The model can place one demonstration inside a 30-second context window and perform the task without gradient updates or fine-tuning. It can also adapt with only 1 to 10 gradient steps on five minutes of task data, while pretraining on large-scale physical interaction experience enables zero-shot physical generalization.
GEN-1.5 is useful for robot manipulation, human-to-robot teaching, sim-to-real transfer, and rapid adaptation to new physical environments. Its one-shot and few-shot behavior reduces the need for task-specific data collection and lengthy retraining for every new skill.
