Gemini Robotics

Google DeepMind, US · March 2025

Foundation VLAClosed weightsAlso written Gemini Robotics 1.0
Parameters
not published
No parameter count was published for Gemini Robotics or for its Gemini 2.0 base model.
GPU memory
not published
weights at bf16, computed
Inference latency
250 ms
Split deployment: cloud-hosted VLA backbone plus a local action decoder on the robot's onboard computer (ALOHA 2 platform).
Weights
closed
no public checkpoint

What it is

Gemini Robotics is the first VLA in the Gemini family, built on Gemini 2.0 through a distilled Gemini Robotics-ER backbone. Its most useful engineering contribution is the split architecture, where the backbone is served from the cloud at under 160 ms and a local decoder on the robot expands action chunks, giving about 250 ms end-to-end latency and an effective control frequency of 50 Hz. It is data efficient in adaptation, reaching over 70 percent success on 7 of 8 new tasks from at most 100 demonstrations. Everything is closed and cloud dependent, and the non-ALOHA embodiment results came from task-specific specialists rather than one generalist. It has been superseded twice since, first by Gemini Robotics 1.5 and then by Gemini Robotics 2.

Architecture

Backbone
A distilled version of Gemini Robotics-ER, which is itself built on Gemini 2.0. Split architecture: the VLA backbone runs in the cloud, a local action decoder runs on the robot's onboard computer.
Action head
Cloud backbone emits an action representation, the on-robot decoder expands it into low-level action chunks containing multiple actions.
Control rate
50 Hz
Parameters
No parameter count was published for Gemini Robotics or for its Gemini 2.0 base model.
Pretraining data
A large-scale teleoperated action dataset collected on a fleet of ALOHA 2 robots over 12 months, consisting of thousands of hours of real-world demonstrations, on top of Gemini 2.0 multimodal pretraining.
Embodiments
ALOHA 2 (primary training platform), Bi-arm Franka with parallel grippers (adapted by fine-tuning), Apptronik Apollo humanoid with five-fingered hands (adapted by fine-tuning)

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.

Published figures. These come from the model authors, not from this site.
Inference latency, reported250 ms on Split deployment: cloud-hosted VLA backbone plus a local action decoder on the robot's onboard computer (ALOHA 2 platform). The report does not name the cloud accelerator or the onboard compute.
Control rate50 Hz

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.

  • Gemini Robotics dexterous long-horizon specialist tasks full long-horizon lunch-box packing
    100%
    success ratesource
  • Gemini Robotics dexterous long-horizon specialist tasks average over the specialist task suite
    79%
    success rateSpecialist models fine-tuned per task family, not the generalist checkpoint.source

Fine tuned tasks

What the model reaches after being adapted to a new task or a new robot, usually from a small number of demonstrations. This is the number that matters if you bring your own data.

  • Gemini Robotics embodiment adaptation bi-arm Franka, all considered tasks
    63%
    average success rateAfter fine-tuning onto the new embodiment. Compared in the report against a single-task diffusion policy baseline.source
  • Gemini Robotics few-shot adaptation 8 new short-horizon tasks with at most 100 demonstrations each
    7 tasks
    tasks reaching above 70 percent successThe report states over 70 percent success on 7 of 8 tasks, reaching 100 percent on two of them. 100 demonstrations correspond to roughly 15 minutes to 1 hour of collection.source

On real hardware

Real hardware throughout. Trained on an ALOHA 2 fleet over 12 months, then adapted to a bi-arm Franka (63 percent average success) and to the Apptronik Apollo humanoid. New short-horizon tasks reached over 70 percent success on 7 of 8 tasks with at most 100 demonstrations. End-to-end latency of roughly 250 ms yields an effective 50 Hz control frequency thanks to action chunking.

Fine tuning it yourself

Fine-tuning is described in the report but is only available to Google and partners. The published recipe uses at most 100 demonstrations per new short-horizon task, equivalent to 15 minutes to 1 hour of teleoperated collection, and a larger set for adapting to a new embodiment.

Where it helps, where it does not

Strengths

  • The split cloud backbone plus on-robot decoder gives 50 Hz effective control despite a large VLM in the loop, with about 250 ms end-to-end latency
  • Documented data efficiency: over 70 percent success on 7 of 8 new tasks from at most 100 demonstrations
  • Adapts to genuinely different embodiments including a full-size humanoid
  • Latency numbers are published explicitly, which is rare in this field

Limits

  • Closed weights, closed code, no self-hosting
  • The backbone runs in the cloud, so a network dependency is structural and offline use is impossible
  • Bi-arm Franka and Apollo results come from post-trained specialists with little generalization beyond trained task variations, as the successor report itself states
  • No parameter count, VRAM or hardware requirement was ever published
  • Superseded by Gemini Robotics 1.5 (September 2025) and Gemini Robotics 2 (July 2026)

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.