Autonomy for Ground Vehicles
Status Report, July 2006
Sanjiv SinghAssociate Research ProfessorField Robotics CenterCarnegie Mellon University
Automation of All Terrain VehiclesMain issues: Path tracking Obstacle Detection at 4-6
m/s Reliable operation Low-cost system Intuitive interface for
Teaching Teleoperation For conducting surveillance
Vehicles
2003 2004
Accomplishments (2003-2005) Retrofitted 3 vehicles for autonomous operation Helped design and Implemented general positioning solution Designed and Implemented optimized design of laser scanner Designed & implemented optimized electronics/power/computing
package to support autonomy Implemented path tracking at 6 m/s Implemented method to detect “stuck condition” Designed and Implemented collision avoidance merging data from
two lasers scanners at 4 m/s Designed and Implemented joint method of collision avoidance and
path tracking. Designed Implemented PDA control of vehicle and feedback of
essential variables. Designed and implemented method to plan smooth paths from via
points specified on base station.
Issues for 2006 Perception
Calibration New interface for sweeping laser Vegetation detection
Guidance Varied Traversability Learned parameters Improvement for tight spaces like tunnels
Experimentation New vehicle testbed
Calibration New calibration method for calibrating relative pose of two
laser scanners Method was improved in Feb 2006.
New Interface to Sweep Laser
Current version
New Interface to Sweep Laser
New version
New Interface to Sweep Laser
Board replaces Sync box.
Vegetation Detection In progress
Example data shown to system
Classification into three components
Vegetation Detection In progress
Example data shown to system
Classification into three components
Vegetation Detection In progress
Example data shown to system
Classification into three components
Varied Traversability Tested in simulation
Vehicle avoids vegetation when possible Vehicle drives over vegetation if necessary
Learned Parameters Have added damping to decrease oscillations. Values are
obtained from observation of human driving Needs to be tested
Vehicle oscillates when near many large obstacles New method (with damping) decreases oscillations
Improved performance for Dodger
For tight spaces with possible GPS shift
QuickTime™ and aCinepak decompressor
are needed to see this picture.
Comparison to other Algorithms
Compared with other methods
Experimentation Testing on real vehicle is very important. Golf cart is not suitable for high speed experimentation Will give us ability to try different configurations Status: ready to test end of July, 2006
Potential Agenda Done
Implementation of design with lower cost components Smaller obstacle detection Smoother steering
In progress Navigation in cluttered environments Immunity to small and medium vegetation Immunity to GPS dropout
No plan yet Road Following Detection of moving obstacles Immunity to large vegetation Water/negative obstacle detection All weather operation
DARPA Urban Challenge 2007 Fully Autonomous operation in Urban Environments Competition to complete mission of 100 km in less than 6 hours Other moving vehicles will be present (no pedestrians) Vehicle must
Plan route automatically/ Replan routes when road is blocked Follow roads avoid collisions with other vehicles Stop accurately at stop signs Pass stopped vehicles Go through intersections Park in parking spot
CMU will enter one vehicle in competition
Learned Parameters
New test vehicle has different steering characteristics than Grizzly
Learned parameters with damping for new vehicle These are used in following vehicle tests
Preventing Off-Path States Add rows of
obstacle points to keep vehicle closer to path
Vehicle is allowed to cross these points
Border obstacles with cost 0.8
Border obstacles with cost 0.3
Border obstacles with cost 0.1
Detected obstacles with cost 1.0
Planning Combined with Dodger
Previously: showed how planning helps prevent stuck situations Planning only invoked when stuck state is predicted
Planning Combined with Dodger
Previously: showed how planning helps prevent stuck situations Planning only invoked when stuck state is predicted
Planning every iteration improves this further Helps deal with desired path offsets (GPS
jumps/outages) Better at finding tunnel openings Shifts goal point to center of tunnels
Planning Combined with Dodger
Planning provides good direction for vehicle to go Dodger provides:
smooth control resistance to infeasible motions in the plan another level of safety in obstacle avoidance
Planning Combined with Dodger
Planned path is jagged, not achievable by vehicle
Dodger uses plan, but drives smoothly and gives more space to
obstacles
Planning Combined with Dodger
Show movies here: Combination of movies taken at LTV on Friday
First three scenarios: MVI_0661 Slalom (fourth): MVI_0667 Wide and big slalom (fifth and sixth): MVI_0668
Bonus footage: 5 m/s complete loop: MVI_0669 6 m/s ¾ of the loop: MVI_0670 Views from the vehicle: MOV05982, MOV05989
dataReplayLoop.avi
Tunnels
Desired path has jumped to side of tunnel.Planner shifts goal point away from the wall.
Tunnels
Regular operation: avoid obstacles that are meters away
In tunnels, walls are less than two meters away Different set of parameters to track paths in a
tunnel Difficulty: When to switch parameter sets
Show movies here: Tunnel movies from LTV: MVI_0678, MVI_0679 dataReplayTunnel.avi
Terrain Classification
System should classify laser data based on: Height How much it looks like a 2-d surface How much it looks like a 3-d cluster
Show system hand-labeled data System learns classifier using the three
attributes
Terrain Classification Results
Tested on data taken at LTV White: confident it’s ground
Blue/purple:
Classified as ground, but with low confidenceGreen: classified
as not ground, confident it’s traversable
Red: confident
it’s not traversable
Terrain Classification Results
GroundFence and bushes
Grass Lowgrass
Terrain Classification Results
Road
Cars
Grass
Terrain Classification Results
Garage wall
Terrain Classification Results
Low grass
BushGrass
Terrain Classification Results
Road
Grass
GrassGrass
Grass
Terrain Classification Results
Small boxes (should be labeled ‘obstacle’) sometimes classified incorrectly
Box classified as traversable
Terrain Classification Results
Small boxes (should be labeled ‘obstacle’) sometimes classified incorrectly
Need more training data with small obstacles