The system built for Dota, applied to tasks in the physical world: . Step towards truly general-purpose AI systems.
My favorite part of the result: the robot's grasps are recognizable and can be labeled with the standard taxonomy of human grasps. These grasps were discovered entirely in simulation from scratch, and run on the physical robot. 
Robotic hands have been on the market for decades, but they have been unusable since no one could program them — in contrast to walking or backflipping robots. Learning has enabled something that was unachievable. Compare also to existing hand results: youtube.com/playlist?list=…
Study testing human children on a related in-hand rotation task: ajot.aota.org/pdfaccess.ashx…. Not trivial!
