Google DeepMind's first robotics model taught robots to manipulate objects on a table. Gemini Robotics 2 teaches them to use their whole body to get there.
Google DeepMind has unveiled Gemini Robotics 2, and it's a notable jump for humanoid robots. The first version taught robots how to manipulate objects on a tabletop. This release extends that to the whole body, so a robot can move through a space and act on it as one continuous behavior. It's built on the vision-language-action approach that's reshaping robotics, now applied to entire humanoids rather than just arms.
Whole-body control, from feet to fingertips
Think about picking up a watering can from a low shelf. You don't just move your arm. You walk over, bend down, reach out, and grasp it, all as one fluid motion. Earlier Gemini Robotics models controlled only a robot's upper body for tabletop work. Gemini Robotics 2 plans that entire sequence, controlling the robot from its feet to its fingertips.
In DeepMind's demo, Apptronik's Apollo 2 humanoid is told to "put the watering can into the green bin on the bottom shelf." The robot walks to the object, picks it up, steps over to the shelves, and places it, coordinating balance, locomotion, and manipulation together. That whole-body coordination is exactly what makes humanoids so hard, and getting it into a single model is the headline advance here.
Five-finger dexterity and where it still falls short
Movement is only part of the story. Once a robot reaches an object, it still has to interact with it, which is the harder half. Gemini Robotics 2 can drive a five-fingered, 22-degree-of-freedom hand (Apollo 2's SharpaWave hand) to do delicate tasks like tying a trash bag, sealing a ziplock bag, and handling a light bulb.
It's worth being honest about the numbers, because they show how hard dexterity remains, which is a theme in why robot hands are so hard. In DeepMind's own reporting, the model unscrewed a bulb successfully about 92% of the time but managed only 36% when screwing one back in, with tying a trash bag around 44% and sealing a bag near 40%. It does noticeably better with simpler two-finger grippers on a Franka Duo, scoring roughly 74% on general pick-and-place and about 90% on precise insertion. Five-finger manipulation is progressing, but it's clearly the frontier.
Longer tasks and multiple robots with Gemini Robotics ER 2
Google also introduced Gemini Robotics ER 2, an embodied reasoning model that acts as the robot's high-level brain. Rather than thinking one step ahead, it plans longer tasks that can run for several minutes and hundreds of decisions, tracks progress through them, and knows when a task starts and ends.
It also enables multi-robot collaboration, letting different types of robots communicate and coordinate on a workflow that a single machine couldn't finish alone, like a multi-step cleanup or organizing a garage. In practice, ER 2 plans and supervises while the vision-language-action model executes the motion, and it can self-correct partway through when a step fails.
On-device models that run without the cloud
DeepMind also released Gemini Robotics On-Device 2, a version optimized to run locally on the robot itself. That matters because many real-world settings can't rely on a constant cloud connection, and network latency can break a delicate, multi-step task.
Running on-device makes the model easier to deploy across different robot platforms, and DeepMind says it can adapt to a completely new robot body with only a few hours of training data. The models have been shown across several embodiments, from the Apollo 2 humanoid to a Franka arm, using the same underlying approach.
What Gemini Robotics 2 gets right and what it doesn't
Gemini Robotics 2 is more than a single model release. It brings movement, manipulation, and planning together in one system: the VLA handles motor control, ER 2 handles reasoning and coordination, and the on-device model handles deployment. DeepMind also paired it with a new safety benchmark, ASIMOV-Agentic, aimed at getting robots to refuse unsafe actions and stop safely when a person is nearby.
The honest picture is a real step forward with clear limits, whole-body tasks and gripper work are solid, five-finger dexterity and movement speed still have a long way to go. But combining all three capabilities in one system moves robots closer to handling real-world tasks from start to finish, which is the whole promise of embodied AI.
FAQ
What is Gemini Robotics 2?
Gemini Robotics 2 is a suite of three AI models from Google DeepMind for controlling robots. It includes a core vision-language-action model for whole-body motor control, Gemini Robotics ER 2 for reasoning and multi-step planning, and Gemini Robotics On-Device 2 for running locally on robot hardware.
What is new in Gemini Robotics 2 compared to the first version?
The original Gemini Robotics controlled a robot's upper body for tabletop manipulation. Gemini Robotics 2 adds whole-body control (walking, crouching, and reaching as one motion), stronger five-finger dexterity, longer multi-step task planning, multi-robot coordination, and on-device operation.
What is Gemini Robotics ER 2?
Gemini Robotics ER 2 is an embodied reasoning model that acts as a robot's high-level brain. It plans multi-minute tasks involving hundreds of decisions, tracks task progress, communicates with people, and coordinates multiple robots working together. It's available in preview on Google AI Studio.
How good is Gemini Robotics 2 at dexterous manipulation?
It's capable but uneven. On five-finger hands it does well at some tasks (around 92% unscrewing a bulb) and poorly at others (about 36% screwing one in). It performs more reliably with two-finger grippers, scoring roughly 74% on pick-and-place. Fine multi-finger dexterity remains the hardest part.
Can you experiment with robotics AI like this in simulation?
Yes. Models like these are developed and validated heavily in simulation before running on real robots. Tools like Drift generate the simulated robots and environments used for that testing, so you can prototype navigation and manipulation behaviours virtually before deploying to hardware.
Top comments (0)