Did robotics become an AI field?

In July 2026, Google DeepMind released Gemini Robotics 2, a model meant to run a robot’s whole body. In April 2026, Physical Intelligence published π0.7, a single model that steers many robots through many tasks. The story everyone tells: robotics has become a branch of artificial intelligence. We can test that against the research itself. We gathered 1,854,756 robotics papers by following their citations, and for each we hold a title and, for most, an abstract. If robotics turned into AI, the papers should show it.

For fifteen years the answer barely moved: about one paper in twenty named a learning method — 4.7% in 2000, 3.9% in 2014. Then it climbed every year: 11.6% in 2018, 19.4% in 2022, 23.6% in 2025. One robotics paper in four now describes a method that learns. The newest wave is steeper still. Papers naming foundation or language models were 0.14% of the total in 2022 and 3.49% in 2025 — in three years, from almost nothing to one paper in 29.

Here is the surprise. The research area built around learning — learning for control — did not grow: 4.1% of all placed robotics papers in 2014, 4.2% in 2025. No AI field swallowed robotics. Instead the method moved into every field. In manipulation and grasping the share of papers using learning went from 4.7% to 46.8%; even in soft, micro and bio-inspired robotics, where it moved least, it rose. Robotics did not become a branch of AI. AI became a tool of every branch of robotics — the way the electric motor did not create an electric industry so much as change every factory.

Most of these papers come from universities: of the 17,403 learning papers of 2025 whose authors name an institution, 8.1% carry a company author. Among the companies, State Grid Corporation of China leads, then Shanghai Electric and China Southern Power Grid. What the words cannot tell us is whether the robots got better. A paper that names a neural network may still be about a gearbox. What they do show is a change in how robotics is done: the method arrived everywhere, quietly, before the models made the news.

The models in the news are a small, new share of robotics research: 5,807 papers in our collection have named a foundation or language model since 2022. Companies take a larger part in them — 12.8% carry a company author, against 8.0% of the learning papers of those years — yet universities and academies write most: Tsinghua University, Chinese Academy of Sciences, Shanghai Jiao Tong University and University of Hong Kong. Of the twenty robot models our earlier research names, eleven have their paper in our collection. Split by the authors’ countries, China holds 27.2% of the model papers and the United States 22.9%: the same race the humanoid story finds in patents.

A model that drives a robot has to learn from robots, and robot data is scarce: someone must steer a robot through a task, again and again. Papers about gathering that data went from 38 in 2018 to 328 in 2025. Consortia build the largest shared datasets: the paper of Open X-Embodiment names 34 institutions. Companies form around it: XDOF raised $70 million for teleoperation that records training data, Mind Robotics $500 million to train robots on a factory’s data. Our records see them only at the edge: we hold all four companies our notes and the press name; two have a funding round in our record, and none a paper or a patent.