Visual odometry

What it is

Visual odometry is a way for a robot to work out how far it has really travelled by watching the ground go past, rather than trusting its wheels. NASA's Jet Propulsion Laboratory describes it as the autonomous detection and tracking of features in the nearby terrain, used to measure how much a rover has slipped while driving on soft or highly sloped ground ^[1]. The distinction matters because a wheel that turns is not necessarily a wheel that carries the vehicle forward: on loose soil or a steep slope the wheels can spin or slide, and a simple count of wheel rotations then overstates the distance covered.

How it works

The rover's cameras take a picture, the rover drives a short distance, and the cameras take another. Software compares the two images, automatically finding dozens of features — corners of rocks, high-contrast corners in the rover's own tracks, sharp changes in soil texture — and tracking how each one moved between the pictures ^[2]. Every tracked feature gives a small clue about how the camera moved; combining the measured three-dimensional motion of all of them lets the rover figure out how much it really moved and turned, much more precisely than simply counting wheel turns ^[2]. The estimate comes from the images rather than from the wheels, so it stays meaningful exactly when the wheels are slipping.

Where it came from

Visual odometry was carried on the Mars Exploration Rovers, and JPL credits those rovers with being the first to demonstrate it ^[1]. It was not part of the baseline software design for the mission. A JPL paper on autonomous navigation results from MER records that visual odometry for position estimation was integrated as an extra credit item, and that it grew in importance as surface operations progressed ^[3]. In areas of high slip, the stereo vision-based version of the technique was used to estimate rover motion ^[3].

What it enabled

The practical payoff was driving on slopes. With visual odometry compensating autonomously for unexpected slip, both rovers were able to drive on slopes from 15 to 30 degrees ^[1]. On Opportunity, the visual odometry software also enabled precision approaches to targets even on slopes with unpredictable slip ^[1], so the rover could pull up close to a rock or a patch of soil that scientists wanted to examine rather than stopping short of it.

Limits

The method depends on there being something to look at. It works by finding dozens of trackable features such as rock corners, track corners and sharp soil texture changes ^[2]; ground that offers few distinct, high-contrast features gives the software less to work with. And measuring slip is not the same as preventing it: visual odometry tells the rover how far it actually moved, which lets it correct its own estimate, but the wheels are still slipping and the slope is still there. What the technique buys is a more accurate picture of where the rover is, which is what makes a drive on uncertain ground a manageable risk rather than a guess.