LAGR
Customer: DARPA
Role: Prime Contractor
Duration: 12/2004 – 11/2008
Purpose: Unmanned Ground Vehicle Navigation
Technologies: Machine learning, Stereo vision, Image processing
Capabilities: Fully autonomous vehicle navigation, long-distance perception, on-line adaptation
Description: The LAGR (Learning Applied to Ground Robots) program was designed to foster application of learning techniques to improving autonomous navigation for ground vehicles in natural terrain.
We were one of eight teams funded to participate in this program. Each team was issued identical robots which had only stereo vision cameras and GPS for sensing. DARPA conducted many series of monthly tests (27 tests in all) at a variety of sites over the course of the program. For these tests, teams sent a FLASH disk with their system, which the LAGR government team (LGT) would plug into one of their (identical) robots to run the test. In general, the teams were not on-site for the tests and had only general information about the course and the type of terrain at the test site. Most tests required the robot to navigate to a GPS waypoint over natural terrain, but some tests were designed to evaluate specific capabilities, such as learning from example or long-distance perception.
We developed the following technologies under this program:
- A convolutional neural network terrain classifier - This neural network takes data from stereo vision and estimates terrain cost, i.e., how "costly" it is to drive over a specific area. Our terrain classifier excels in difficult environments: those with tall grass, overhangs, brush, and low vegetation.
- A Bayesian learning method for learning from example - From a short (5-10 second) tele-operated run, our system can learn to identify low-cost terrain. This component of our system has been very successful in learning to drive on paths and dirt roads.
- Image-based planning - We have developed a technique for planning a path in a monocular image to the projected goal point in the image. The resulting path is used to guide our Cartesian path planner. This was one of our approaches to long-distance perception, i.e., extracting information from well beyond the limited stereo-vision range in order to avoid myopic behavior.
- Cartesian path planner - We developed a component that plans a path for the robot over a cost map, applying limitations on the path curvature in order to produce smooth robot behavior.
- Dynamics-based controller - We developed a path-following controller for the robot that incorporates a dynamic model of the robot.
We are continuing work on the LAGR robot platform to develop a system that can negotiate dynamic environments.