TECHNOLOGY · JUL 30, 2026
Gemini Robotics 2 runs a whole humanoid body and 22 joint hands
Google DeepMind's new models walk, crouch and handle small objects on Apptronik's Apollo 2. Its own numbers show many fine hand tasks still fail more often than not.
By Riley Hart · 3 sources linked below
This is robot software for companies, not something you can buy for your home. It shows humanoids getting better at the small, fiddly jobs a household needs, like bagging trash. DeepMind's own results show hand tasks failing more than half the time, so home ready humanoids built on it are still a way off.
Google DeepMind announced Gemini Robotics 2 on July 30, 2026. It is a family of three AI models for robots: one that turns sight and spoken instructions into movement, one that plans multi step tasks, and a smaller version that runs on the robot itself without the cloud.
The main new skill is whole body control. Earlier robot models mostly moved arms on a fixed base. DeepMind says the new model controls a full humanoid as it walks, crouches and handles objects, for example carrying a watering can to a bin on a shelf.
DeepMind showed it on Apptronik's Apollo 2 humanoid fitted with SharpaWave hands, which have five fingers and 22 joints, or degrees of freedom. Tasks included tying a trash bag, sealing a ziplock bag, unscrewing a light bulb and using a dustpan. A Franka two arm robot with simple grippers was also tested, and a humanoid and a two arm robot were shown working together on one task.
DeepMind also says the on device model can adapt to a new two arm robot in a few hours, usually with fewer than 200 examples. If that holds up outside the lab, robot makers would not need to collect huge amounts of new data for each machine.
DeepMind's own charts show how hard hand work still is. On Apollo with the Sharpa hands, it reports 92% success unscrewing a bulb but 36% screwing one in, 44% tying a trash bag, 40% sealing a ziplock bag and 32% using a dustpan. For whole body pick up tasks, it reports 76% at a shelf, 68% at a table and 46% from the floor. These are company figures, and no independent tests have been published. DeepMind says the robots still need to get faster and that multi finger handling remains a challenge. The Robot Report notes that its coverage relies on DeepMind's statements.
On safety, DeepMind added a test called ASIMOV-Agentic. It checks whether a robot refuses unsafe actions, judges whether a task is possible and asks a person for help when unsure. DeepMind calls the planning model its safest robotics model so far on its own benchmarks.
Access is limited. The planning model, Gemini Robotics ER 2, is available to developers in Google AI Studio. The movement and on device models are only open to early access partners, which DeepMind names as Apptronik, Boston Dynamics and Agile Robots.
Apptronik's Apollo is not on our humanoid tracker, which lists robots meant for homes. The models could reach home robots later through partners, but DeepMind has announced no consumer product or date.