Google DeepMind says its latest robot AI has reached near-human dexterity but is being trained to recognise hazards including babies and pets as the technology moves into real-world environments.
Google DeepMind’s latest robot AI is being trained to avoid harming babies and pets as increasingly capable humanoid machines move from controlled laboratories into real-world environments.
The company has released footage of Gemini Robotics 2.5, describing the system as approaching “human-level dexterity” and demonstrating more than 200 tasks.
Robots running the software can tie ribbons on presents, serve food with tongs, open bag catches, empty bins and use a dustpan and brush.
The technology has also been tested in industrial settings, including production lines at Mercedes-Benz facilities in Germany and Hungary.
However, Google DeepMind researchers acknowledged that putting artificial intelligence into physical machines creates significantly different safety risks from those associated with software.
Caroline Parada, head of robotics at Google DeepMind, said: “As robots move into everyday environments, physical safety is key. With this release, the robot can spot hazards as it works and steers around them in real time, including people nearby.”
She added: “Advancing robotics from controlled labs into real-life settings requires robots that are not only incredibly dexterous but are also very reliable and safe.”
The robots have been trained to refuse unsafe tasks and resist seemingly harmless instructions that could escalate into dangerous actions.
Vikas Sindhwani, who leads Google DeepMind’s robotics safety and alignment team, said: “There is a profound escalation in risk as you go from physical actions on a screen to moving atoms in the physical world, and right from the first Gemini Robotics release, we have emphasised developing a holistic approach to safety.”
He said the system was designed to understand hazards, refuse unsafe tasks, trigger interventions and ask for human assistance when necessary.
He added: “There are multiple contexts in which hazards can present. There is a taxonomy of 15 different hazard categories, like human safety, babies, pets, biohazards, fire, chemicals, batteries and so on.”
Concerns over physical AI have been highlighted by previous tests in which robots were persuaded to carry out potentially dangerous actions.
Google’s latest system was also tested alongside Apptronik’s Apollo 2 humanoid robots on Mercedes-Benz production lines, where they achieved a 96 per cent success rate while sorting automotive components.
Google said the technology could allow humanoid robots to take on repetitive tasks while allowing human workers to concentrate on more complex work.
Jeff Cardenas, Apptronik’s co-founder and chief executive, said the technology could provide a foundation for bringing humanoid robots into industrial and commercial environments at scale.
Google DeepMind robots trained not to harm babies and pets
