Recent breakthroughs in artificial intelligence are catalysing a revolution in autonomous robotics. While neural networks enable robots to execute increasingly complex tasks, scaling these capabilities exposes a critical bottleneck: data scarcity. Unlike the abundant text corpora utilised to train large language models, the action data required for robotic control is heavily reliant on slow and prohibitively expensive tele-operation. In this talk, Dr William Dally demonstrates how to overcome this limitation through advanced augmentation of human action data, effectively bridging the gap for robust robot training. Furthermore, practical deployment of these AI-driven systems demands highly efficient inference. This involves both efficient hardware and efficient models. Dr Dally will illustrate many of the techniques involved with an example of an efficient video model.