Life in the Atacama
As a Masters student at the Robotics Institute I conducted research for the Life in the Atacama (LITA) project, advised by David Wettergreen. In 2003-2005 the Zoë rover autonomously surveyed microscopic life in the Atacama Desert in Chile. In 2013 Zoë returned to the Atacama equipped with a drill, spectrometers, and other sensors to detect subsurface life.
Optical Flow Odometry
When not on Earth, planetary rovers must navigate without GPS. Some rely on wheel odometry for position estimation, but are prone to errors caused by wheel slip on loose soil. Visual odometry (VO) is unaffected by slip but has its own challenges. The Mars Exploration Rovers used stereo cameras for VO, but processing a single update took up to 3 minutes, and inadvertently tracking the robot's own moving shadow would cause errors (Maimone et al. 2007).
I developed a computationally cheap, optical flow odometry algorithm that is robust to self-shadowing. Instead of a forward-facing stereo camera pair, my algorithm uses a single downward-facing camera. Features on the ground plane are tracked frame-to-frame using the pyramidal Lucas-Kanade algorithm. Their locations are transformed from image to robot coordinates by a homography. Most VO systems use RANSAC or something similar to select the largest set of features that move as a rigid pattern, then estimate robot motion using just this “inlier” set. This approach fails when a majority of features lie on the edges of the robot's own shadow. I addressed this by detecting and dynamically masking shadow edges in the image when selecting features to track. As a result, my algorithm accurately estimated robot motion when RANSAC alone failed. Robustness to self-shadowing also allows greater freedom in camera placement.
I validated my algorithm on multiple robots and terrain types. The videos below show feature tracking results with and without shadow masking for the Lunar All Terrain Utility Vehicle (LATUV) on concrete. Without masking, most features lie on the shadow edge. With masking, only a few features lie on the shadow edge and these are correctly discarded as outliers. Quantitative results are provided in our IROS 2011 paper. This algorithm was also used in another RI project on the Icebreaker robot, which uses a plow control slip on steep slopes.
Control of a Passively-steered Rover
The Zoë rover has a unique passive-steering design in which axle steer angles are controlled indirectly by varying left/right wheel speeds. This design reduces mass and power consumption by eliminating steering actuators. Unlike skid-steer designs, it also allows turning without wheel slip and loss of energy to soil work. The trade-off is the need for careful kinematic modeling in the controller.
While the old controller used a 2D model, I developed one based on 3D kinematics. On uneven terrain, Zoë's axles roll to keep wheels in contact; my 3D controller accounted for these suspension deflections when computing steer angle and wheel speed commands. Our IROS 2011 paper demonstrated improved path following accuracy in simulation and physical experiments. This was a stepping-stone to my thesis research on WMR dynamic modeling.
Prior to Zoë's return to the Atacama in 2013 I also reprogrammed the low-level motor controller to better avoid and recover from axle limit errors.
Neal Seegmiller and David Wettergreen, Optical Flow Odometry with Robustness to Self-shadowing, Proc. IEEE International Conference on Intelligent Robots and Systems (IROS), September, 2011. |pub|pdf|
Neal Seegmiller and David Wettergreen, Control of a passively steered rover using 3-D kinematics Proc. IEEE International Conference on Intelligent Robots and Systems (IROS), September, 2011. |pub|pdf|
This research was made with Government support under and awarded by DoD, Air Force Office of Scientific Research, National Defense Science and Engineering Graduate (NDSEG) Fellowship, 32 CFR 168a. NASA funded the Life in the Atacama project and development of the Zoë rover (NAG5-12890)
Copyright © 2017 Neal Seegmiller