Stanley : The robot that won grand challengePresentation Transcript
STANLEYBilly OkalMachine Learning Seminar 2012
Outline• Task overview• Robot and Team description• Robot performance/results• Challenges/critique
Task (in brief)• Drive across a desert autonomously for ~140 miles• Course (RDDF) given only 2 hours before race• US $2 Million in prize money (for 2nd run in 2005)• DARPA already mentioned that a 4WD pickup is physically capable of ﬁnishing the race
Stanley Team • Stanford University (headed by Thrun) • Volkswagen ERL • Mohr Davidow Research • Intel Research • OthersImages: Sebastian Thrun’s presentation
Basic Vehicle Specs• Volkswagen Touareg R5, 4WD• Variable height air suspension, auto and electronics locking differentials• Added skid plates, front bumper• Custom interface to throttle and brakes• DC Motor Steering, Gears (drive, reverse, packing)• CAN bus for communication• Custom roof-rack for carrying sensors
Stanley (Nov, 04)Images: Sebastian Thrun’s presentation
Sensors• Five SICK Laser range ﬁnders different angles)• Two 24 Hz Radar sensors (cover up to 200m)• Color Camera for long range perception• GPS positioning unit and compass• IMU (Inertial measurement unit)
Computational Resources• Six Pentium M computers (only 4 used in actual race)• Gigabit Ethernet switch• Custom power with backup batteries• Special air ducts from AC for cooling• Total power requirement of 500 W• Linux OS
Software ArchitectureImages: Sebastian Thrun’s presentation
Terrain classiﬁcation Planning Vision LearningImages: Sebastian Thrun’s presentation
Key concepts• Modular architecture• Probabilistic Error Models• Data driven parameter learning ( e.g for terrain classiﬁcation)• UKF Position estimation• ...
Beyond the Race• Stanley now rests at a Museum• Junior competed in 2007 Urban challenge (didn’t win :( )• Sebastian now leads Google’s driverless cars campaign• Modern “versions” of Stanley now are smaller, less expensive.
Challenges• Stanley cannot pass a moving big object• Stanley had ‘no global planner’. Solved in successors• Only off-road use• Only static race environments• Critical time issues (500 msec delay caused problems)
References• S. Thrun et. al. ; ‘Stanley: The Robot that won the DARPA Grand Challenge’, Journal of Field Robotics 23(9), pp 661-692• S. Thrun, M. Montemerlo and A. Aron; ‘Probabilistic Terrain Analysis for high speed desert driving’, Robotics: Science and Systems (RSS) 2006• Stanley web page: http:// robots.stanford.edu/talks/stanley/