DeepMind puts Gemini 2.0 into robots with two new models

Big AI models are great at answering trivia or writing code, but the moment you ask them to fold an origami crane or stuff a snack into a Ziploc bag, physical reality gets messy.

Google DeepMind unveiled two new AI models built on Gemini 2.0 designed specifically to give machines physical common sense: Gemini Robotics and Gemini Robotics-ER. The goal is straightforward: bring multimodal intelligence out of chat interfaces and directly into the physical world.

Hands-on control meets spatial reasoning

The two releases tackle real-world automation from complementary angles.

Gemini Robotics is a vision-language-action (VLA) model. It treats physical robot movements as a native output format alongside text or code. Trained primarily on the dual-arm ALOHA 2 platform and Franka robotic arms, the model more than doubles the average performance of competing SOTA VLA models on broad generalization benchmarks. It takes conversational prompts in multiple languages, reacts to physical interruptions on the fly, and manages delicate dexterity tests like multi-step paper folding.

Meanwhile, Gemini Robotics-ER (Embodied Reasoning) functions as an advanced spatial brain. Instead of directly moving motors, it handles perception, 3D object detection, trajectory planning, and code generation. Roboticists can plug it straight into their existing low-level safety controllers. DeepMind reports that Robotics-ER reaches a 2x to 3x higher success rate in end-to-end tasks compared to stock Gemini 2.0. If you show it a mug, it intuitively calculates a safe approach trajectory and code for a two-finger pinch around the handle.

Hardware partnerships are already in motion. DeepMind is collaborating with Apptronik to embed the system into its Apollo humanoid robot, while trusted tester access is rolling out to Boston Dynamics, Agility Robots, Agile Robots, and Enchanted Tools. To address safety, DeepMind also released the ASIMOV evaluation dataset alongside a natural-language framework to automatically generate safety constitutions for physical agents.

Why it matters

For years, robotics was bottlenecked by brittle, hand-coded policies. Move an object three inches to the left, and the script would fail. By leveraging Gemini 2.0's general world knowledge, Gemini Robotics allows hardware to handle unseen objects and unexpected environments without extensive retraining.

We think DeepMind's two-tier architecture is especially clever. Fully end-to-end black-box control is still risky when multi-hundred-pound metal limbs are moving near people. By shipping Gemini Robotics-ER alongside raw action models, DeepMind allows hardware makers to lean on Gemini for spatial reasoning and code generation while keeping deterministic, low-level collision-avoidance loops intact.

That said, high-latency cloud reasoning remains a challenge for dynamic real-world tasks. Folding paper in a demo is impressive, but messy domestic chores and fast-moving factory floors demand millisecond-level responsiveness. For now, access is locked behind an enterprise preview, but Google just threw down a serious gauntlet for everyone trying to build embodied general-purpose robots.