Gemini Robotics 2: When AI Learns to Move, Think, and Team Up

For decades, the dream of robots that seamlessly step into our world and lend a hand has felt perpetually five years away. Google DeepMind just moved the timeline. With the launch of Gemini Robotics 2, the company has delivered what it calls an "intelligence layer" for the next generation of adaptable robots — and the implications extend far beyond the lab.

The announcement, which shot to the top of Hacker News with over 560 points and 445 comments, introduces three models that together represent a significant leap in physical AI: a vision-language-action (VLA) model for whole-body humanoid control, an embodied reasoning (ER) model for multi-step task planning, and an on-device model that adapts to entirely new robot bodies in just hours.

From Feet to Fingertips: Whole-Body Control

The headline feature of Gemini Robotics 2 is full humanoid control. Previous iterations could manage upper-body table-top tasks. Now, the model can drive an entire robot — walking, crouching, stretching, and manipulating objects in sequence.

DeepMind demonstrated this with Apptronik's Apollo 2 humanoid robot. When given the instruction to "put the watering can into the green bin in the bottom shelf," Apollo walks to the table, picks up the can, navigates to the shelves, and places it precisely. It sounds simple, but this kind of whole-body coordination — locomotion plus manipulation plus spatial reasoning — has been one of the hardest problems in robotics.

The world is built for human movement. Shelves are at human heights. Doorways are human-width. Robots that can only reach but not walk, or walk but not grasp, will always be limited. Gemini Robotics 2 closes that gap.

Dexterity That Actually Matters

Beyond locomotion, the model brings a new level of fine dexterity. Using the five-fingered, 22-degree-of-freedom SharpaWave hand on Apollo 2, the robot can tie knots and seal ziplock bags. It can also operate standard two-fingered grippers on a Franka Duo platform for tasks like tight packing.

This matters because the gap between "impressive demo" and "actually useful" has always been dexterity. A robot that can pick up a box is neat. A robot that can seal a bag and pack it tightly next to other items is genuinely useful. We are watching that gap close in real time.

The Brain Behind the Body: Embodied Reasoning

Physical capability without intelligence is just a pre-programmed machine. What sets Gemini Robotics 2 apart is its embodied reasoning model, Gemini Robotics ER 2, which acts as the robot's high-level brain. It processes user instructions, observes the room, reasons about the steps needed, coordinates with the VLA model to execute actions, and tracks progress until the task is complete.

DeepMind claims robots can now execute multi-step tasks lasting several minutes and involving hundreds of decisions. The ER model understands when tasks begin and end, can self-correct if a step fails, and generalizes to novel situations. This is the difference between a robot that follows a script and one that actually thinks through a problem.

Robots That Work as Teams

Perhaps the most surprising feature is multi-robot collaboration. Different types of robots can now communicate and work together to solve workflows that a single robot could not handle alone. One robot picks items, another packs them. One navigates, another manipulates.

This is a frontier that most robotics companies have not even begun to approach. The coordination challenges — communication protocols, shared state, task allocation — are significant. If DeepMind's implementation works reliably, it could transform how warehouses, factories, and eventually homes are automated.

On-Device: Fast Adaptation Without the Cloud

The On-Device 2 model is perhaps the most practically important of the three. Running locally on robotic hardware without internet connectivity, it can adapt to completely new robot embodiments with just a few hours of data — typically fewer than 200 examples.

This means a manufacturer could take a brand-new robot design, collect a few hours of demonstration data, and have a capable AI controller ready. No cloud dependency. No massive retraining. The implications for deployment in remote environments, factories, and homes are enormous.

• Adapts to new robot bodies with fewer than 200 examples
• Works with drastically different shapes, sensors, and degrees of freedom
• No network latency or internet dependency required
• Tested on Dexmate, SO101, and Trossen platforms

Safety: The Make-or-Break Question

With great physical capability comes great responsibility. DeepMind is introducing ASIMOV-Agentic, a new benchmark for agentic safety orchestration. It measures whether the reasoning model can refuse unsafe tool calls from the VLA, predict whether a task is possible, and proactively request human intervention when uncertain.

Gemini Robotics ER 2 is described as DeepMind's safest robotics model to date in safety constraint following and human proximity benchmarks. It can detect when humans are nearby, trigger safety tool calls, and bring the robot to a safe stop. These are not nice-to-haves — they are prerequisites for any robot that will share space with people.

What This Means for the Rest of Us

Gemini Robotics 2 is not shipping to consumers. The VLA and On-Device models are available to early-access partners, and the ER model is in preview on Google AI Studio. But the trajectory is clear.

We are watching the transition from single-task automation to general-purpose physical AI. The combination of whole-body control, human-like dexterity, multi-step reasoning, team collaboration, and fast on-device adaptation is a convergence that has never existed before in a single platform. DeepMind explicitly frames this as a milestone on the path to solving AGI in the physical world.

The competition is intensifying too. Figure, Tesla Optimus, Boston Dynamics, and Apptronik are all racing toward commercial humanoid robots. But the differentiator will not be the hardware — it will be the intelligence layer. With Gemini Robotics 2, DeepMind just made a very strong claim to owning that layer.

For developers and researchers, the ER model being available on Google AI Studio means you can experiment with embodied reasoning today. For everyone else, the countdown to robots in warehouses, factories, and eventually homes just got a lot shorter.