Perceptron trained Isaac on huge amounts of cheap video alongside expensive robot demonstrations, and their experiments suggest that enough video can drastically reduce how much teleoperation data is needed to reach the same action-learning quality. https://t.co/iSTSZCipLe