Robots and AI: How We Got Here, What It Means and What Comes Next

For most of the last sixty years, robots meant one thing: a caged steel arm repeating the same weld on the same car door, all day, without complaint. Artificial intelligence meant something separate, a research field that lived in labs and, later, on screens. In 2026 those two stories have finally merged. The machines are getting a brain, and the brain is getting a body.
That shift is real, but it is also surrounded by hype, confusing numbers and a lot of confident predictions that do not agree with each other. This piece walks through how we got here, where things actually stand today, who is affected and what the sensible expectations are for the next decade.
From programmed arms to learning machines
The first industrial robot went to work on a General Motors line in 1961. It did what it was told, in a fixed order, in a space where no person was allowed to wander. For decades that was the template. Robots were fast, strong and precise, but they were also blind to anything that had not been written into their code. Move the part three centimetres to the left and the whole process failed.
AI followed a different path. Chess programs, speech recognition and early machine vision improved slowly, then quickly once neural networks met cheap computing power in the early 2010s. Language models arrived at scale in 2022 and changed what the public thought a computer could do. For a while the two fields mostly ignored each other. Software could write an essay but could not fold a towel. Robots could lift an engine block but could not tell a ripe tomato from a bruised one.
The gap is closing because researchers started treating physical action as one more thing a model can output. Instead of hand-coding every movement, they train a single system on images, language and recorded actions, and let it work out how to connect them. Google DeepMind described this approach in its Gemini Robotics models, which it presents as a way to add physical actions to a model that already understands text and images. Nvidia has pushed the same idea from the infrastructure side with open robot models and simulation tools, so that a machine can practise millions of times in a virtual world before it touches a real object.
Where things stand in late 2026
Start with the unglamorous part, because that is where the money and the working machines are. The International Federation of Robotics (IFR) published its World Robotics 2026 report in September. Factories worldwide installed 603,307 industrial robots in 2025, an 11% increase and the first time annual installations have climbed clearly past the 600,000 mark. The total number of industrial robots in operation reached a record 5 million.
China dominates this picture. It installed 354,000 units in 2025, which is roughly three out of every five robots deployed anywhere. The United States installed about 38,400, up 12%, and moved ahead of Japan as the second-largest market. The IFR expects installations to rise another 9% to 655,000 in 2026 and to reach 806,000 by 2029. It points to labour shortages and better vision and sensing as part of the reason.
One line in the IFR's own summary deserves more attention than it usually gets. The organisation says general-purpose "physical AI" systems, the kind that learn from simulation and experience, are not yet industry-grade. In other words, the robots doing real, paid work at scale today are still mostly the classic kind. The learning machines are arriving, but they are arriving in pilots and early deployments, not in the bulk of the 5 million.
The humanoid boom, and how to read the numbers
Humanoid robots get the headlines, so they also get the most exaggeration. The honest answer to "how many are out there?" is that nobody agrees, and the gap between estimates is wide.
Counterpoint Research reported that more than 22,000 humanoids shipped worldwide in the first half of 2026, nearly four times the figure a year earlier. Other trackers put the first-half total anywhere from about 19,000 to 25,000. A state-backed report presented around the World Robot Conference claimed a much higher number, and independent counts do not support it. If you see a single confident figure, ask who produced it and how they counted.
What most sources do agree on is who is shipping. Chinese companies account for the large majority of units. Counterpoint estimated that AgiBot shipped around 9,700 humanoids in the first half and Unitree more than 7,000, with UBTech, Galbot and Leju also in the top five. American companies are better known in the press and raise bigger rounds, yet they ship far fewer machines to outside customers. Figure AI has reportedly been valued at $39 billion, while its independently tracked sales are tiny by comparison.
Tesla's Optimus is the clearest example of the gap between promise and proof. The company has talked about large production targets, and some outlets repeat figures of thousands of units. Tesla has not published an audited production count, and several trackers say the robot is not available for purchase by anyone outside the company, with outside sales not expected before late 2027. Treat claims about Optimus volume as unconfirmed until Tesla publishes real numbers.
It also matters what these robots are doing. Counterpoint's breakdown suggests that entertainment, research, education and data collection made up more than 60% of demand in the first half. That is not a failure, because data collection is exactly how these machines get better. But it is a long way from a humanoid cleaning your kitchen. Home robots are still at the pre-order and early-trial stage. Companies such as 1X have taken deposits and promised deliveries, and UBTech began home deliveries in China in September at prices starting around $16,500, but anything resembling a mass market is further out.
In factories the evidence is more concrete. Figure has run its robots on a BMW production line in South Carolina, with reported figures of more than a thousand hours and tens of thousands of parts loaded. Agility Robotics has deployed its Digit robot in warehouses. These are narrow tasks, moving totes and loading parts, done in controlled spaces. That is the realistic near-term job description for a humanoid: a flexible machine for dull, repetitive material handling.
Why the "brain" matters more than the body
Robot bodies have been good enough for years. What held them back was adaptability. A system that handles one task in one setting is a product; a system that handles a new task after a short demonstration is a platform. Foundation models are the reason people believe the second one is possible.
DeepMind released Gemini Robotics-ER 1.6 in April 2026, a model focused on spatial reasoning and on understanding a scene from several camera views at once. It was developed with Boston Dynamics and offered to developers through Google's AI tools, so a small startup can use it without training something comparable from scratch. A newer Gemini Robotics 2 followed in the summer. Nvidia's GR00T models and its Cosmos world models aim at a similar goal from a different direction.
The practical effect is a change in who can build a useful robot. When the intelligence is a service you can call, the hard part moves toward hardware reliability, safety certification and getting enough real-world data. That favours companies with factories, customers and cash, which is part of why the competition looks more like a race between industrial ecosystems than between clever labs.
There is a catch, and researchers say it openly. Language models can make mistakes that cost nothing more than an embarrassing paragraph. A robot that misjudges the weight of an object or the position of a person's hand can break something or hurt someone. Reliability in the physical world has to be far higher than reliability in a chat window, and nobody has shown it can be reached cheaply for open-ended tasks.
What it means for work
The employment question is where predictions diverge most, so it helps to separate what each number actually measures.
- Exposure is not replacement. A widely repeated Goldman Sachs estimate from 2023 said generative AI could expose the equivalent of 300 million full-time jobs to automation. Exposed means some tasks could be automated, not that the job disappears.
- Displacement estimates are smaller and keep shifting. Goldman's later baseline put displacement at about 6 to 7% of US workers if AI is widely adopted. In a note reported on 22 October 2026, the bank said around 9% of US workers could be displaced over a decade, and that a recession could speed that up.
- Net effects can be positive on paper. The World Economic Forum's Future of Jobs Report 2025 projected 92 million jobs displaced and 170 million created by 2030, a net gain of 78 million. Those figures cover all major trends, not AI alone, and the new jobs will not necessarily be in the same places or require the same skills as the lost ones.
Two patterns stand out. First, the first wave hits tasks done at a desk, such as data entry, claims processing, routine customer service and junior analysis, more than it hits plumbers or electricians. Physical dexterity in unpredictable settings is still one of the hardest things to automate. Second, entry-level workers feel it first. When a model can draft the first version of a report, the traditional apprenticeship ladder gets shorter, and some analysts worry about what that does to training the next generation of experts.
Robots add a geographic layer. Factory and warehouse automation lands in specific towns, so national averages can look calm while one community loses most of its jobs. Yet labour shortages are also a driver of adoption, so in some industries the machines fill roles employers could not staff.
The risks people are actually arguing about
Beyond jobs, three questions will shape how this goes.
Safety around people. Industrial robots work inside fences. A humanoid that shares a corridor with workers needs different standards, and regulators in the US, EU and elsewhere are still working out what to require. Expect rules on testing, incident reporting and, eventually, liability when a learning system makes a bad decision.
Concentration of power. If a few firms provide the models, the chips and the training data, they set the terms for everyone building on top. Shared, open models help, but the best-funded companies still hold the advantage in compute and real-world data.
Geopolitics. China leads in installed industrial robots and in humanoid shipments, while the US leads in frontier AI models and chips. Trade restrictions, subsidies and import rules are already part of the story, and they will affect prices and availability more than any technical breakthrough this decade.
What to expect next
Forecasts for humanoids vary enormously. Different analysts project annual shipments in the hundreds of thousands by 2030, and others far lower. Treat all of them as scenarios, not schedules. They depend on three things that nobody can yet guarantee: costs falling quickly, reliability reaching the level customers demand, and enough useful work to justify the purchase.
Here is a more grounded way to think about the next five years.
- Industrial robots keep growing steadily. The IFR's forecast of 806,000 annual installations by 2029 is a mainstream, data-based projection, and it needs no breakthrough to come true.
- Humanoids find narrow jobs first. Material handling, parts feeding and inspection in factories and warehouses are the likeliest early wins.
- AI makes existing robots more flexible. The quieter change is that ordinary arms and mobile robots become easier to set up and re-task, because the software understands what it sees.
- Homes come last. Messy, unpredictable, private spaces are the hardest environment. Early home robots will be limited, supervised and expensive.
The bottom line
The fair summary is neither "robots are about to take every job" nor "it's all hype." Millions of industrial robots already work, their numbers are growing at a record pace, and AI is making the next generation more adaptable. Humanoids are real products now, shipped mostly by Chinese makers and used mostly for research, training and early factory tasks.
What is still missing is proof that general-purpose machines can be cheap, safe and reliable enough to leave the pilot stage. The next two or three years of independent shipment data, audited production numbers and real customer results will say more than any keynote. If you want to follow the story, watch those, and be cautious with any round number that arrives without a source.
Sources: International Federation of Robotics, World Robotics 2026 press releases and executive summary (24 September 2026); Counterpoint Research H1 2026 humanoid shipment figures as reported by trade press; Google DeepMind (Gemini Robotics announcements, March 2025 and April 2026); Nvidia robotics model announcements; Goldman Sachs Research estimates as reported by Fortune and IBTimes (January and October 2026); World Economic Forum, Future of Jobs Report 2025. Shipment and production figures vary between trackers and several company claims are unaudited.
تعليقات
إرسال تعليق