From the ROS blog:
The Robotics Institute at CMU has been developing systems to learn from humans. Using a Machine Learning class of techniques called Imitation Learning the group has developed AI software for a small commercially available off-the-shelf ARdrone to autonomously fly through the dense trees for over 3.4 km in experimental runs. They are also developing methods to do longer range planning with such purely vision-guided UAVs. Such technology has a lot of potential impact for surveillance, search and rescue and allowing UAVs to safely share airspace with manned airspace.
Watch the flight in unstructured environments below.
Comments
This is amazing for several reasons. 1st they only used a camera as the main sensor. 2nd the optical flow routine & vision based learning algorithm looks simple. 3rd the payload capacity of the Ardrone is very limited. I assume the autopilot was based on the px4 fmu like the one that 3dr has & the px4ioar. CMU article at
http://www.ri.cmu.edu/pub_files/2012/11/1211.1690v1.pdf
Thanks Chris & CMU team.