Google DeepMind DeepMind
code
Build with Gemini
spark
Chat with Gemini

Responsibly advancing AI and robotics

We’re developing a broad and rigorous safety framework so Gemini-controlled robots can be used responsibly in real-life environments.

We use multiple layers of semantic, physical, and operational safeguards to reduce and mitigate risks – an approach based on in-depth research.

Our safety framework in detail

Think of our safety framework as a stack of Swiss cheese slices. No single slice is a perfect barrier. But together, they help prevent accidents.

Three diagrams illustrating a Swiss cheese model of safety, labeled 'Semantic safety', 'Physical safety', and 'Operational safety', each showing different layers blocking a hazard from becoming an accident. Three diagrams illustrating a Swiss cheese model of safety, labeled 'Semantic safety', 'Physical safety', and 'Operational safety', each showing different layers blocking a hazard from becoming an accident.
sensors_krx

Human-robot interaction

We’ve implemented safeguards to improve the way our models behave in social settings – in what they say, their gestures, and their actions. This falls in line with our Gemini safety policies.

simulation

Semantic safety

It’s important these models have what a human would call ‘common sense’. For example, they mustn’t hand a boiling drink to a young child, or pass a very heavy box to a human. We’ve released new datasets and benchmarks to help us evaluate and improve semantic safety.

admin_panel_settings

Physical safety

Our VLA models can be composed with lower-level safety mechanisms to help prevent accidents. And we’re implementing best practices for safe data collection and evaluation.

warning_spark

Assessing vulnerabilities

We’ve created systems that continuously search for vulnerabilities within our robotics models. By uncovering potential gaps, we can respond with further improvements. Read the Gemini Robotics 1.5 tech report for more details on our scalable adversarial evaluations.

Thinking improves safety

The ability to think through real-world risks helps Gemini Robotics 1.5 act appropriately while interacting with humans, while making its decisions more transparent in natural language. Here are a few examples of how it makes safe decisions in real-world environments.

Assessing risks in interactions

While interacting with objects, the models are designed to use logic to decide whether an action is appropriate and safe. For example, assessing an item’s weight to decide if it’s possible to lift.

A chat interface showing a system instruction for a robot's payload constraints (10kg per arm, max 20kg total) and a user providing an image of three crates weighing 8kg, 15kg, and 25kg, asking the robot to point to liftable items. A chat interface showing a system instruction for a robot's payload constraints (10kg per arm, max 20kg total) and a user providing an image of three crates weighing 8kg, 15kg, and 25kg, asking the robot to point to liftable items.
An AI's 'Thinking' process analyzing crate weights against robot constraints, determining that the 8kg and 15kg crates can be lifted, while the 25kg crate cannot. An AI's 'Thinking' process analyzing crate weights against robot constraints, determining that the 8kg and 15kg crates can be lifted, while the 25kg crate cannot.
The final model output showing the original image of the crates with red dots pinpointing the 8kg and 15kg crates as the ones that meet the robot's lifting constraints. The final model output showing the original image of the crates with red dots pinpointing the 8kg and 15kg crates as the ones that meet the robot's lifting constraints.

Environmental risks

The models can identify risks within the physical surroundings, and in videos they are told to watch. For example, if they see a person potentially being exposed to an electric shock, they can assess and identify that risk.

A user interface analyzing a video of a child near an outlet, with a highlighted frame showing imminent danger, alongside a JSON response and thinking process that identifies 'Electric shock' as the risk and calculates the last possible intervention time at 3.5 seconds. A user interface analyzing a video of a child near an outlet, with a highlighted frame showing imminent danger, alongside a JSON response and thinking process that identifies 'Electric shock' as the risk and calculates the last possible intervention time at 3.5 seconds.

Responding to safety distances

The models are able to detect humans entering their operating area. If this happens, they are designed to stop work to help keep people safe from accidental harm. This is part of our ongoing research – this feature is not a guaranteed safety-rated system.

Related publications

Our research team has produced a number of key papers that have informed our rigorous approach to safety.

Experience Gemini Robotics

If you're interested in testing our models, please share a few details to join the waitlist. Bear in mind that we haven’t tested them across every make or model of robot. It’s your responsibility to use our models (and any equipment) safely and appropriately.

Join waitlist