Google DeepMind has introduced two new AI models designed to enhance robotic capabilities in physical environments.
The announcement on 12 March 2025 revealed Gemini Robotics and Gemini Robotics-ER (embodied reasoning), both developed using the company’s Gemini 2.0 foundation.
The models aim to expand the range of tasks robots can perform. Gemini Robotics incorporates physical actions as an output modality for direct robot control, while Gemini Robotics-ER focuses on spatial understanding for roboticists implementing their own programs.
During a press briefing, Carolina Parada, senior director and head of robotics at Google DeepMind, stated the models improve performance in three areas: generality, interactivity, and dexterity. The company claims Gemini Robotics doubles performance on generalization benchmarks compared to existing systems.
“While we have made progress in each one of these areas individually in the past with general robotics, we’re bringing drastically increasing performance in all three areas with a single model,” Parada said.
Testing demonstrations show robots using the technology to complete tasks such as paper folding, placing glasses in cases, and manipulating objects in response to voice commands. The company reports that Gemini Robotics allows robots to perform well even in environments not included in the training data.
According to Google DeepMind, the models can understand conversational language instructions and adapt to changing environments. When objects move or slip from a robot’s grasp, the system can quickly adjust its approach and continue with the task.
On safety considerations, Google DeepMind researcher Vikas Sindhwani explained that the company has implemented systems to evaluate whether actions are safe in given scenarios. “Gemini Robotics-ER models are trained to evaluate whether or not a potential action is safe to perform in a given scenario,” he noted.
The company has also released a dataset called ASIMOV for measuring safety implications of robotic actions, building on previous work around robot safety protocols.
Google DeepMind has partnered with Apptronik for humanoid robot development and granted access to its Gemini Robotics-ER model to selected partners including Agile Robots, Agility Robotics, Boston Dynamics, and Enchanted Tools.
The announcement comes amid growing advancements in the robotics sector, with increasing AI capabilities accelerating commercial applications, particularly in industrial settings. Google’s launch follows recent developments in the field, including Figure AI’s reported breakthrough in AI for robots last month.
While trained primarily on data from the ALOHA 2 bi-arm robotic platform, the company states that Gemini Robotics can be adapted for different robot types, including more complex humanoid systems.
