UGCV PerceptOR
Integration (UPI)


Field Tests at Fort Carson,
Colorado, August 2006

Over hills, under trees, and through washes and sagebrush! The UPI team spent two grueling weeks in August at Fort Carson, Colorado testing Crusher’s autonomy and perception systems. Goals included faster vehicle speed, better obstacle and vegetation detection, lower latencies, and improved far-range perception. Here’s a snapshot of the results.

Highlights

Test duration:	11 days
Number of different courses:	10
Total miles (in both Crusher vehicles):	134.6 miles (216km)
Total distance traveled by Autonomy System (both vehicles):	60.5 miles (97.5km)
Typical daily average speeds (including RC operations):	1.7 - 3.3 m/s
Typical daily autonomous average speeds:	2.1 – 2.7 m/s

Overall, this round of autonomy testing met program goals and identified areas for improvement. The next incarnation of Crusher’s autonomy system will be tested at Fort Bliss, Texas from January 22 to February 2.

Test Results

•

Improvements in near-range perception (up to 10-15 meters from the vehicle) reduced the perception system's latency. Reducing latency is crucial for supporting higher vehicle speeds. Drifting vehicle pose sometimes made it difficult to estimate where the ground was located and caused the vehicle to believe it was surrounded by obstacles. Using infrared cameras to detect vegetation showed improvement over previous tests, although sloped ground still presented a challenge.

•

Far-range online learning (FROLL) was tested with the vehicle’s far-range perception system (up to 60-80 meters from the vehicle). FROLL uses either stereo cameras or laser range finders to locate and learn to identify distant terrain features. Laser FROLL identified obstacles at greater distances and increased the range at which roads could be detected by 20 meters. Stereo FROLL also detected obstacles at long range, but needed assistance to determine costs for vegetation, roads, and areas around trees.

•

Because good prior overhead terrain data isn’t always available, it’s important to find out how Crusher performs using data from other sources. A series of tests over different courses compared Crusher’s performance using combinations of low-resolution and high-resolution prior data, near-range perception, FROLL, and map online learning (MOLL). MOLL learns how to interpret prior data on the fly by observing the perception system’s classifications and by applying them to a larger area of the map.

°

With high-resolution prior data, Crusher traveled at faster average speeds, planned smoother and safer paths that avoided extreme terrain, had fewer deviations from its originally-planned paths, and needed fewer operator interventions.

°

With low-resolution prior data, Crusher performed similarly in many respects but deviated into extreme terrain (such as steep slopes where its roll limit was exceeded) because its elevation data was not as good. It depended more on its sensors for local obstacle avoidance and required more operator interventions.

°

When Crusher operated only with near-range perception and no prior data, path planning was more of a challenge. It hunted its way through courses and ventured into cul-de-sacs that sometimes required operator intervention to exit.

°

Laser FROLL gave Crusher a better vantage point than near-range perception and enabled it to plan smoother paths and avoid some cul-de-sacs. Stereo FROLL had better range than laser FROLL, but wasn’t as accurate. Overall, planning based on FROLL lacked the overarching perspective of planning with prior overhead data of any resolution.

°

Using MOLL, Crusher learned how to classify roads, grass and trees and was able to avoid obstacles. It planned paths that took advantage of trails and other easily-navigable areas.

•

Cameras and ladar sensors cannot always directly sense drop-offs, ditches, holes and other negative hazards. Instead, the perception system reasons about occlusions to estimate (or infer) their slope and depth. By analyzing sensor data and Crusher's behavior when it encountered negative hazards, the UPI team learned how to make better use of inferred data.

More photos (Fort Carson field test photos are identified on the ensuing page.)

Video