Google's new AI lets robots perform tasks without prior training
What's the story
Google DeepMind has unveiled two groundbreaking AI models, Gemini Robotics and Gemini Robotics-ER, to improve robots' physical capabilities.
The first model, Gemini Robotics, is a vision-language-action system that can adapt to new situations without any prior training. This sophisticated model is based on Gemini 2.0, Google's flagship AI model.
Carolina Parada from Google DeepMind said this new tech "draws from Gemini's multimodal world understanding and transfers it to the real world by adding physical actions as a new modality."
Model capabilities
Gemini Robotics excels in 3 key areas
Gemini Robotics marks a major leap in three important areas for robotics to work: generality, interactivity, and dexterity.
The system can adapt to new scenarios better than its predecessors. It also interacts better with humans and its environment, showing improved precision in tasks like folding paper or opening a bottle cap.
Parada stressed these advancements make it possible to create robots that are "more capable, responsive, and robust to changes in their environment."
Advanced reasoning
Gemini Robotics-ER: A step forward in visual language understanding
The second model, Gemini Robotics-ER, is an advanced visual language model that can comprehend complex and dynamic environments.
Parada explained that this tech would be capable of tasks like packing a lunchbox by identifying items on a table, knowing how to open the lunchbox and grasp items correctly.
This system is meant for roboticists to connect with existing low-level controllers — systems controlling robot movements — enabling new capabilities powered by Gemini Robotics-ER.
Safety measures
Google DeepMind prioritizes safety in AI model development
In terms of safety, Google DeepMind is taking a "layered-approach" for its AI models. The company has also introduced new benchmarks and frameworks to further enhance safety research within the AI industry.
Last year, it launched its "Robot Constitution," a set of rules inspired by Isaac Asimov's work for its robots to follow.
Parada said they are "very focused on building the intelligence that is going to be able to understand the physical world and act on it."