Deep reinforcement learning.

One paradox of AI is that things humans find challenging, such as complex calculations, machines can typically deal with easily, and yet tasks easily mastered by a human toddler often remain extremely challenging for machines to figure out.

In this example, a deep reinforcement learner learns how to dress itself through practice, trial and error. No doubt in a year or two similar systems will figure out how to tie shoelaces or a bow-tie. Procedural knowledge developed within a simulation can then be applied to embodied AI systems in the physical world.

An emerging further level of technology beyond Deep Learning is Deep Meta-learning systems. These train to a novice level across multiple tasks, which helps them to generalise and transfer learning between multiple domains, thus making an agent that is capable of navigating complex dynamic environments, albeit with less efficiency in a particular domain than a narrowly specialised one.

There's a super cute outtake right at the very end. Derp.