Google unveils AI robotics models | LinkedIn
Google offered a peek at its latest ventures in artificial intelligence-powered robotics this week, unveiling two new models which run on its Gemini 2.0 AI platform. In a video demonstration, researchers from Google's DeepMind division showcased robots performing tasks such as packing lunchboxes and dunking toy basketballs in response to human commands, as well as adapting to changing environmental situations. Google also announced that it will partner with robotics developer Apptronik to build newer models.
First-Hand Experience with AGI
2 years ago, there was a question on what it would look like to bring the predict-next-token philosophy to robotics and put it in scale while finetuning Large Language Model at Google Search. Now at GDM Robotics, watching it come to life in person has been an exciting journey. Here are some first-hand AGI moments experienced:
- Super fast adaptation, showcasing the model's ability to pick up socks from anywhere on the table after only 5 demos.
- Generalization capability demonstrated through using mirror actions immediately after training with a single hand.
- Gemini's ability to draw letters and perform random tasks on a casually arranged table scene.
For more details, refer to the tech report here.
Gemini Robotics: Advancements and Capabilities
Gemini Robotics, the latest AI models by Google, are designed to usher in a new era of helpful robots. These models can perform various complex tasks that demand dexterity, such as folding origami, packing a lunch box, playing tic-tac-toe, and more.
Gemini Robotics-ER brings advanced world understanding to robotics, enabling tasks that the robots haven't been specifically trained on. This is achieved through perception, spatial understanding, planning, and more.
Google looks forward to further developing these models and collaborating with a range of trusted testers to integrate Gemini into the physical world.
Find out more about Gemini Robotics here.
