Gemini Robotics 2

Learning Objectives

• Explain what Gemini Robotics 2 is and how its three models divide the work
• Describe whole-body control and why it is harder than arm-only manipulation
• Evaluate which parts of the family you can actually build on today

What Is Gemini Robotics 2?

Gemini Robotics 2 is Google DeepMind's embodied-AI model family, announced on July 30, 2026. Rather than one model that does everything, it splits the problem three ways:

• Gemini Robotics ER 2 — the embodied reasoning model. It perceives a scene, plans a sequence of steps, orchestrates them, and self-corrects when something goes wrong. This is the layer you talk to.
• A vision-language-action model — the motor-control layer that turns an instruction into joint movements.
• An on-device variant — a smaller model for robots that cannot rely on a network round-trip.

The split matters because it lets ER 2 treat the low-level controller as a tool. Developers can declare a vision-language-action model as something ER 2 calls, the same way a text agent calls a function — so you can swap the motor layer without rewriting the planner.

Whole-Body Control

The capability Google leads with is intelligent whole-body control: a humanoid walking, crouching, reaching and manipulating an object as one coordinated motion rather than as separate phases.

This is harder than it sounds. Most robot manipulation research treats the base as fixed — the robot drives somewhere, stops, then the arm does the work. Whole-body control means the balance problem and the manipulation problem are solved together, because reaching for a low shelf changes the robot's center of mass while it is still holding itself up. Getting that wrong does not produce a slightly worse grasp; it produces a fall.

The family also advances dexterous hand control and multi-robot collaboration, where different machines share a semantic understanding of a task well enough to hand work between them. Google demonstrated two dissimilar robots — Apptronik's Apollo 2 humanoid and a Franka F3 Duo arm — cooperating on a single task.

Benchmark Results

Google published benchmarks across quite different task families, and the spread is wide enough to be worth reading carefully.

Read the ranges rather than the top numbers. A multi-finger success rate that runs from 32 percent to 92 percent depending on the task is not a solved capability — it is a capability that works well on some manipulations and poorly on others, and finding out which is your integration work.

Safety Behavior

Google reports meaningful gains over the prior ER 1.6 generation on two safety benchmarks, Safety Instruction Following and Human Proximity. The demonstration behavior is concrete: the model halts a humanoid when a person comes near and resumes autonomously only once the area is clear.

That is the right shape for a safety property — enforced as a behavior the model exhibits, and measured on a benchmark, rather than asserted in a prompt.

Pricing

Google has not published pricing specific to the robotics models; ER 2 is reached through the ordinary Gemini API surface.

Getting Started

1. Open Google AI Studio and select Gemini Robotics ER 2 — you can exercise the reasoning and planning behavior against images and video without any robot hardware.
2. Move to the Gemini API once you have a task worth automating, and wire ER 2's plan output into whatever controller you already run.
3. Declare your existing low-level controller as a tool ER 2 can call, rather than waiting on Google's vision-language-action model.
4. Apply for early access only if the motor-control layer is genuinely the piece you are missing.

Strengths

• The reasoning layer is actually available — ER 2 ships through the Gemini API and AI Studio, not a waitlist
• Controller-agnostic by design — declaring a vision-language-action model as a tool means you are not locked to Google's motor layer
• Whole-body control — coordinated locomotion and manipulation rather than drive-then-reach
• Multi-robot handoff — demonstrated across dissimilar hardware, not just two of the same robot
• Measured safety gains — human-proximity and safety-instruction benchmarks improved over ER 1.6

Limitations

• Two of three models are early-access only — the vision-language-action and on-device models require an application and carry no public availability date
• Uneven benchmark results — multi-finger task accuracy spans 32 percent to 92 percent depending on the manipulation
• No published robotics pricing — billing runs through the general Gemini API, which makes per-task cost hard to model in advance
• Hardware partnerships, not hardware — Google supplies models; you still source and integrate the robot
• Enterprise platform is private preview — the managed deployment path is not open

Key Takeaways

• Gemini Robotics 2 splits embodied AI into a reasoning model, a motor-control model, and an on-device variant, letting the planner call the controller as a tool
• Gemini Robotics ER 2 is the generally available piece, through the Gemini API and Google AI Studio; the other two models are early-access only
• Whole-body control coordinates balance and manipulation together, which is what makes a humanoid crouching to reach a low shelf a different problem from an arm on a fixed base
• Benchmark ranges matter more than headline numbers here — dexterous manipulation still varies sharply by task