Google details how Gemini Robotics teaches robots to play ball and fold origami

Put a random shoe on a table and ask a robotic arm to drop pens inside it. Then hand it a toy basketball hoop it has never seen before and say, "Slam dunk the ball."

Without prior training on basketball toys, the robot paused, figured out the spatial puzzle, and nailed the dunk on its first try. That was the breakthrough moment Google DeepMind’s robotics team witnessed before unveiling its new Gemini 2.0 robotics family.

Announced with backing from Google CEO Sundar Pichai, the new lineup builds directly on Gemini 2.0, extending its multimodal outputs beyond text, video, and audio into physical movement.

From folding origami to humanoid bodies

The release splits the workload between two primary models, tackling how robots think and how they move.

First is Gemini Robotics-ER, an embodied reasoning model built on Gemini 2.0 Flash. It specializes in understanding the physical scene in front of it: identifying 3D objects, locating parts, and predicting the trajectory and grip required before writing executable code. Google is rolling this model out to trusted testers and partners.

Then there is Gemini Robotics, DeepMind’s flagship vision-language-action model focused on dexterity. Hand manipulation has long been a brick wall in robotics, requiring split-second spatial judgment alongside precise motor control.

DeepMind ditched the traditional approach of training machines repeatedly on single tasks. Instead, it exposed the model to a massive variety of activities, gambling that generalization would emerge naturally.

It worked. In internal tests, machines powered by Gemini Robotics prepped salads, packed kids’ lunchboxes, played Tic-Tac-Toe, and even folded an origami fox.

Crucially, these models are not tied to a single machine. DeepMind confirmed they run on dual-arm ALOHA research setups as well as Apollo, the full-size humanoid robot developed by partner Apptronik.

Why it matters

We think this release signals a genuine turning point in how physical machines learn.

For decades, robotics has suffered from narrow specialization: teach an arm to weld a specific door frame, and it breaks down if the frame shifts two inches. By feeding Gemini 2.0 massive task diversity alongside robot-specific fine-tuning, Google is giving hardware a baseline of physical common sense.

For developers and robotics startups, getting access to models like Gemini Robotics-ER means skipping years of building basic scene-understanding pipelines from scratch.

Still, Parada herself cautions that robots handling real-world household chores remain years away. A controlled lab test folding an origami fox looks brilliant on video, but navigating unpredictable living rooms with dynamic obstacles and fragile objects is a completely different beast. Hardware costs, joint wear, and latency remain massive hurdles. Expect these generalist models to land on high-mix assembly lines and industrial settings long before one does your laundry.