In science fiction stories, artificial intelligence often runs all kinds of smart, capable, and sometimes killing robots. The obvious limitation of today's best AI is that for now it remains straight locked inside the chat window.
Google Deepmind signaled plans to change today (presumably minus the murderous portion) by unveiling a new version of Gemini AI model, which drives a range of more capable, adaptive, and potentially useful robots by fusing language, vision, and physical behavior.
In a series of demo videos, the company will showcase several robots equipped with a new model called Gemini Robotics, and manipulate items in response to voice commands. I handed over the robot arms to fold paper, handing over vegetables, gently putting glasses in the case to complete the other tasks. The robot relies on the new model to connect items that are displayed with possible actions to do what is displayed. This model is trained to generalize its operation on very different hardware.
Google Deepmind has also released a version of a model called Gemini Robotics-ER (for embodied inference). The idea is for other robot researchers to use this model to train their own models to control the actions of the robot.
In the video demonstration, researchers at Google Deepmind used this model to control a humanoid robot called Apollo from startup Apptronik. The robot talks to humans and moves letters around the table top when instructed.
“We have been able to bring global understanding of Gemini 2.0, a general concept to robotics,” said Kanishka Rao, robot researcher at Google Deepmind, in a briefing ahead of today's presentation.
Google Deepmind said the new model can successfully control various robots in hundreds of specific scenarios that were not previously included in training. “Having a general conceptual understanding of the robotic models can be much more common and useful,” Rao said.
The breakthroughs that created powerful chatbots, including Openai's ChatGpt and Google's Gemini, have raised hopes for a similar revolution in robotics in recent years, but major hurdles remain.
