Google DeepMind has announced two new AI models, Gemini Robotics and Gemini Robotics-ER, that significantly enhance robotic intelligence and dexterity, enabling machines to perform a broader range of real-world tasks with greater precision and adaptability.
Built on the foundation of Gemini 2.0, Google’s latest multimodal AI model, these innovations aim to bridge the gap between robot perception and physical action, making robots more interactive, adaptable, and responsive to human environments. “We’re drastically increasing performance in three key areas - generality, interactivity, and dexterity, with a single model,” said Carolina Parada, Senior Director and Head of Robotics at Google DeepMind. “This enables us to build robots that are more capable, more responsive, and more robust to changes in their environment.”
Gemini Robotics is a vision-language-action model that allows robots to process and act on new situations, even those they haven’t been explicitly trained for. Unlike previous AI-driven robotic systems, Gemini Robotics:
• Understands new environments dynamically without prior training. • Interacts with humans and surroundings in a more natural and responsive way. • Performs precise physical tasks, such as folding paper or removing a bottle cap, improving robotic dexterity.
Google DeepMind sees Gemini Robotics as a leap forward in creating general-purpose robots, capable of adapting to real-world scenarios autonomously.
Alongside Gemini Robotics, Google DeepMind also introduced Gemini Robotics-ER (Embodied Reasoning), an advanced visual language model designed to help robots understand and interact with complex, real-world environments. “For example, if you’re packing a lunchbox, you need to know where everything is, how to open the lunchbox, how to grasp items, and where to place them. That’s the kind of reasoning Gemini Robotics-ER enables,” explained Parada.
Gemini Robotics-ER is designed for roboticists to integrate with existing low-level controllers, making it easier to develop new capabilities powered by AI-driven reasoning.
With robots gaining greater autonomy, safety remains a top priority. Google DeepMind researcher Vikas Sindhwani emphasised that the company is developing a layered safety approach to ensure responsible AI deployment. “Gemini Robotics-ER models are trained to evaluate whether or not a potential action is safe to perform in a given scenario,” Sindhwani said.
DeepMind is also releasing new benchmarks and frameworks to advance safety research in the AI industry. This builds upon last year’s launch of Google DeepMind’s “Robot Constitution”, a set of AI safety rules inspired by Isaac Asimov’s laws of robotics.
Google DeepMind is working with top robotics firms, including Apptronik, Agile Robots, Agility Robotics, Boston Dynamics, and Enchanted Tools, as part of its efforts to build the next generation of humanoid robots. “We’re very focused on building the intelligence that understands the physical world and acts upon it,” Parada said. “We’re excited to leverage these models across multiple embodiments and applications.”