DeepMind Builds Gato, a Single AI Agent for Hundreds of Text and Robotics Tasks
DeepMind introduces Gato, a generalist AI agent that uses a single neural network to perform over 600 diverse tasks ranging from text chat to robotic block stacking.
Researchers at DeepMind introduce Gato, a single generalist agent that handles a wide variety of tasks across different modalities and embodiments. Inspired by the success of large-scale language models, this system uses a single neural network with shared weights to process diverse data types and actions.
Gato performs over 600 distinct tasks, demonstrating remarkable versatility. The same model plays Atari games, captions images, engages in conversational chat, and controls a real robot arm to stack blocks. It seamlessly decides what type of output to produce—whether text, joint torques, or button presses—based entirely on the context it receives.
This multi-modal, multi-task approach represents a significant shift away from specialized, single-purpose AI systems. By documenting Gato's current architecture and capabilities, the research team highlights the potential of training a single generalist policy to master a broad spectrum of real-world and virtual environments.