Learning Applied to Ground Robotics (LAGR)

The DARPA LAGR program has the challenging aim of producing a robot that can autonomously navigate in varied outdoor environments from tall grass fields to heavily wooded terrain. The robot platform on which this task is to be accomplished has two stereo pairs, which are its only means of perception.

My primary task in the project was the creation of a 6-DOF pose estimation algorithm that samples various measurement streams, such as GPS, IMU, and visual odometry, and computes an optimal estimate of the robot pose. Since the frequency and latency of the measurement streams differ, the estimate is computed using fixed-lag smoothing. Fast linearization and matrix factorization methods are used to enable the filter to run at upto 40 Hz, making it more than adequate for real-time use. The pose estimator has clocked hundreds of hours of testing on the robot.

I also worked on a grid-mapping algorithm using depth measurements from stereo.

Publications

•Fast 3D Pose Estimation with Out-of-Sequence Measurements
Ananth Ranganathan, Michael Kaess, and Frank Dellaert, IROS 2007

When various sensors provide measurements at different rates, and the estimate of the output quantity is required at a rate which is less than the latency of some of the sensors, we can in effect use measurements from the “future” to provide the best estimates of the output quantity. This is performed using fixed-lag smoothing in a completely Bayesian framework, which can handle varying latencies in the sensors, so that a “delayed” measurement arrives out of chronological sequence. The resulting algorithm operates at upwards of 40 Hz and has clocked hundreds of hours of successful operation.
•Place Recognition-based Fixed-Lag Smoothing for Environments with Unreliable GPS
Roozbeh Mottaghi, Michael Kaess, Ananth Ranganathan, Richard Roberts, and Frank Dellaert, ICRA 2008

next >

< previous