Introduction
Generalist AI introduced the GEN-1.5 model, which aims to enable robots to learn new physical tasks from a single demonstration. The model can apply 3- to 12-second demonstrations together with sensor, video, language, position and motion data, without fine-tuning or gradient updates. Using a 30-second context window, GEN-1.5 generates 100 motion commands per second.
Developments
GEN-1.5 achieved an average success rate of 59% from a single demonstration across 10 short-horizon manipulation tasks. With five minutes of data containing approximately 50 demonstrations per task and 10 gradient steps, this rate rose to 83%.
Details
A success rate of 66.5% was measured with one minute of data and a single gradient step.
The model can combine different demonstrations to create long task sequences, develop intermediate actions on its own and transfer demonstrations in simulation to a real robot zero-shot. It was stated that some of the model’s in-context learning and improvisation capabilities emerged spontaneously during its pretraining, which has continued for more than eight months.