Simple & efficient trajectory learning for smooth navigation

Video showing the simple and efficient trajectory learning method used for reactive obstacle avoidance in the DARPA LAGR (Learning Applied to Ground Robots) project by New York University team. This technique is compared to a trivial linear steering method. The robot never hits any obstacles with the trajectory learning method while requiring very little compute power.
More details at http://www.cs.nyu.edu/~yann/research/... and http://sermanet.free.fr/

• Category

  Science & Technology

[nyulagr]

That is precisely the point of the video, it compares 2 methods of driving the robot given the stereo map of obstacles. The first one is the trivial linear way, i.e. the commands sent to the motors are proportional to the direction you want to go. This methods has a red caption on top in the video, and that's when the robot crashes. The main problem is that the robot has inertia. The second method "learns" that among other things and does not crash anymore (green caption in the video).

[Golden Lion]

The stereographic vision works to determine distance. Why does the robot sometimes seem to navigate the obstacle barriers with easy and other this becomes confused, crashing into the barriers, and then recalibrating a new route?