DeepMind RT-1: the first Transformer trained on real robotics data

DeepMind trained an AI model directly on physical robots, collecting 130,000 real demonstrations over 17 months. The result is RT-1, a transformer that understands natural language instructions and executes them on a robotic arm.

The surprising part is that the model works on tasks it never saw during training — this is called generalization, meaning the robot learned something deeper than individual movements.

Before RT-1, teaching a robot a new task required manual programming or task-specific training. Here, data and a transformer are enough.