General robot AI: first releases in 2024, controlling whole humanoid bodies by 2026, makers say
AI that turns camera images and instructions into robot movement has improved every few months. Nearly all results come from the makers.
A vision-language-action (VLA) model is the robot version of a chatbot. It takes in camera images and a written instruction such as “fold the towel” and outputs motor commands. The aim is one general model that can be adapted to many robots and tasks, instead of programming each job by hand.
The first wave arrived between late 2024 and spring 2025. Physical Intelligence released π0 in October 2024 and made it open source in February 2025. Google DeepMind’s Gemini Robotics and NVIDIA’s openly available GR00T N1 for humanoids followed in March 2025. In November 2025 Physical Intelligence’s π*0.6 learned from its own practice runs, which the company says more than doubled throughput on some of its hardest tasks. One robot running it made espresso drinks from 5:30 am to 11:30 pm.
In 2026 the models moved to whole humanoid bodies and to outside customers. Figure says its Helix 02 (January 2026) gives its robot full-body autonomy. Physical Intelligence’s π0.7 (April 2026) folded laundry on a robot arm it had no laundry-folding data for. DeepMind’s Gemini Robotics 2 (30 July 2026) adds whole-body control and drove Apptronik’s Apollo 2 humanoid. In October 2026 Physical Intelligence opened a test API, so other companies can adapt its models with their own data and run them on their own robots.
What it means
Robot makers are moving from programming each task by hand to adapting one general model, and some now offer theirs to other companies. But with no shared test, nobody outside these companies can check how well the models work or compare them. A long demo shows what a robot can do on a good day, not how often it succeeds.
Caveats
- Almost every number here comes from the company that built the model, measured on tasks it chose. There is no shared benchmark that all these models are tested on.
- Long-running demos are selected examples, not full statistics. Physical Intelligence labels its long-run videos “evaluation highlights”.
- NVIDIA says the GR00T N2 model it previewed in March 2026 succeeds at new tasks in unseen places more than twice as often as leading models, but has not said which models or how it tested.
- Access is limited. Gemini Robotics 2 is available only to selected early-access partners.