Gemini Robotics 2: The AI That Gives Robots a Brain

A humanoid robot walks to a table. It picks up a watering can, crosses the room, and sets the can down on a shelf. No task-specific script. No hand-coded motion path. Just a single model reading the room and deciding what to do next.

This is Gemini Robotics 2, Google DeepMind's latest robotics model family, announced on July 30, 2026. It marks the point where robots stop repeating fixed routines and start reasoning through open-ended physical tasks.

What robots could not do before Gemini Robotics 2

Most robots today run on narrow, pre-programmed sequences. A warehouse arm picks the same box shape all day. A demo humanoid follows a scripted path. Change the object, the lighting, or the room layout, and the system breaks.

Skill transfer between robot bodies has been just as hard. A model trained on one arm rarely works on another arm, let alone a full humanoid with legs and a torso. Every new embodiment meant starting the training process from scratch.

Gemini Robotics 2 targets both problems directly. It generalizes to tasks it never saw during training, and one model checkpoint can drive multiple robot bodies, from tabletop arms to full humanoids with legs and hands.

How Gemini Robotics 2 actually works

Gemini Robotics 2 ships as three separate models that work as one system, not one monolithic model trying to do everything.

  Gemini Robotics 2 stack

Gemini Robotics 2 is the vision-language-action model, or VLA. It converts what the robot sees and what it's told into motor commands. This is the layer that actually moves the joints, fingers, and feet.

Gemini Robotics ER 2 is the embodied reasoning model. Think of it as the robot's high-level brain. It watches the scene, talks with a human in natural language, and plans a sequence of steps that can run for several minutes and involve hundreds of decisions. It hands off the physical execution to the VLA below it.

Gemini Robotics On-Device 2 is a lightweight VLA built to run locally on the robot itself, with no cloud connection required. It inherits motion transfer techniques from Gemini Robotics 1.5, letting it adapt to a brand new bi-arm robot body in a few hours, typically using fewer than 200 examples.

The reasoning layer treats the action layer as a callable tool. ER 2 can also call other tools natively, like Google Search or a user-defined function, and it streams video, audio, and text through the Gemini Live API's bidirectional streaming endpoint. That streaming design is what removes the old stop-and-think delay between perception and action.

Multi-robot collaboration builds on this same structure. Two different robots, say a wheeled rover and a humanoid, share a semantic understanding of the task and divide the work between them. Apptronik's Apollo 2 and the Franka FR3 Duo demonstrate this handoff in DeepMind's published demo.

What Gemini Robotics 2 can physically do

The model controls entire humanoid bodies, not just an upper torso and arms. When told to put a watering can into a bin on a bottom shelf, Apollo 2 walks to the table, picks up the can, crosses the room, and places it precisely, coordinating legs, balance, and grip in one continuous motion.

Dexterity gets a real upgrade too. The model drives the five-fingered, 22 degree-of-freedom SharpaWave hand on Apollo 2 for delicate work like tying knots and sealing a ziplock bag. It also runs standard two-fingered parallel grippers on a Franka Duo platform for tighter, more precise packing tasks.

Video Source

The numbers behind Gemini Robotics 2

DeepMind published specific success rates rather than vague claims, and the gaps between task types are worth reading closely.

  General whole body manipulation

For whole-body manipulation on Apollo 2 with Inspire hands, the model hit 68.4% on picking objects up from a table, 45.7% from the floor, and 76.3% from a shelf. Floor pickups clearly remain the harder balance-and-reach problem.

Multi-finger dexterity numbers, tested on Apollo 2 with SharpaWave hands, show a wide spread: 92% on unscrewing a light bulb, but only 36% on screwing one in, 44% on tying a trash bag, 32% on using a dustpan, and 40% on sealing a ziplock bag. Fine five-finger coordination is still the frontier, not a solved problem.

Gripper-based dexterity on the Franka Duo scored higher and more consistently: 74.2% on general pick and place, 78.9% on diverse tool kitting, and 89.6% on precise insertion tasks.

On the reasoning side, Gemini Robotics ER 2 reached 57.4% accuracy on progress classification, the task of judging how far along a robot is in completing an action. On moment-finding, the task of pinpointing the exact video frame where a key event happens, it hit 91.3% accuracy with a 0.96 second mean absolute distance, running at 4x the execution speed of comparable models.

Calling Gemini Robotics ER 2 through the API

Gemini Robotics ER 2 is publicly available today through the Gemini API and Google AI Studio. Here's a minimal Python example using the official google-genai SDK.

from google import genai

client = genai.Client(api_key="YOUR_API_KEY")

response = client.models.generate_content(
    model="gemini-robotics-er-2-preview",
    contents=[
        "Look at this workspace image. Identify the light bulb "
        "and describe the steps needed to unscrew it safely.",
    ],
)

print(response.text)

For real-time video and low-latency tool calls, DeepMind's own examples use the Live API's bidirectional streaming setup. Google's robotics-samples repository on GitHub includes a working notebook for this exact configuration, along with a Boston Dynamics Spot demo that fetches objects on command.

How the three models compare

Where this leaves robotics

Gemini Robotics 2 does not claim robots have solved dexterity or balance. The published numbers show exactly where the system holds up, gripper tasks and shelf pickups, and where it still struggles, five-finger manipulation and floor-level reach. That kind of transparency is rare in a field full of polished demo reels.

What matters more is the shift in structure. One reasoning model, one action model, and one on-device model, working as separable layers that can drive different robot bodies without retraining from zero. That separation is what lets a Boston Dynamics Spot and an Apptronik humanoid run on the same underlying intelligence.

Robotics has spent decades chasing general-purpose machines that adapt the way humans do. Gemini Robotics 2 does not close that gap completely, but it narrows it in a way developers can already test through the Gemini API. For anyone building physical AI, that access matters more than any single benchmark score.

FAQs