How is pricing calculated?

Pricing is based on the hourly rate of operators in your selected region. You only pay for the time operators spend controlling your robots.

What robots are supported?

We support 9+ robot arms including SO-100, Koch, Franka FR3, Franka Panda, WidowX 250, ViperX 300, ViperX 300s, and ALOHA.

What is the latency for teleoperation?

Our platform provides ultra-low latency of 30-70ms worldwide, enabling real-time responsive robot control.

Gemini Robotics 2

Google DeepMind, US · July 2026

Foundation VLAClosed weightsAlso written GR 2

Parameters: not published; No parameter count published.
GPU memory: not published; weights at bf16, computed
Inference latency: not published; per action step
Weights: closed; closed, private preview with waitlist

What it is

Gemini Robotics 2 is DeepMind's July 2026 VLA and the first to control a full humanoid body, legs, torso, arms and multi-finger hands, under one policy. It drives the 22 DoF SharpaWave hand on Apptronik Apollo 2 and reaches 89.6 percent on precise insertion and 78.9 percent on tool kitting with a Franka Duo. The failure pattern is informative: unscrewing a bulb succeeds 92 percent of the time while screwing one in succeeds 36 percent, and picking from the floor lands at 45.7 percent, so balance-critical and fine bimanual work remain open. All figures come from Google's own evaluations, with no technical report published. It is closed, in private preview, and gated behind a waitlist.

Architecture

Backbone: Not published. Released together with Gemini Robotics ER 2, which is based on Gemini 3.5 Flash.
Action head: Vision-language-action model producing whole-body motor control, covering legs, torso, arms and multi-finger hands under one policy. Inputs are text and image, output is robot action.
Parameters: No parameter count published.
Pretraining data: Not published.
Embodiments: Apptronik Apollo 2 humanoid with Inspire hands, Apptronik Apollo 2 humanoid with SharpaWave 22 DoF five-fingered hands, Franka Duo with Robotiq gripper, Dexmate, SO101, Trossen

What hardware it needs

This is the section most people came for, so it is worth being precise about which numbers are measured and which are arithmetic.

The authors publish no memory or latency figure for this model. Everything above is computed from the parameter count.

Reported results

Grouped by what the evaluation actually asked. Values from different suites are not comparable with each other, so each one keeps its suite and split.

Simulation

No results in this category are published for this model.

Real world

Rollouts on physical hardware. The setups differ, so read these as evidence, not as a ranking.

Multi-finger dexterity (Google DeepMind internal) Apollo 2 with SharpaWave hands, unscrew bulb
92%
success ratesource
Gripper dexterity (Google DeepMind internal) Franka Duo, precise insertion tasks
89.6%
success ratesource
Gripper dexterity (Google DeepMind internal) Franka Duo, diverse tool kitting
78.9%
success ratesource
General whole-body manipulation (Google DeepMind internal) Apollo 2 with Inspire hands, pick up from shelf
76.3%
success ratesource
General whole-body manipulation (Google DeepMind internal) Apollo 2 with Inspire hands, pick up from floor
45.7%
success rateThe weakest whole-body split, since picking from the floor requires squatting and balance recovery.source
Multi-finger dexterity (Google DeepMind internal) Apollo 2 with SharpaWave hands, screw bulb
36%
success rateScrewing in a bulb scores far below unscrewing it (92 percent), which shows how asymmetric the dexterity still is.source

Fine tuned tasks

No results in this category are published for this model.

On real hardware

Demonstrated on the Apptronik Apollo 2 humanoid walking to a table, picking up a watering can, stepping to a shelf and placing it. Controls the 22 degree-of-freedom SharpaWave five-fingered hand. Bi-arm results on Franka Duo reach 89.6 percent on precise insertion and 78.9 percent on tool kitting. All numbers are first-party internal evaluations.

Fine tuning it yourself

No public fine-tuning path. Access is by waitlist into an early-access partner program. Named partners include Apptronik, Franka and Agile Robots.

Where it helps, where it does not

Strengths

First model in the family to control a full humanoid, legs, torso, arms and fingers, under a single learned policy
Handles 22 DoF five-fingered hands, well beyond the parallel grippers of earlier Gemini Robotics models
Strong bi-arm industrial numbers: 89.6 percent precise insertion and 78.9 percent tool kitting on Franka Duo
Scales across a wide embodiment range, from SO101 class tabletop arms to humanoids

Limits

Every published number is a first-party internal benchmark with no technical report and no external replication
Whole-body picking from the floor is only 45.7 percent, so balance-critical manipulation is not solved
Screwing in a bulb reaches 36 percent against 92 percent for unscrewing, and tying a trash bag 44 percent, so fine bimanual dexterity remains weak
Closed weights, private preview with a waitlist
No parameter count, latency, control frequency or hardware requirement published

Sources

Everything on this page was taken from these documents. Where they disagree with what you read here, they win.

Entry last checked 2026-08-11.

Try a policy on a real arm

A physical SO-100 is online and free to drive in the browser, and five of the models in this arena can be fine tuned on a dataset you record with your own.

Download the client Drive a real arm now