# Google DeepMind launches Gemini Robotics 2 for cross-hardware robot control > Source: > Published: 2026-07-30 15:32:08+00:00 Google DeepMind's robotics group, led by Carolina Parada, launched Gemini Robotics 2 on July 30 as a three-model stack for planning, motor control and local inference across robot bodies ranging from dual-arm workstations to humanoids. For startups building robot hands, humanoids, foundation models or deployment tooling, the release provides a public reasoning model to test alongside a more capable action model that remains restricted. The stack comprises [Gemini Robotics 2](https://deepmind.google/models/gemini-robotics/vla/?ref=runtimewire), Gemini Robotics ER 2 and Gemini Robotics On-Device 2. The flagship control model can direct complete humanoid bodies from feet to fingertips, according to Google DeepMind, including balance adjustments while stepping, squatting or bending. It is in private preview, with developers able to apply for early access. In its [X announcement](https://x.com/GoogleDeepMind/status/2082844170998182350?ref=runtimewire), Google DeepMind said Gemini Robotics 2 can control a five-fingered hand to tie a knot or screw in a lightbulb while also managing parallel grippers. An attached image labels a dual-arm platform as Duo and a humanoid as Apollo, with colored point-cloud overlays marking objects and surfaces. ### What founders can test [Gemini Robotics ER 2](https://deepmind.google/models/gemini-robotics/embodied-reasoning/?ref=runtimewire), the stack's embodied-reasoning component, is available in public preview through Google AI Studio and the Gemini API. A separate private preview is available through the Gemini Enterprise Agent Platform. Google says ER 2 adds success tracking, tool use and coordination between robots with different physical capabilities. That access gives robotics teams a way to evaluate the planning layer immediately. The motor-control model has a narrower distribution: Google DeepMind says it is working with more than 100 trusted testers, including research partners Agile Robots, Apptronik and Boston Dynamics. Google has also published internal task results. A Franka Duo system running Gemini Robotics 2 scored 74.2% on general pick-and-place tasks, 78.9% on tool kitting and 89.6% on precision insertion. An Apollo humanoid equipped with Sharpa hands screwed in a lightbulb in 36% of trials, tied a trash bag in 44%, used a dustpan in 32% and handled a zip-lock bag in 40%. Those figures offer a more useful reference than demonstration footage, though they remain Google-run evaluations rather than independent comparisons. Teams assessing the system still need results for their own hardware, objects and operating conditions, including recovery after failed grasps and performance under changes in lighting or object placement. [Gemini Robotics On-Device 2](https://deepmind.google/models/gemini-robotics/on-device/?ref=runtimewire) addresses local inference. Google DeepMind says it can adapt to new hardware with fewer than 200 examples and several hours of training. Access remains limited to trusted testers. Google DeepMind has not published pricing, model sizes, training-data details or a broad release schedule for the flagship control model. Its private-preview status also prevents most startups from reproducing the reported manipulation results on their own systems. ### The competitive reference point Other robotics developers describe similarly broad ambitions. [Skild AI calls its model "omni-bodied"](https://www.skild.ai/blogs/building-the-general-purpose-robotic-brain?ref=runtimewire), covering humanoids, quadrupeds, arms and mobile manipulators. Figure describes [Helix 02](https://www.figure.ai/news/helix-02?ref=runtimewire) as a neural system for walking, balancing and manipulation directly from pixels. Gemini Robotics 2 gives founders published task results and immediate access to its reasoning component. Startups can test ER 2 through Google's developer tools now, while treating the flagship model's private-preview benchmarks as reference data until they can evaluate its action policies on their own hardware.