Gemini Robotics On-Device
Google DeepMind, US · June 2025
- Parameters
- not published
- No parameter count published for the 1.0 on-device model.
- GPU memory
- not published
- weights at bf16, computed
- Inference latency
- not published
- per action step
- Weights
- closed
- closed, trusted tester program only
What it is
Gemini Robotics On-Device is the first compact VLA in the Gemini family, built to run locally on bi-arm robots without a network connection. Its practical selling point is adaptation from 50 to 100 demonstrations plus dexterous behaviors such as unzipping bags and folding clothes, demonstrated on ALOHA, a bi-arm Franka FR3 and the Apollo humanoid. For anyone doing hardware selection it is effectively opaque, because no parameter count, latency figure or minimum compute specification was ever published. Access requires the trusted tester program. It was replaced by Gemini Robotics On-Device 2 in July 2026.
Architecture
- Backbone
- A compact derivative of the Gemini Robotics VLA, engineered to require minimal computational resources and to run locally on the robot. No public technical report accompanies it.
- Action head
- Vision-language-action model for bi-arm manipulation. The release material does not specify the action head or chunk size.
- Parameters
- No parameter count published for the 1.0 on-device model.
- Pretraining data
- Derived from the Gemini Robotics training data, primarily ALOHA demonstrations. No corpus size published.
- Embodiments
- ALOHA (primary training platform), Bi-arm Franka FR3, Apptronik Apollo humanoid
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.
No minimum hardware specification was published. The blog only states that the model is engineered to require minimal computational resources.
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
No results in this category are published for this model.
Fine tuned tasks
No results in this category are published for this model.
On real hardware
Runs on the robot itself with no network dependency and performs dexterous tasks such as unzipping bags and folding clothes. Adapts to new tasks from as few as 50 to 100 demonstrations. It was demonstrated on ALOHA, a bi-arm Franka FR3 and the Apollo humanoid. The release blog reports generalization results only as a chart, so exact per-split numbers are not quoted here.
Fine tuning it yourself
Adaptation to a new task is reported to need as few as 50 to 100 demonstrations. The tooling is the Gemini Robotics SDK, which is available only after acceptance into the trusted tester program.
Where it helps, where it does not
Strengths
- Runs entirely on the robot, which removes the network dependency that constrains the cloud Gemini Robotics models
- Adapts to a new task from 50 to 100 demonstrations
- A Gemini Robotics SDK is offered so testers can evaluate it on their own tasks and environments
- Demonstrated on three different embodiments including a humanoid
Limits
- Closed weights, gated behind a trusted tester program, so it cannot simply be downloaded
- No parameter count, latency figure, memory requirement or target compute module was ever published, which makes hardware planning impossible from public data
- No technical report, only a blog post and a model card
- Bi-arm manipulation only
- Superseded by Gemini Robotics On-Device 2 on 30 July 2026
Sources
Everything on this page was taken from these documents. Where they disagree with what you read here, they win.
- https://deepmind.google/blog/gemini-robotics-on-device-brings-ai-to-local-robotic-devices/
- https://arxiv.org/abs/2510.03342
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.