Gemini Robotics On-Device 2
Google DeepMind, US · July 2026
- 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, trusted tester program only
What it is
Gemini Robotics On-Device 2 is the compact VLA of the July 2026 release, built on the on-device Gemma models with Gemini Robotics 1.5 technology and intended to run locally on bi-arm robots without a network. Its practical claim is fast embodiment adaptation, a few hours and typically fewer than 200 examples, with named evaluations on SO101 and Dexmate. Google states its own limits plainly: weak out-of-distribution generalization, poor control of high degree-of-freedom robots, and mobile or whole-body control out of scope. No benchmark numbers, parameter count, latency or hardware requirement are published, so it cannot be sized against open alternatives. Access is restricted to trusted testers.
Architecture
- Backbone
- Based on the on-device Gemma models and on Gemini Robotics 1.5 technology, per the model card. Trained on Google TPUs.
- Action head
- Vision-language-action model for bi-arm manipulation. Inputs are image, text and action, output is a stabilization action per the product page. No chunk size published.
- Parameters
- No parameter count published. The model card names the on-device Gemma models as the base but gives no size.
- Pretraining data
- Datasets of images, text, and robot sensor and action data, processed with deduplication, safety filtering and quality filtering. No corpus size published.
- Embodiments
- Bi-arm robots in general, SO101 (named in the model card evaluation), Dexmate (named in the model card evaluation), Trossen (named in the release blog)
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 published. The model card states only that it runs efficiently on local devices.
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
Adapts to a new bi-arm embodiment in a few hours of adaptation time, typically with fewer than 200 examples. Evaluations named in the model card cover SO101 and Dexmate platforms. No numeric benchmark table was published for this model.
Fine tuning it yourself
Adaptation to a new bi-arm embodiment is reported to take a few hours and typically fewer than 200 examples. Access requires acceptance into the Trusted Tester Program.
Where it helps, where it does not
Strengths
- Runs locally on the robot, so it works with intermittent or zero connectivity
- Adapts to a new bi-arm embodiment with fewer than 200 examples and a few hours of training
- Explicitly evaluated on low-cost platforms such as SO101 and Dexmate, which is unusual for a frontier lab model
- Inherits Gemini Robotics 1.5 technology in a compact package
Limits
- Google itself states it is limited in generalizing to out-of-distribution tasks and in controlling high degree-of-freedom robots
- Mobile platforms and whole-body control are explicitly outside its scope, so humanoids need Gemini Robotics 2 instead
- No numeric benchmarks, parameter count, latency or minimum hardware specification published, which makes it impossible to plan compute from public data
- Closed weights, gated behind the trusted tester program with a waitlist
Sources
Everything on this page was taken from these documents. Where they disagree with what you read here, they win.
- https://deepmind.google/models/model-cards/gemini-robotics-on-device-2/
- https://deepmind.google/models/gemini-robotics/on-device/
- https://deepmind.google/blog/gemini-robotics-2-brings-whole-body-intelligence-to-robots/
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.