Gemini Robotics 2 Review — Google DeepMind's Whole-Body Intelligence for Robots

Overview

On July 30, 2026, Google DeepMind unveiled Gemini Robotics 2 — a significant leap forward in physical AI. Building on the foundation of Gemini Robotics 1.5, this new generation introduces whole-body intelligence for humanoid robots, advanced dexterity for both robotic hands and grippers, and multi-robot collaboration.

The announcement comes at a time when the robotics industry is at an inflection point — multimodal AI models are finally becoming capable enough to serve as the “brain” for physical systems. Gemini Robotics 2 directly addresses the three hardest problems in robotics: generalization across embodiments, fine motor control, and long-horizon task planning.

Three Models, Three Capability Tiers

Gemini Robotics 2 ships as three distinct models, each optimized for different parts of the robotics stack:

1. Gemini Robotics 2 (VLA)

This is the vision-language-action model that converts visual input and language instructions directly into motor commands. It’s the “doer” — the model that actually controls the robot’s body.

Key capabilities:

• Full humanoid control — feet, legs, torso, arms, hands, fingers
• 22-degree-of-freedom hand control (SharkWave hand on Apollo 2)
• Fine dexterity: tying knots, sealing ziplock bags
• Standard gripper operation: Franka Duo bi-arm platform for tight packing

How it works: The model takes camera images + text instructions → outputs joint positions and torques directly. No intermediate motion planning layer needed. This end-to-end approach lets the model learn movement strategies that humans would never design manually.

2. Gemini Robotics ER 2 (Embodied Reasoning)

This is the vision-language model that serves as the robot’s high-level reasoning brain. It plans multi-step tasks, communicates with humans, monitors progress, and self-corrects when things go wrong.

Key capabilities:

• Multi-step task planning spanning several minutes
• Self-correction on failure — if a step fails, it replans
• Task completion awareness — knows when tasks begin and end
• Multi-robot coordination — different robots communicate and collaborate

Available now: ER 2 is available on Google AI Studio and in private preview on Gemini Enterprise Agent Platform. You can test the reasoning model without owning a robot.

3. Gemini Robotics On-Device 2

This is the efficient VLA optimized to run locally on robotic devices without network connectivity. It’s the model for real-time, low-latency applications where cloud dependency isn’t acceptable.

Key capabilities:

• Multi-embodiment native — works across different robot body types
• Fast adaptation: new robot embodiments in a few hours with <200 examples
• Works with drastically different shapes, sensors, and degrees of freedom
• Local inference — no cloud dependency for core motor control

Whole-Body Control: The First Real Humanoid Brains

The standout feature of Gemini Robotics 2 is whole-body control. Previous robotic AI models could only control upper-body movements — pick-and-place tasks on tables. Gemini Robotics 2 is the first VLA model that can control a full humanoid from feet to fingertips.

In DeepMind’s demonstrations, the Apptronik Apollo 2 humanoid performs tasks requiring coordinated whole-body movement:

1. “Put the watering can into the green bin on the bottom shelf” — Apollo 2 walks to the table, picks up the watering can, walks to the shelves, crouches, and places it precisely.

2. “Clean up this cluttered room” — Apollo 2 navigates obstacles, picks up objects at various heights, and places them in designated containers.

3. Tying knots with a five-fingered hand — The 22-DOF SharkWave hand performs delicate manipulation that was previously impossible for AI-controlled robots.

This is not just incremental improvement. Whole-body control requires the model to simultaneously reason about balance, reach, trajectory, force, and object interaction — a coordination challenge that neural networks have struggled with until now.

Dexterity: Hands That Work

Gemini Robotics 2 works with two types of end effectors:

Anthropomorphic Hands

The SharkWave five-fingered hand (22 degrees of freedom) enables tasks like:

• Tying knots in rope
• Sealing ziplock bags
• Grasping irregular objects (tools, fruit, electronics)

Standard Grippers

The Franka Duo bi-arm platform uses standard two-fingered parallel grippers for:

• Tight packing of objects into containers
• Precise placement in confined spaces
• Coordinated bi-manual manipulation

The key insight: one model handles both types. The VLA learns a unified representation of manipulation that generalizes across hand types, rather than requiring separate models for each.

Multi-Robot Collaboration

Gemini Robotics ER 2 introduces multi-robot collaboration — different types of robots communicating and working together on tasks that no single robot could complete alone.

In DeepMind’s demo, a wheeled robot and a humanoid work together to clean a room. The wheeled robot navigates quickly to fetch objects, while the humanoid handles the fine manipulation of placing them on high shelves. They coordinate through the ER 2 reasoning model, which assigns subtasks and monitors progress.

This is significant because real-world environments are heterogeneous — a factory might have floor bots, arm robots, and mobile manipulators. Making them collaborate under a single reasoning layer is the path to practical deployment.

Real-World Benchmarks

While DeepMind hasn’t published standardized benchmark scores yet, here’s what we can observe from their demonstrations:

Capability	Previous Best (Gemini Robotics 1.5)	Gemini Robotics 2
Body control	Upper body only	Full body (feet to fingertips)
Task duration	<1 minute	Several minutes
Self-correction	Limited	Full re-planning on failure
Multi-robot	Not supported	Supported (ER 2)
Embodiment transfer	~1 week	<1 day, <200 examples
Hand dexterity	Basic grasp	Knots, ziplock, precision
On-device inference	No	Yes (On-Device 2)

How to Access

Model	Access	Platform
Gemini Robotics ER 2	Public preview	Google AI Studio, Gemini Enterprise Agent Platform
Gemini Robotics 2 (VLA)	Early-access partners	Contact via application form
Gemini Robotics On-Device 2	Early-access partners	Contact via application form

The ER 2 model is the most accessible — you can try it in Google AI Studio right now. For the VLA and On-Device models, you’ll need to apply for the early-access partnership.

Pricing

• Gemini Robotics ER 2: Billed through Google AI Studio consumption (standard Gemini API pricing per token)
• VLA and On-Device models: Partnership-negotiated pricing
• Hardware costs: Apptronik Apollo 2 humanoid ($50K+), Franka Duo ($30K+), compute hardware for on-device inference (NVIDIA Jetson or equivalent, ~$5-15K)

Competitive Landscape

Feature	Gemini Robotics 2	RT-2 (Google)	Octo (Physical Intelligence)	π0 (Physical Intelligence)
Whole-body control	✅ Yes	❌ No	❌ No	❌ No
Fine dexterity	✅ Hands + grippers	❌ Basic	❌ Basic	❌ Basic
Multi-robot collaboration	✅ Yes	❌ No	❌ No	❌ No
On-device inference	✅ Yes	❌ No	❌ No	❌ No
Embodiment transfer	Hours	❌ Weeks	Limited	Limited
Reasoning integration	✅ ER 2	❌ Scope	❌ Scope	❌ Scope

The closest competitor is Physical Intelligence’s π0 model, which also uses a VLA approach. However, Gemini Robotics 2’s whole-body control and multi-robot collaboration are unique differentiators that no other platform currently offers.

Verdict

8.8/10 — Groundbreaking, but early-stage affordable access is limited.

Gemini Robotics 2 represents a genuine breakthrough in how AI models interact with the physical world. Whole-body control, fine dexterity, and multi-robot collaboration are capabilities that the robotics community has been pursuing for years. Google DeepMind has delivered them in a unified model family.

The main limitations are availability (VLA and On-Device are restricted to partners) and hardware cost (humanoid robots remain expensive). For research labs and deep-pocketed enterprises, this is transformative. For the broader developer community, the ER 2 reasoning model in AI Studio offers a taste of what’s coming.

As hardware costs decline and the models become more broadly available, Gemini Robotics 2 could be remembered as the moment when “general purpose robots” went from sci-fi to shipping reality.