LAGR - Learning Applied to Ground Robots

Customer: DARPA
Role: Prime Contractor
Duration: 12/2004 – 4/2008
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 program aims to develop completely autonomous off-road ground vehicle navigation using machine learning techniques. Eight research teams were funded to participate in this program, and each research team is issued a standard ground vehicle platform. For this program, energy-emitting sensors (like ladar, sonar, or radar) are prohibited; instead, the vehicle must rely purely on passive sensing, such as video imagery and GPS receivers.

Each month, DARPA conducts tests of all the systems on a new test course of their own design.  Each robot is started in the same location, and is given a selected GPS waypoint as the goal. The LGT scores each vehicle on the basis of the time taken to reach the goal, (or, if the robot fails to achieve the goal in the allotted time, the remaining distance to the goal).  The teams participating in the program provide the LAGR Government Team (LGT) with a flash disk containing their executables before each test. The LGT has identical robotic platforms to those of the participants, and so can simply plug in the teams’ flash disks and conduct tests at sites selected by the LGT. Participants are generally unaware of the exact nature of the test courses, so there is little opportunity for tuning a system for particular terrain.

API’s accomplishments during the thirteen tests conducted during Phase I include:

  • Most runs completed, and highest average score, of any team over all the tests in Phase I, including five first-place finishes (Tests 2, 4, 6, 8, 10) and five second-place finishes (Tests 3, 5, 7, 9, 12).
  • Successful development of neural-net based local obstacle detection.
  • Successful development of Bayesian framework for learning-from-example. System chose the correct route on all 3 runs on first learning-from-example test (Test #8) and all 4 runs of the second  learning-from-example test (Test #12).
  • Successful development of on-line image-to-geometry learning for long-distance and monocular vision. Fewest bumper hits and best score of any team on monocular learning test (Test #10).
  • Clear ability to label long-distance terrain costs in image space. Demonstrated ability to label and avoid obstacles at distances up to three times farther than stereo range.
  • API has recently been selected for Phase II participation in this program, and has also been asked by DARPA to begin transitioning some of this technology to the NREC Crusher vehicle.
CURRENT PROJECTS:
Processed stereo data showing grassy terrain and a tree. Terrain traversal cost is shown in grayscale, from low cost (gray) to high cost (white).
Two plots of stereo data taken from a path through the forest. The top plot is shown in natural color; the bottom plot is encoded with terrain traversal cost, from low cost (gray) to high cost (white).
The LAGR test vehicle.
On-line learned terrain cost labeling in image space at long distance. Low cost terrain is shown in black, and high-cost terrain is shown in white.
HOME | ABOUT US  | CONTACT |  NEWS | PROJECTS | SERVICES | JAUS