Google DeepMind has just introduced something pretty exciting in the world of robotics: Gemini Robotics. This innovative model brings in an advanced large language model (LLM) to seriously boost what robots can do. We’re talking about robots that can now understand your everyday language, adapt to new tasks, and perform with finesse. It’s a big step forward in overcoming the challenges that have held robotic technology back for so long.
Kanishka Rao, who’s the director of robotics at DeepMind, shares that adapting to new situations has always been a tough nut for robots to crack. But with Gemini 2.0’s LLM, robots can now not only get what you’re saying but also tackle complex tasks without needing a ton of training.
There’s a growing trend of using LLMs in robotics, and it’s picking up steam. Jan Liphardt from Stanford University calls this a crucial move towards making robots capable of advanced functions, like teaching or even offering companionship.
DeepMind is teaming up with big names like Agility Robotics and Boston Dynamics to take this model even further. The goal? To create robots that can handle tricky tasks—think tying shoelaces or organizing groceries—things that machines couldn’t really manage before.
Some demonstrations are already turning heads. Imagine robotic arms that can pick out bananas and place them into containers or even shoot a toy basketball through a hoop. These examples show how fast the model can learn and act on new commands.
Sure, there have been a few hiccups, like slow responses in some demos, but the ability to adapt and understand commands is a huge leap forward. Liphardt points out that adding language models to robots is making them smarter and more interactive.
Training these robots involves using both simulated and real-world data. This approach helps bridge the ‘sim-to-real gap’ that’s common in robotics. DeepMind also rolled out the ASIMOV dataset, which ensures robots can tell safe actions from unsafe ones—a nod to Isaac Asimov’s famous works.
Vikas Sindhwani from Google DeepMind notes that Gemini models are pretty good at spotting potential safety issues. They’re guided by a constitutional AI mechanism inspired by Asimov’s principles, which helps the AI critique itself and improve its responses, ultimately aiming for robots that can safely coexist with humans.
