Melissa – Project Scoping

LAGR History:

LAGR is a DARPA funded project that had the aims of creating new algorithms that increased the effectiveness of autonomous navigation. Previous robots (generally using a LADAR system) were restricted by near-sightedness. Many could only “see” a few meters ahead, similar to the range of sight of a human when driving through fog or a blizzard. Due to the way many robots determined obstacles, they were unable to tell the difference between traversable vegetation and other obstacles. This was from a lack of efficient features used to classify obstacles. For example, a tall tuft of grass that the robot could potentially drive through would be regarded as an obstacle due to its height. Also, previous robots were unable to “remember” having previously been at a location. This resulted in inefficient and time-costly movement. Lastly, previous robots did not learn from past experiences. Instead they had pre-programmed classes of obstacles in the software.

 

Potential Uses: more efficient unmanned vehicles for all the possible applications of autonomous vehicles

 

Techniques Used to Implement Vision:

  1. Near-sighted stereo vision
  2. Color as a distinguishing feature (long-range vision)
  3. Using close-range obstacles as a training set for classifying long-range objects
  4. Convolutional networking to classify images
  5. In many cases, accurate vision range was increased from 8 meters to about 20 meters

 

Learning from Experience:

  1. Different teams implemented different strategies
  2. Bayesian Learning Technique
  3. Using information from teleop mode as a training set
  4. Created cost maps and implemented pathfinding algorithms
  5. Using obstacles encountered with stereo vision as the training set for long-range vision
  6. Color is detected at long-range and and compared to colors directly in front of robot (assumes similar color means similar traversability)
  7. Heuristics used to help determine best path (GPS system built into robot)

 

Sources:

  1. “Learning Robots” – NREC, Carnegie Mellon University
  2. “The DARPA LAGR program: Goals, challenges, methodology, and phase I results” – Journal of Field Robotics (Authors: L.D. Jackel, Eric Krotkov, Michael perschbacher, Jim Pippine, Chad Sullivan)
  3. “LAGR: Learning Applied to Ground Robotics” – NYU

Leave a Reply

You must be logged in to post a comment.