Watch DeepMind's AI robot slam-dunk a basketball
Artificial-intelligence company Google DeepMind has recently incorporated a version of its most advanced large language model (LLM), Gemini, into robots. This advancement allows machines to complete tasks such as slam-dunking a miniature basketball through a desktop hoop without prior observation of another robot performing the action.
Advances in AI
Multiple companies, including Google DeepMind, are exploring the potential of harnessing artificial intelligence (AI) to develop general-purpose robots. Despite the exciting possibilities, there are concerns regarding the safety of these AI models due to their ability to generate incorrect or harmful outputs.
The goal is to create intuitive machines capable of handling various physical tasks independently, without human intervention or preprogramming. The integration of Gemini's robotic models enables developers to enhance their robots' capabilities, allowing them to comprehend natural language and have a deeper understanding of the physical world.
Gemini Robotics Model
The model, known as Gemini Robotics, represents a crucial step towards achieving this objective. Developed by Google DeepMind, the model leverages Gemini 2.0 — the company's most advanced vision and language model — which was trained by analyzing vast amounts of data to excel at tasks involving 3D physical and spatial understanding.
Furthermore, the model underwent additional training using data from numerous hours of real, remote-operated robot demonstrations. This training enabled the robotic 'brain' to execute real actions by leveraging learned associations, similar to how LLMs predict the next word in a sentence.
Performance Testing
The team at Google DeepMind rigorously tested Gemini Robotics on humanoid robots and robotic arms, evaluating their performance on both familiar and unfamiliar tasks. Results showed that robots utilizing the model consistently outperformed state-of-the-art competitors when tested on new tasks and modified versions of familiar activities.
For more information, you can access the articles on Gemini Robotics here and the technical paper.
