How we built the new family of Gemini Robotics models
Powered by Gemini Robotics models, robots can learn complex actions like preparing salads, playing games like Tic-Tac-Toe and even folding an origami fox.
As Google DeepMind prepared for its recent announcement of a new family of Gemini 2.0 models designed specifically for robots, its head of robotics, Carolina Parada, gathered her team for another check of the tech’s capabilities.
They asked a bi-arm ALOHA robot — a duo of limber metal appendages with multiple joints and pincer-like hands used widely in research — to perform tasks it hadn’t done before, using objects it hadn’t seen. “We did random things like put my shoe on the table and ask it to put some pens inside,” Carolina says. “The robot took a moment to understand the task, then did it.”
For the next request, they found a toy basketball hoop and ball and asked the robot to do a “slam dunk.” Carolina watched, proud and delighted, as it did just that.
