Google DeepMind launches Gemini Robotics, turning Gemini 2.0 into a robot controller and signing Apptronik as its humanoid partner

Google DeepMind on 12 March 2025 introduced two robot models built on its Gemini 2.0 family, the most direct attempt so far by a frontier AI lab to turn a general multimodal model into the controller of a physical machine. Gemini Robotics is a vision-language-action (VLA) model that adds physical actions as an output, so the same network that reads a camera feed and a spoken instruction also drives a robot's arms. Gemini Robotics-ER, short for embodied reasoning, is a vision-language model tuned for spatial understanding that robot makers can connect to their own controllers.

The company gave two headline numbers. In the announcement, written by Carolina Parada, Google DeepMind said Gemini Robotics more than doubles performance on a comprehensive generalization benchmark compared with other state-of-the-art VLA models. It said Gemini Robotics-ER, when used end to end to perceive, plan and generate code for a robot, reaches a success rate two to three times that of Gemini 2.0. The technical report released with the models says the action system runs at an effective 50 Hz, with a cloud-hosted backbone answering in under 160 milliseconds and end-to-end latency from raw observation to a chunk of low-level actions of about 250 milliseconds.

Most of the training robot data came from Google DeepMind's own fleet of ALOHA 2 bi-arm robots. According to the report, teleoperators collected thousands of hours of expert demonstrations across thousands of tasks over 12 months, and the lab mixed that with web documents, code, images, audio and video. In demonstrations the model folded an origami fox and packed a snack into a Ziploc bag. The report says the model learns new short-horizon tasks from as few as 100 demonstrations; after specialisation on 2,000 to 5,000 episodes per task it packed a lunch box with a 100% success rate over 20 trials. Google DeepMind also showed it adapted to a bi-arm Franka platform and to Apptronik's Apollo humanoid.

For the reasoning model the lab built a new open benchmark, ERQA, with 400 multiple-choice visual questions about the physical world, published on GitHub under a CC BY 4.0 licence. In the report's baseline results Gemini 2.0 Flash scored 46.3%, Gemini 2.0 Pro Experimental 48.3%, OpenAI's GPT-4o 47.0% and Anthropic's Claude 3.5 Sonnet 35.5%; chain-of-thought prompting lifted Gemini 2.0 Pro Experimental to 54.8%. On pointing, the skill a robot needs to know where to grasp, Gemini Robotics-ER scored 71.3 on Paco-LVIS against 46.1 for Gemini 2.0 Flash and 16.2 for GPT-4o, and it reached 48.3 AP@15 on SUN-RGBD 3D detection.

The commercial wrapper is a partner programme rather than a product. Google DeepMind said it is working with Apptronik, the US humanoid maker, to build the next generation of humanoid robots with Gemini 2.0. Apptronik had raised a US$350 million Series A on 13 February 2025, led by B Capital and Capital Factory with Google participating, and described a strategic collaboration with Google DeepMind's robotics team. Gemini Robotics-ER went to a group of trusted testers: Germany's Agile Robots, Agility Robotics, Boston Dynamics and France's Enchanted Tools. The VLA itself was not made generally available, and no pricing was given.

Google DeepMind, Alphabet's London-based AI lab, has run robot learning research for years, and the data behind this release came from its own ALOHA 2 bi-arm robots. What changed in March 2025 is the starting point. Instead of training a robot policy and attaching language later, the lab began from Gemini 2.0, which already knows what a Ziploc bag is and how a lunch box closes, and taught it to act.

The race it joined is crowded. Figure AI published Helix on 20 February 2025, a two-part model in which a 7-billion-parameter vision-language model reasons at 7 to 9 Hz and an 80-million-parameter policy drives a 35-degree-of-freedom upper body at 200 Hz, trained on about 500 hours of teleoperation and running entirely on embedded GPUs inside the robot. Physical Intelligence distributes its π0 models through the openpi repository with base checkpoints pretrained on more than 10,000 hours of robot data; the Gemini Robotics report used a re-implementation of π0 as one of its baselines. Six days after the Google launch, on 18 March, NVIDIA released Isaac GR00T N1 at GTC as an open, customizable humanoid model and said it generated 780,000 synthetic trajectories in 11 hours, improving performance by 40% when combined with real data.

Plainly put, there are two ways to use the release. A robot maker can hand the whole job to Gemini Robotics, which takes pixels and words in and puts joint commands out. Or it can keep its own motion control and ask Gemini Robotics-ER the questions that classical software struggles with, such as where to put a gripper on a mug, which way a drawer opens or what path to follow, and receive points, boxes and trajectories back. The second path is the one Google opened to outside companies first, which says something about how cautious the lab is about handing over direct motor control.

For the global industry the significance is the business model on display. Google DeepMind is positioning itself as a supplier of robot intelligence to many hardware makers at once, including companies that compete with each other, much as Android supplied phone software. That puts it in direct contest with Figure, which builds its own model for its own robot, and with NVIDIA, which gives models away to sell chips and simulation tools.

The open questions are substantial. The headline generalization claim is relative and measured on Google's own benchmark against baselines the authors chose, including their own re-implementation of a rival's model. The split architecture depends on a cloud backbone and a network link, which factories and homes cannot always guarantee. Not every benchmark went Google's way. On Where2Place pointing, the open Molmo 72B model scored 63.8 against 45.0 for Gemini Robotics-ER. Google DeepMind addressed safety with a dataset it calls ASIMOV for evaluating semantic safety and with what it calls data-driven constitutions, but neither substitutes for certified low-level safety controllers on real machines.