Researchers Peter Abbeel and John Schulman discuss their use of reinforcement learning to make a helicopter fly upside down and do stunts, including the challenges they faced with GPS units and positioning.