Google DeepMind splits its robot brain into three models

Google DeepMind released the second generation of its physical-AI stack on 30 July 2026, and the notable part is the shape of it: instead of one robotics model there are three. Gemini Robotics ER 2 does the thinking, Gemini Robotics 2 turns plans into whole-body motion, and Gemini Robotics On-Device 2 runs on the robot itself so the machine keeps working with no network. The reasoning model is built on Gemini 3.5 Flash and, according to its model card, accepts interleaved text, images, video and audio inside a 128,000-token context, returning up to 64,000 tokens. The shift from the ER 1.x line is temporal rather than cosmetic: earlier versions judged a still frame or a short clip, while ER 2 reasons over a continuous video stream and keeps track of what it has already done across minutes of a task — the difference between recognising a scene and remembering a job. Access is split too. ER 2 is available through the Gemini API and Google AI Studio, with the Gemini Enterprise Agent Platform in private preview, while the two action models reach robot makers through partnerships; weights stay closed in all three cases. DeepMind's card is unusually blunt about limits, ruling out safety-critical deployment in healthcare, transport and anything where a malfunction could hurt someone, and it reports separate evaluations for safety instruction following and human-safety monitoring. The wujec.ai catalogue now carries the 2026 line as its own profile; the 2025 first generation stays where it was.