Autonomous Offroad Navigation Under Poor GPS Conditions

1
Autonomous Offroad Navigation Under Poor GPS
Conditions

• Thorsten Luettel, Michael Himmelsbach, Felix von
  Hundelshausen, Michael Manz, André Mueller,
  Hans-Joachim Joe Wuensche
• Institute for Autonomous Systems Technology (TAS)
• Department of Aerospace Engineering
• University of the Bundeswehr Munich

2
Outline

• Introduction
• Demonstration Platform MuCAR-3
• Sensor Data
• Driving With Tentacles
• Autonomous Navigation Scenario of the 2009
  C-ELROB
• Impressions from C-ELROB
• Results

3
Introduction

• Autonomous Navigation Scenario at Civilian
  European Land Robot Trial, June 2009 in
  Oulu/Finland
• track was defined through sparse GPS waypoints
• ? insufficient shape of track if using point to
  point connections
• hard access path, leading through the forest,
  some very narrow passages at the end (footpath
  through the Botanic Gardens)
• bad GPS conditions? never trust GPS

4
Demonstration Platform MuCAR-3

• Munich Cognitive Autonomous Robot Car 3rd
  Generation
• VW Touareg with full drive-by-wire modifications

• Inertial Navigation System
• INS IMU (D)GPS
• For inertial compensation of sensor data,
  egomotion estimation and global localization

• Camera Platform MarVEye 8
• Multi focal, active / reactive Vehicle Eye, 8th
  generation
• Fast sakkades in yaw direction
• Inertial pitch stabilisation
• Three HDRC CMOS RGB cameras

• Velodyne High Definition LIDAR
• 360 horizontal, 26 vertical field-of-view (FOV)
• 64 beams rotating at 10 Hz, 1 Mio 3D points /
  second

5
LIDAR Data Occupancy Grid Mapping

• 2.5-D ego-centered occupancy grid of dimension
  200m x 200m
• Each cell (0.15m x 0.15m) stores a single value
  expressing the degree of how occupied that cell
  is
• Computed from inertially corrected LIDAR scan
• Accumulating obstacles from multiple scans
  results in a probabilistic occupancy value for
  each cell

cell size exaggerated
6
Driving With Tentacles

• Approach to robot navigation was driven by the
  search for simplicity
• What is the simplest approach that lets a robot
  safely drive in an unknown environment?
• MuCAR-3 moves within an unknown environment
  similarly to how a beetle would crawl around and
  uses its feelers to avoid obstacles
• Feelers converted to target trajectories
  representing the basic driving options of MuCAR-3
• We named these target trajectories tentacles

Tentacles A fixed number of precalculated motion
and sensing primitives
Felix von Hundelshausen, Michael Himmelsbach,
Falk Hecker, Andre Müller, and Hans-Joachim
Wünsche.Driving with Tentacles Integral
Structures for Sensing and Motion. In
International Journal of Field Robotics Research,
2008.
7
The Details of a Tentacle

• Classify tentacles into the two classes
• Drivable no obstacles within velocity-dependent
  safety distance
• Not drivable
• Normalized weighting factors
• distance to the first obstacle
• averaged weighted sum of the occupancy-grid cells
  under the respective support area
• distance and angle to target track
• visual drivability analysis (camera image)
• From all drivable tentacles Select the tentacle
  with the minimum decision value

8
Visual Evaluation of Tentacles

• For visible tentacles the 3D classification area
  boundaries are sampled and projected into the
  camera frame
• Evaluate color saturation as a clue to drivable
  areas in wooded environments
• Sum all weighted saturation pixel values located
  within the support area using line-oriented
  integral images ? lower sum better

9
Autonomous Navigation Scenario

• Now to the Autonomous Navigation Scenario of the
  2009 C-ELROB
• 60 min time limit
• two laps to go, 5.2 km (3.2 miles) in total
• 17 UTM waypoints given (per lap)
• Point to point connections give only an
  inaccurate description of the track
• some waypoints seem to be in the forest

10
Improved Target Track Generation

• commercially available data from a GIS (Geo
  Information System)
• GIS provides a lot of additional information, we
  only use the road network in this approach
• 1st we match given UTM waypoints to GIS road
  network
• 2nd we add intersection points from GIS as
  additional waypoints
• Finally we replace the point to point connections
  with the more detailed GIS shape? improved
  target track

11
Bad GPS

• forest environments cause many GPS signal outages
  and reflections
• GPS position errors reach 50m
• Therefore we use an INS and enhanced egomotion
  estimation for
• inertial correction of sensor data, and
• localization
• egomotion estimation
• utilizes INS measurements, steering angle,
  odometry, air suspension levels,
• provides pose in fixed global coordinate system
  and also in drift space coordinate system

12
Reduced GPS Weight for Navigation

• Due to
• GPS position offsets, and
• Typical differences between GIS target track and
  real world
• strict GPS path following is not advisable
• ? Very low weight on target track in our tentacle
  approach during normal driving
• However, at some places GPS or track information
  is needed
• Direction to take at crossings
• Where to go on large free areas
• ? Consider GPS as a factor in the tentacle
  approach only within a given radius around few
  waypoints and crossing points

13
Challenging Parts from C-ELROB
14
Challenging Parts from C-ELROB
15
Challenging Parts from C-ELROB
16
Fast driving
17
MuCAR-3 frequently stops
18
MuCAR-3 frequently stops
19
Results

• MuCAR-3 was the only vehicle to reach the finish
  line,
• while demonstrating a high level of autonomy
• 95 of distance
• 81 of time(incl. post-processing after the
  race)
• critical parts
• paths narrower than the vehicle? high grass was
  perceived as an obstacle
• turn-offs into the green were challenging

20

• Thank you for your attention!
• Any Questions?