UGCV PerceptOR
Integration (UPI)


Field Tests at Fort Carson,
Colorado, August 2007

The UPI team was in the field at Ft. Carson from August 17-30, 2007, testing in the same area as it did in 2006.   The 2007 field tests built upon the results of the 2006 tests.  In 2006, the UPI team evaluated the performance of several new components of Crusher’s autonomy system. In 2007, the team put the best-in-class technologies from the 2006 tests (now with a year’s worth of improvement) through rigorous field experiments to assess the performance of the entire autonomy system.  The UPI team also continued its investigation into the effect that prior data has on Crusher’s overall performance.

By the end of two weeks in the field, Crusher showed significant performance improvements over the previous year – and with just one vehicle instead of two (the other being in use on a different project).   The general impression from observers was that Crusher was much faster than the previous year.  Crusher also enjoyed good reliability during the 2007 tests. 

Highlights

Days of field testing	14
Number of different courses	6
Total distance traveled	114.8 kilometers/71.36 miles

Performance comparison	Average speed (meters/second)	Average number of interventions
All runs	2.25	1 per 3 km traveled
Runs with DTED 4 and 5 prior data	2.51	1 per 11 km traveled

Quality of Prior Elevation Data

Ft. Carson is the UPI program’s toughest test site.   It has a little bit of everything: grassy flatlands, rugged hills and washes, forests with mature trees, and areas of dense brush.  Its many elevation changes – steep slopes, deep gullies and ditches, and sudden drop-offs – are particularly challenging for Crusher’s planning and perception systems.   

The 2007 Ft. Carson experiments explored how the quality of prior data affected Crusher’s performance in this rugged, highly variable terrain.   Crusher performed a series of runs using DTED 3, DTED 4 and DTED 5 terrain maps as prior data.  Each DTED level contains elevation data of successively higher resolution. The vehicle did six runs with maps at each DTED level.

As the quality of terrain data improved, Crusher’s performance became significantly better.  Its average speed increased and the number of human interventions during autonomous runs dropped.  The following table summarizes the results from each set of runs.

Max Margin Planning (MMP) Based Maps

The 2007 Ft. Carson tests also saw the first use of Max Margin Planning (MMP) based maps, a machine learning-based method for processing prior terrain data.  All prior maps are run through a set of algorithms to extract terrain features from the map.  These extracted features form the basis of a cost map that the planning system uses to find the safest, most efficient routes between waypoints on a test course. 

Earlier UPI experiments made use of human-engineered maps, which applied a set of human-based rules to combine the extracted features into a cost map.  The 2007 tests replaced human-engineered maps with cost maps that were generated using MMP.   MMP uses machine learning techniques to map the extracted features to terrain costs based on ideal sample paths that are drawn between waypoints.  The MMP-based maps assign costs to terrain features according to how these paths are drawn and can learn different cost mappings based on different example paths.

Best-of-Class Configuration for Autonomy System

The 2007 Ft. Carson tests did not specifically compare different perception systems.  However, the vehicle’s perception package had many improvements from earlier tests.  Several perception systems that were first tested in experimental “science” runs in 2006 have now become part of Crusher’s standard perception package. 

Laser-based Far-Range Online Learning (FROLL)

Laser-based FROLL (Far Range Online Learning) was in use constantly during the 2007 tests at Ft. Carson.  It enables Crusher to detect positive hazards that are 20 to 60 meters away from the vehicle.  Laser-based FROLL was selected for the 2007 tests because it gave more accurate results than stereo-based FROLL during the 2006 field tests.

Like MMP, laser-based FROLL uses machine learning techniques to map terrain features to mobility costs.  It enables Crusher to learn new cost mappings “on the fly” as it is driving around.  Crusher’s near-range perception system tags range data from its laser sensors with color and near-infrared (NIR) data from its cameras.  FROLL computes features from the colorized, tagged laser range data.  During autonomous runs, it learns to assign mobility costs to these features by using cost estimates from the perception system as ground truth cost estimates.   This avoids the complex, time-consuming, and error-prone process of manually mapping features to costs.  It also enables Crusher to easily adapt to new, unfamiliar surroundings.

Far Range Ground Module

The Far Range Ground Module is another formerly experimental perception system that saw constant use during the 2007 Ft. Carson tests.  Figuring out the exact location of the ground surface capable of supporting Crusher’s weight is not an easy task.  On paved roads and sand, rock or dirt surfaces, the vehicle’s laser sensors can detect the actual ground surface.  However, in dense vegetation, these sensors can only perceive the top of the surrounding vegetation, not the actual supporting surface.  Without being able to determine where the ground surface is, Crusher would effectively be driving blind. 

The Far Range Ground Module finds the location of the underlying ground surface in both cases.  The module ray-traces data points from Crusher’s laser sensors to build a map of the ground plane, walls, extended slopes, and other obstacles.  To find the true ground plane, the module applies a probabilistic model to smooth out the laser data and estimate the location of the ground surface.  The laser data gets more smoothing in areas with vegetation and less in areas without it, reflecting the sensor’s ability to detect ground in these places.

New Autonomy System Technologies

The 2007 field tests also evaluated several new technologies that were incorporated into Crusher’s autonomy system.

Learned Vehicle Model

The Learned Vehicle Model bridges the gap between Crusher’s intended and actual movement.  The movement that Crusher actually executes almost never quite matches its intended trajectory.  For example, when Crusher turns, it tries to drive in a perfect arc.  However, it often cannot follow the arc as closely as its autonomy system would like: its wheels may slip, its physical attitude and position may not be optimal, or it may be on too steep of a slope.  The result is that Crusher deviates from the arc, forcing its autonomy system to constantly correct its path.

The Learned Vehicle Model applies machine learning techniques to improve prediction and control of Crusher’s movement.  Instead of assuming that Crusher always executes its moves perfectly, the model takes the vehicle’s physical state and surroundings into account to estimate its actual trajectory.  The vehicle’s pitch and roll, the positions of its suspension arms, the terrain where it’s located, and the commands it’s been given are fed into the model.  On the basis of earlier predicted and actual results, the model predicts how much Crusher will deviate from its desired path and determines its final location.  This enables the autonomy system to control Crusher more accurately.

Learned Odometry

GPS is not available in many places at the Ft. Carson test site, such as the bottoms of deep washes and forests with dense tree canopy. When Crusher’s positioning system cannot see the GPS satellites, it relies on data from its IMUs to estimate its position.  However, these IMU-based estimates will drift over time from Crusher’s actual position, introducing errors. 

Learned Odometry uses machine learning techniques to reduce this drift.  It feeds an estimate of the vehicle’s true velocity into the positioning system to compensate for drift-related positioning errors.  A machine learning algorithm uses Crusher’s command history and a number of physical factors (including the vehicle’s current posture, the angles of its suspension arms, and the wheel velocity of its six wheels) to generate this estimate of its true overall velocity.  This is not a trivial problem.  Because Crusher is skid steered, each of its six wheels has a different wheel velocity and slips at a different rate. 

Range Image Based Terrain Classification

Crusher’s near-range perception system detects both ground hazards and above-ground hazards up to 20 meters away from the vehicle. It classifies terrain into three categories: roads, vegetation and obstacles.  Previous tests used a voxel-based representation to classify terrain.  The 2007 field tests used range image based classification.  It identifies roads, vegetation and obstacles by using laser data that shows the local shape of the terrain (for example, whether it is flat, smooth, spherical, etc.) and information from Crusher’s color and NDVI cameras. 

Urban Terrain Mapping

Data from Crusher’s laser sensors is also used for urban terrain mapping.  As it drives through an urban area, Crusher’s sensors build a three-dimensional map of the surrounding environment.  This map shows the positions of structures, vehicles, and other features of the urban landscape.  Three-dimensional urban terrain mapping is useful in Military Operations on Urban Terrain (MOUT) and similar applications – for instance, to update aerial or satellite data with the current locations of buildings, walls, and so forth.

Additional Pictures