Autonomous robots that can learn using neural networks are advancing at an incredible pace. The DARPA LAGR Program aims to advance learning algorithms in off road robotics. The main challenge in off road robotics is the highly changing and unstructured environment. Learning algorithms can adapt to changes in lighting conditions and obstacles appearance where conventional algorithms often fail.

The LAGR vehicle comes with four cameras allowing passive stereo vision, a GPS receiver used to reach a GPS goal, and four onboard computers. Stereo vision is typically reliable only up to about 4 meters. Increasing this range with passive vision remains highly challenging. A short sighted robot makes poor decisions and will get stuck in couldesacks.

NetScale Technologies and New York University have focused their efforts on developing machine learning techniques for long vision along with a robust and collision free software navigation system. This long range vision system increases the vision range of the robot to 50 or even 100 meters or more.

The autonomous robot first finds the horizon using a ground plain extracted from stereo points. In then builds a series of normalized image bands at different distances so that it can train on and classify objects of all sizes. The neural network is trained both online and offline using a short range stereo label such as traversable and obstacle to learn the appearance of natural obstacles which are then classified in the long range patches of the image. This technique has been dubbed Near Far Learning and seems to be the most promising approach for true autonomous robots of the future.

Neural networks are particularly well suited for autonomous robots because they extracting features from images while providing scale and translation in variance. Two layers of convolution separated by a max pooling layer produce a 100 dimensional feature vector for each point in the image bands. The feature extractor is trained offline using logs captured in the full duration of the program.

Below is a video showing various autonomous robots, with the advanced learning algorithm, in action. For more information about the latest in robotic learning algorithms go to http://www.cs.nyu.edu/~yann/research/lagr/index.html

Duration : 0:5:1