Example: Machine Learning for Robotics

Net-Scale has worked together with New York University (Courant Institute of Mathematical Sciences) on several research projects with the goal of improving autonomous ground vehicle navigation through the application of machine learning techniques.

Autonomous ground vehicle control has potential vast commercial and military benefits but performance of current systems is not sufficiently robust to allow deployment. A key deficiency in current systems is that they cannot respond intelligently to the unexpected. Advances in machine learning for pattern recognition, control problems, and, in particular, end-to-end learning hold promise of a breakthrough.

The vehicle on the left was built and tested during the DAVE project (DARPA Autonomous Vehicle). It successfully learned to avoid natural obstacles on its own. The pixels from two cameras were directly fed into a neural network and its output went straight to steering and speed. No preprocessing or postprocessing nor manual tweaking of any system components took place. The only human input was through the hand labeled training data. For more information see an example video clip (MPEG, 9.2 MByte) filmed through the vehicle's stereo camera set or download the technical project report (PDF format, 3.1 MByte).

In the DARPA LAGR program (Learning Applied to Ground Robots) our team successfully implemented and field tested a new approach to complex off-road navigation based on near-to-far learning and the separation of tactical from strategic path planning. For more information, see the project home page or view the robot video (AVI, 5 Min, 117 MByte).