Gemini Robotics 2 | What is available for AI robotics now

What actually changed in AI robotics in the last 24 hours? Google DeepMind made Gemini Robotics ER 2 available through Gemini API and AI Studio. It is a planning layer that can inspect continuous video, call tools and coordinate control models. It does not make a home robot autonomous from a prompt. Low-level control, physical limits and the emergency stop still belong in your build.

The part you can use now

DeepMind announced three parts under Gemini Robotics 2. ER 2 is the public one. It is designed to choose action sequences, use APIs and hand a task to a vision-language-action model or an existing controller.

Gemini Robotics 2 for whole-body control and Gemini Robotics On-Device 2 remain in early access for partners. That distinction matters. You can experiment with a high-level planner today, but the model for moving an entire robot is not openly available.

DeepMind’s results are useful, not independent proof

DeepMind reports 68.4% success clearing a table, 45.7% clearing the floor and 76.3% working on a shelf with Apollo 2. Its multi-finger-hand results range from 32% to 92%, depending on the task. Those are the company’s own evaluations, not an outside replication. The spread also points at a practical limit: fine manipulation does not become reliable simply by adding a larger model to a hand.

The safety report sets another boundary. It does not evaluate a certified functional-safety architecture. For a builder, the takeaway is straightforward. Keep the emergency stop, mechanical limits, speed limits and a controller that can halt an actuator even when the model makes a poor decision.

If you want to test the public layer, start with a narrow task. Let ER 2 choose among actions that are already safe and measurable, such as opening a gripper, moving an arm to approved positions or requesting a sensor reading. Measure before applying power. The model can organize the sequence, but it cannot replace the base that keeps a build from going wrong.