Autonomous Driving and Intelligent Vehicles at Volkswagen Dirk Langer, Ph.D.
VW Autonomous Driving Story 2000 2003 2006 Robot Klaus Purpose: Replace test drivers on poor test tracks (job safety) Robot Klaus-Dieter Purpose: Reproducible testing of stability limit and vehicle dynamics applications GTI 53 +1 Purpose: Reproducible testing at stability limit SAFETY AND TESTING COMFORT AND DRIVER ASSISTANCE 2005 2007 2008 Stanley Purpose: Off road of autonomous driving Junior Purpose: Advanced testing of driver assistant systems Paul Purpose: Parking without driver in tight parking spots
VW Autonomous Driving Story AUTONOMOUS PASSENGER VEHICLE VaMoRs VaMP / VITA 2 ALV PARTIALLY AUTONOMOUS VEHICLE APPLICATION NavLab UGV ESP KNOWLEDGE ALLOCATION AUTONOMOUS VEHICLE FOR TESTING PURPOSE Source: Mueller Bessler, Bernhard ; Stock, Gregor ; Hoffmann, Juergen: Customer oriented safety and handling evaluation via adjusted driver model using real vehicle. Munich: FISITA, 2008 AHS KNOWLEDGE PREDEVELOPMENT Klaus ACC ABD Steering Robot Stanley Grand Challenge 2004 MuCar3 Junior Caroline Grand Challenge 2005 Urban Challenge 2007 European Land Robot Trial 2007 Gespannstabilisierung Active Steering Klaus Dieter ACC follow to stop LKA PLA DSR PSB Volkswagen Workshop Driving Dynamics 2006 GTI 53+1 Track Trainer Porsche Steering Robot TIME BAR 1990 1995 2000 2005 research
Robot Klaus-Dieter REPRODUCIBLE CLOSED LOOP TESTING
Robot Klaus-Dieter
Robot Klaus-Dieter Goal: Reproducible closed loop lateral vehicle dynamics evaluation Arbitrary test vehicles equipped with steering control (approx. 50 cars tested up to now) Vehicle follows a given trajectory (e.g. double lane change as standardized closed loop maneuver) through a set of cones using DGPS Method ensures reproducible steering behaviour. Vehicle speed is controlled by the driver. In case of double lane change driver removes foot from throttle pedal upon entering the cones in order to eliminate longitudinal dynamic effects Publication: Mueller-Bessler, Bernhard ; Henze, Roman ; Küçükay, Ferit: Reproducible transverse dynamics vehicle evaluation in the double lange change. In: ATZ04/2008, pp. 358-365 365
GTI 53+1 REPRODUCIBLE CLOSED LOOP TESTING
GTI 53+1 Goal: Reproducible closed loop vehicle dynamics behaviour evaluation Test Vehicle: VW GTI equipped with steering, brake and throttle control A driving course is defined by pairs of cones Process steps Using DGPS and a laser scanner, the vehicle is slowly autonomously driven through the cones, recording cone pair locations. The recorded data is processed off-line, computing an ideal driving trajectory in between the cone pairs. The GTI is then driven fully autonomously at the highest speed possible on the pre-computed p ideal trajectory to traverse the course in minimum time. Publication Mueller-Bessler, Bernhard ; Stock, Gregor ; Hoffmann, Juergen: Reproduzierbares Fahren im Grenzbereich. Munich: race.tech, 2006
GTI 53+1
Stanley Grand Challenge Junior Urban Challenge
Grand Challenge 2005 Autonomous Off-Road Driving in 2005 Grand Challenge Finished in first place among 23 contenders Autonomous Urban Driving in 2007 Urban Challenge Finished in second place among 89 teams, 11 finalists and one of only 6 teams to complete the race. Next Steps: Extract relevant technologies for future driver assistance systems
Inside Stanley GPS GPS compass 6 Computers E-stop 5 Lasers Camera Radar Drive by wire IMU Screen Popular Mechanics
2005: Primm, NV
The Laser Teaches the camera Which pixels outside the laser brick look similar to the pixels inside the brick?
Robot Perception: Stanley Obstacle detection with laser rangefinders Computer vision-based road detection
What did we learn from Stanley? Incorporate GPS, but drive with your sensors A simple world model will get you a surprisingly long way Off-road dtrail ildi driving, i not unstructured off-road driving Driving is a software problem!
Websites: Stanford Racing Team: http://www.stanfordracing.org PBS NOVA Special Great Robot Race http://www.pbs.org/wgbh/nova/darpa/
2005 Final Race Standings
Challenge: Motion Planning
Challenges of Urban Driving Moving obstacles Hazards 360 degrees around dthe vehicle Faster closing speeds Behavior of other drivers must be tracked and predicted In order for robots to make intelligent t driving i decisions i in urban environments, they must understand their environment, not just perceive it
2007 DARPA Urban Challenge 60 mile urban race Moving traffic Obey California traffic laws Intersections, merging, passing, traffic circles, parking, u-turns
Typical Urban Driving Scene
Motion Planning Driving constraints help limit the size of the planning search space
Motion Planning Driving constraints help us predict the behavior of other vehicles
Grand Challenge 2005 Stanley
NQE Area B, Parking
NQE Area B, Passing
Lane Change (30mph)
Final Results 1 st : CMU s Boss 4h 10min 20sec 2 nd : Stanford s Junior 4h 29min 28sec 3 rd : Virginia Tech s Odin 4h 36min 38sec 4 th?: MIT s Talos?: Cornell s Skynet?: Penn s Little Ben
Paul
Paul Goals Parking in tight perpendicular parking spaces. In order to successfully navigate into perpendicular parking places, a precise measurement of the space is essential. Monocular cameras are used here along with modern image processing techniques.
2. Motivation Paul Volkswagen offers a semi-automatic parking assistance system (Park Assist) for parallel parking. Perpendicular parking is a much greater challenge for the sensors. The relationship between car size and parking space size is growing to be a problem. Especially in the case of perpendicular parking spaces one increasingly has difficulty accessing the automobile.
Paul Technical Solution Car Configuration Sensors Parking without a driver - Automatic perpendicular parking Video cameras are the main sensors. Ultrasound sensors for scanning for obstructions Camera Wheel speed sensors Ultrasound sensors for scanning for obstructions - The parking space is measured by two cameras mounted in the mirrors - Series ultrasound sensors scan for unexpected obstructions Camera Wheel speed sensors Remote activation key for initiating automatic parking Actuators electromechanical power steering brake booster ECU for the parking strategy computer for video stream processing
3. Technical Solution Image Processing Paul In order to detect a parking space, the video stream is processed using a method called Structure-From- Motion (for explanation see figure) The figure shows the camera perspective from two video frames. Assuming the surroundings are static, these two frames can be used to produce stereo depth information.
THE END Spin offs Park Assist
THE END Future Source: YouTube-Film irobot -Scene