Where robots and AI stand in 2026

In January, Jensen Huang opened the robotics year with one sentence: “The ChatGPT moment for robotics is here.” Nine months later, the money and the models suggest he was early rather than wrong. Robots that learn from video and language improve every few months. The robots that do paid work every day are, with few exceptions, still the ones that do not look like us.

Here is what is interesting right now, what is real, and what is still a demo.

New brains for machines

The change behind the headlines is in the software. A classic robot is programmed step by step for one task. Robot foundation models learn from demonstrations, video and language, much as chatbots learned from text, and one model can drive different machines. Google DeepMind describes the old world plainly: “Most robots are pre-programmed or teleoperated for narrow, repetitive task sequences.”

In July, DeepMind released Gemini Robotics 2, which controls a whole humanoid “from feet to fingertips”, shown on Apptronik's Apollo 2, with Apptronik, Boston Dynamics and Agile Robots as partners. Its planning model is open to developers, while the model that moves the robot goes to early partners only. DeepMind also published its limits. Picking things up from the floor worked 45.7 percent of the time and screwing in a bulb 36 percent, and it writes that dexterous manipulation with several fingers “remains challenging”.

Nvidia supplies much of the layer underneath: the chips, and open models such as Isaac GR00T N1.6 for humanoids and Cosmos for generating synthetic training data.

The humanoids

Figure took its robot into 30 homes in the Bay Area where it had collected no data, for three jobs: tidying living rooms, folding towels and making beds. A policy trained from scratch succeeded in 9 percent of these attempts. One pretrained on Index, Figure's global dataset of human behaviour on video, succeeded in 56 percent. Figure says it has committed 3.5 billion dollars of compute to training its model, Helix. These are the company's own numbers, and 56 percent is progress, not yet a product.

Tesla is setting up an Optimus production line in Fremont, where the Model S and X used to be built. On the July earnings call, Elon Musk said production would start soon, that the early ramp would be “quite flat and long”, and that the aspiration for a later generation is 10 million units a year. He was also frank about the field: “those demonstrations you're seeing are preprogrammed or remote controlled”, and “there is no humanoid robot that is actually able to do generalized tasks.”

1X sells NEO, a humanoid for the home, for 20'000 dollars with priority delivery in 2026, or later on a subscription of 499 dollars a month. Deliveries start mainly in the US. For chores it does not know yet, owners can schedule a 1X Expert to guide it.

The volume is in China. Unitree sold 5'500 humanoids in 2025, roughly a third of global humanoid sales, Rest of World reported as the company prepared to list in Shanghai, and over 70 percent of them went to research and education. Its average humanoid price fell from about 593'400 yuan in 2023 to 167'600 yuan, about 25'000 dollars, in 2025.

What is actually working

The robots that earn their keep are mostly not humanoids. The International Federation of Robotics counts a record 5 million industrial robots now working in factories, after more than 600'000 were installed in 2025, 59 percent of them in China. Almost 250'000 professional service robots were sold that year, nearly half of them for transport and logistics, along with almost 34.2 million consumer robots.

Against that, about 7'000 full size humanoids were sold worldwide in 2025. The IFR notes that “most current applications remain specialized and often require human teleoperation”, and names the barriers: safety standards, high training and maintenance costs, and the lack of a strong business case.

There are exceptions. Agility Robotics' Digit, a humanoid on two legs, works at nine customer sites, including GXO, Schaeffler, Toyota's plant in Canada and Mercado Libre, and Agility says it has more than 300 million dollars in multi year orders for its new model. And the largest robot fleet most people will meet has no arms at all: Waymo runs roughly 4'000 robotaxis and averages 500'000 paid rides a week in 15 US cities, up from three cities in September 2024.

How to read a robot demo

Robot videos are now a marketing genre, so it helps to ask five questions. Was the robot acting on its own, or was a person steering it from elsewhere? How many attempts did the clip take, and at what speed was it played? Were the room, the objects and the task new to the robot, and what is the success rate over a hundred tries rather than one? And who pays for the work today?

A robot that folds a towel on camera is a research result. A robot that moves totes, shift after shift, for a paying customer is a deployment. Both matter, but only the second belongs in a business plan.

What it means for a company

For most companies, the useful robots of the next few years will be specialised: arms, carts, drones, cleaning machines and vehicles that do one job reliably. Humanoids are worth watching, and worth piloting where the work is repetitive and the setting controlled. On today's evidence, they are not yet something to plan a business around.

The same technology is reaching physical products earlier, in the design work around them. In Forge, which takes a company's own product from the idea to a specification a manufacturer can build from, AI researches, sketches and models in 3D at speed, people make every decision, and specialists join where a product needs them. The robots will reach the factory floor in their own time. The way a physical product is conceived is changing now.