A small quadrupedal robot built by Boston Dynamics for the DARPA Learning Locomotion program

One of the simplest models of `walking'

a simple 3D dynamic biped that was able to quickly and reliably learn to walk. The beginning of the video (above) demonstrates the robot's ability to walk passively downhill on a treadmill with the computer turned off. The robot was then placed on flat terrain with the computer switched on, and tasked with acquiring the same gait without assistance from gravity, but rather by learning a feedback controller. This learning occurred in less than 20 minutes, using only trials implemented on the real robot (no simulations). The learning algorithm continues to quickly adapt as the robot walks over different terrain.

The first video (above) shows our glider (no propellors) landing on a perch. The glider enters the frame at 6m/s, and must come to rest on the "perch" that is just 3.5m away. The entire trajectory is just under 1 second; the video was shot with a high-speed camera and is played back 11x slower than real-time. The second video (above) show some preliminary flow visualization, which demonstrates the additional complexity in the fluid dynamics at high angles of attack.
 

The first video (above) shows the first flight of one of our robotic birds (piloted by a human). This bird was built in our lab, but is a replica of a hobbyist ornithopter designed by Sean Kinkade. The second video (above) shows a slow-motion shot of the flapping, which reveals the complicated interaction between the flexible wing and the surrounding flow. The third video (above) shows the first autonomous flight of our Phoenix ornithopter - a substantially redesigned machine with onboard sensing and computation.

The following video shows our experiments on an apparatus built by collaborator Jun Zhang at NYU. The experiment uses a symmetric flat plate as the simplest model of a flapping wing, and this video shows a variant with a flexible trailing edge. Using policy gradient learning run directly in the real experiment (no models), we have designed a control system which can achieve both higher speeds and a higher energy effectiveness than the initial sinusoidal gait.