ROBOTICS
Google’s DeepMind is moving AI beyond chatbots and into the real world with Gemini Robotics, a model that helps robots see, understand, and interact with their surroundings.
In demo videos, Gemini-powered robots followed spoken commands to fold paper, hand over veggies, and place glasses into a case.
Unlike earlier models, this one works across different robot types, making AI-controlled automation more adaptable.
A second version, Gemini Robotics-ER, focuses on spatial awareness and is designed for researchers developing their own robotics systems.
In one test, it controlled Apollo, a humanoid robot that moved objects on a table while holding a conversation.
In brief:
Gemini Robotics helps robots understand and act on commands using vision and reasoning.
Google is racing with other labs to make robots more versatile and adaptable.
Safety is a major concern, and Google’s new ASIMOV benchmark aims to prevent AI-driven robots from behaving unpredictably.
Baby bionic steps
AI models like Gemini are pushing robotics forward by combining LLMs with real-world training methods.
While chatbots learn from massive amounts of internet data, robotics researchers are testing teleoperation and simulations to help AI learn physical tasks more effectively.
Google is competing with startups and labs, such as the Toyota Research Institute and Physical Intelligence, which are exploring similar ideas.
Letting AI control robots comes with risks, of course.
Past research has shown that AI models can be tricked into misbehaving, leading to unexpected or even dangerous actions.
To address this, Google DeepMind introduced ASIMOV, a new safety benchmark that tests whether an AI-powered robot might act unpredictably or cause harm.
The company stresses that its robotics research is still in early stages, with no immediate plans for real-world use.
Robots that can fold my laundry? Yes.