), M — secondary source or one cross-checked against several, L — estimate, claim by a company or official, a single unconfirmed source or our calculation. The data were gathered and cross-checked by 18 research agents using search; key figures were checked against the source and refuted ones removed; some of the banks’ forecasts and companies’ claims could not be checked against the primary source (see “Honesty and limitations”). All sums are in US dollars unless stated otherwise.
- IFR — History of robots. ifr.org H
- Unitree — R1 and G1 price list. unitree.com H
- IFR — World Robotics 2026, press release. ifr.org H
- Hans Moravec, Mind Children, 1988 (Wikipedia). en.wikipedia.org M
- Survey of vision-language-action models, 2025: robot data volume. arxiv.org M
- Open X-Embodiment, 2023. arxiv.org H
- Brohan et al., RT-1, 2022. ar5iv.labs.arxiv.org H
- NVIDIA — GR00T N1, 2025. arxiv.org H
- Physical Intelligence — Robot Olympics test of π0.6. pi.website H
- Google DeepMind — Gemini Robotics 2, 2026. deepmind.google M
- IFR — World Robotics 2026: service robots. ifr.org H
- IDC — cleaning robot shipments 2025. idc.com M
- Counterpoint Research — humanoid shipments, H1 2026. counterpointresearch.com H
- Data-collection operator postings at Tesla, Figure (Business Insider and others). businessinsider.nl L
- Modern War Institute (West Point) — Ukraine’s ground robots. mwi.westpoint.edu M
- Defence Leaders — over 22,000 UGVs contracted. defenceleaders.com M
- Goldman Sachs Research — humanoid robot market, February 2024. goldmansachs.com M
- Fortune — Bank of America humanoid forecast, 13 March 2026. fortune.com M
- IFR — World Robotics 2010. web.archive.org H
- Interact Analysis — industrial robot market 2025–2030. interact-analysis.prezly.com M
- IFR — World Robotics 2025, executive summary. ifr.org H
- IFR — World Robotics 2006: robot prices 1990–2005. cep.lse.ac.uk H
- Interact Analysis (via a vendor guide) — Chinese robot prices. evsint.com L
- IFR — robot density, press release, 8 April 2026. ifr.org H
- iRobot — Chapter 11 filing, December 2025. theregister.com M
- Al-Jazari, Book of Knowledge of Ingenious Mechanical Devices (Wikipedia). en.wikipedia.org M
- Jacquard machine (Wikipedia). en.wikipedia.org M
- Karel Čapek, R.U.R. (Wikipedia). en.wikipedia.org M
- Isaac Asimov, Runaround, 1942 (Wikipedia). en.wikipedia.org M
- US patent 2,988,237, Programmed Article Transfer. patents.google.com H
- UN/ECE and IFR — World Industrial Robots, 1998–2000 summaries (Industrial Robot). emerald.com M
- SRI — Shakey the robot. sri.com H
- Waseda University — WABOT-1. waseda.jp H
- History of the Biocybernetics Department, Institute of Cybernetics, Kyiv (TAIR, MAVR), cyberneticzoo compilation. cyberneticzoo.com M
- DARPA Grand Challenge 2004–2007 (Wikipedia). en.wikipedia.org M
- Intuitive Surgical — Form 10-K, FY2025. sec.gov H
- NASA — Opportunity. science.nasa.gov H
- Universal Robots / IFR — UR5, 2008. ifr.org H
- ROS: first commit and version 1.0 (Wikipedia). en.wikipedia.org M
- Amazon buys Kiva Systems, 2012. techcrunch.com M
- IFR — World Robotics 2019, executive summary. web.archive.org H
- DARPA Robotics Challenge Finals, 2015. en.wikipedia.org M
- SoftBank Pepper — production pause (Reuters). en.wikipedia.org M
- Levine et al., learning hand-eye coordination for grasping, 2016. arxiv.org H
- OpenAI — solving a Rubik’s cube with a robot hand, 2019. arxiv.org M
- OpenAI: robotics team closed and rebuilt. venturebeat.com M
- Google DeepMind — RT-2, 2023. arxiv.org H
- Physical Intelligence — π0, 2024. arxiv.org H
- Figure — Helix, 2025. figure.ai H
- Google DeepMind — Gemini Robotics 1.5. deepmind.google M
- Beijing humanoid half marathon 2025–2026. english.beijing.gov.cn H
- Intuitive Surgical — Q3 2026 results. sec.gov H
- Medtronic — FDA clearance for Hugo, December 2025. news.medtronic.com H
- J&J MedTech — OTTAVA De Novo, 2026. rttnews.com M
- Shanghai to Kuwait telesurgery, MicroPort Toumai. pubmed.ncbi.nlm.nih.gov H
- Alphabet, Waymo — weekly paid rides (via CNBC, AFP). webull.com M
- Waymo Safety Impact. waymo.com H
- Baidu — Q1 2026 results. sec.gov H
- Alphabet — Other Bets segment, Q2 2026. webull.com M
- Apollo Go outage in Wuhan and permit freeze (Bloomberg and others). techcrunch.com M
- TechCrunch — Amazon deploys its one millionth robot, 2025. techcrunch.com M
- Starship Technologies — 10 million deliveries, 2026. starship.xyz H
- Zipline — 2 million deliveries, 2026. zipline.com H
- Cruise: permit suspension and shutdown. en.wikipedia.org M
- Amazon — Blue Jay, update of 25 February 2026. aboutamazon.com H
- Ocado — FY2025 results. cdn.prod.website-files.com H
- Omdia (via TechNode) — humanoids 2025. technode.com M
- IDC (via IT Brief) — humanoids 2025. itbrief.co.uk M
- Unitree — IPO prospectus (via press). hwbusters.com M
- Dealroom — venture investment in humanoids. dealroom.co H
- Figure AI — Figure 03 and the BotQ factory. en.wikipedia.org M
- Tesla — Q3 2026 shareholder update. assets-ir.tesla.com H
- China’s NDRC — humanoid bubble warning. unite.ai M
- Hyundai at CES 2026 (via InsideEVs). insideevs.com M
- Agility Robotics — Digit unit economics in SPAC materials, September 2026. letsdatascience.com M
- Tesla Optimus: promises and the We, Robot event (Wikipedia and press). en.wikipedia.org L
- 1X — NEO terms. businesswire.com M
- Grace et al., Thousands of AI Authors on the Future of AI, 2024. arxiv.org M
- HumanEgo, arXiv preprint, 2026. arxiv.org H
- Physical Intelligence — π0.7, 2026. pi.website H
- Physical Intelligence — Series B, November 2025. komo.ai M
- Skild AI — January 2026 round (Dow Jones). marketscreener.com M
- Physical Intelligence — Pi API, October 2026. pi.website H
- IFR — top five robotics trends 2026. ifr.org H
- Google DeepMind — Genie 3, 2025. bdtechtalks.com M
- Meta — V-JEPA 2, 2025. ai.meta.com M
- Sharpa Wave (CNX Software). cnx-software.com M
- Meta and GelSight — Digit 360, 2024. arxiv.org H
- Elon Musk on the Optimus price (Davos, January 2026). optimusk.blog L
- Amazon — Vulcan robot, 2025. aboutamazon.eu M
- Kim et al., SRT-H, Science Robotics, 2025. pubmed.ncbi.nlm.nih.gov H
- UCSF, ECoG interface and robotic arm, Cell, 2025. pubmed.ncbi.nlm.nih.gov H
- UC San Diego — Surgie humanoid, Nature, 2026. hoodline.com M
- Zoox — paid rides in Las Vegas. automotiveworld.com M
- Lawsuit by Figure’s former head of product safety, November 2025. en.wikipedia.org M
- China’s humanoid policy 2025–2026. investinchina.chinadaily.com.cn L
- Acome et al., HASEL, Science, 2018. doi.org H
- Shi et al., dielectric elastomer muscle, Science, 2025. doi.org H
- Lee et al., self-healing electronic skin, Science Advances, 2025. doi.org H
- Yu et al., microrobots for sinusitis, Science Robotics, 2025. doi.org H
- ETH Zurich — magnetic microrobots, Science, 2025. pubmed.ncbi.nlm.nih.gov H
- University of Pennsylvania and University of Michigan — microrobots, 2025. engineering.upenn.edu M
- Kriegman et al., xenobots, PNAS, 2020–2021. doi.org H
- Gumuskaya et al., anthrobots, 2023. doi.org H
- University of Tokyo — biohybrid hand, 2025. doi.org H
- NUS — OstraBot, Nature Communications, 2026. cde.nus.edu.sg M
- Ma et al., RoboBee, Science, 2013. doi.org H
- Jafferis et al., RoboBee X-Wing, Nature, 2019. doi.org H
- Hsiao et al., MIT, Science Advances, 2025. news.mit.edu H
- Carnegie Mellon — EEG control of robotic fingers, 2025. pmc.ncbi.nlm.nih.gov H
- Hochberg et al., BrainGate, Nature, 2012. pubmed.ncbi.nlm.nih.gov H
- Hines, Fischer et al., LENS, Science Robotics, 2025. doi.org H
- Burger et al., a mobile robotic chemist, Nature, 2020. doi.org H
- Szymanski et al., A-Lab, Nature, 2023. doi.org H
- Critique of A-Lab, PRX Energy, 2024. journals.aps.org H
- Defense News — Ukraine orders 25,000 ground robots. defensenews.com M
- UNITED24 Media — units using ground robots. united24media.com M
- 3rd Assault Brigade — share of logistics by robots. mwi.westpoint.edu M
- Casualty evacuation by the Maul robot. mwi.westpoint.edu M
- Zmiy demining robots. mwi.westpoint.edu M
- Kyiv Independent — the Lyptsi assault, December 2024. kyivindependent.com M
- Statement by the President of Ukraine, 13 April 2026. the-decoder.com L
- CSIS — AI-enabled autonomous warfare in Ukraine, March 2025. csis.org H
- Foreign Policy — the cost of Ukraine’s ground robots. foreignpolicy.com M
- Breaking Defense — US Army halts the RCV, May 2025. breakingdefense.com M
- Number of Ukrainian ground-robot makers (Brave1). defensenews.com M
- Defence Ukraine — first specification-based UGV tender, August 2026. censor.net M
- End of the VAT exemption for UGVs, January 2026. sofx.com M
- Change of defence minister, 2026. euromaidanpress.com M
- UN Secretary-General, A New Agenda for Peace, 2023. stopkillerrobots.org M
- French MFA — GGE on lethal autonomous weapons, September 2026. diplomatie.gouv.fr H
- Milrem Robotics and VDL — THeMIS line in the Netherlands. defensenews.com M
- Defense News — Magura at Balikatan 2026. defensenews.com M
- RoboCup — the 2050 goal. 2017.robocup.org M
- Morgan Stanley Research — humanoid market to 2050 (via secondary sources). morganstanley.com L
- Citi GPS — The Rise of AI Robots, 2024. citigroup.com M
- Barclays Research — Equity Gilt Study 2026. businesswire.com M
- BLS — Employer Costs for Employee Compensation, June 2026. bls.gov H
- Eurostat — hourly labour costs, 2025. getsix.eu M
- State Statistics Service of Ukraine — average wage; labour-cost calculation. gtai.de M
- Acemoglu and Restrepo, Robots and Jobs, NBER, 2017. nber.org H
- Dauth et al. — robots and employment in Germany. boeckler.de M
- French firm-level evidence on robot adoption 2010–2015. repec.cepr.org M
- Goldman Sachs — robotaxis to 2035 (2026 update). goldmansachs.com M
- US National Intelligence Council — Global Trends 2040. archive.dni.gov H
- Science Robotics — special issue on humanoids, August 2026. doi.org M