Robotics Sebastian Thrun AI Course
<p>Dec 25, 2011 07:52 PM
 
Self Driving Cars...  can be safely driven...
in the world every year we LOSE over 1 million people... deaths per year/ world.
 
started in 2004 after  DARPA Grand Challenge...  Calif to Nevada... Mojave...
100 difft entries...  no team made it more than 5% of total course..
CMU made it abt 8 miles only in March 13 2004...
 
CS294  stfd racing team at Stfd...  VW Tureg...
roof of parking garage at Med Center...  within less than a week they chose to NOT park there anymore...
 
60 km/hour...  avoid shrubs...  NYT  shot a photo.. crash into a tree... get stuck in the mud...
 
he shows all the maps that Stanley id's drivable terrain....
 
see coloring... drivable terrain in the desert...
2M dollar prize...
 
Stanley passed  CMU  car....  v dangerous passes...  cliffs...
won 2M dollars...
 
then built Junior... DARPA Urban Challenge...
LIDAR.. range vision... probabilistic localization
probabilistic car tracking using particle filters...  track other cars... bounding boxes...
 
waits for a gap in traffic to make a L turn...
 
Google Self-Driving  Car... Prius as fast as it can go....
nobdoy even notices... it   halts for pedestrians.... toll booth
in PA passing palm trees...  see obstacles being flagged...
colored path (green path)
Lombard Street...
 
car does a screaming park going backward!!!!
 
*
robotics is  partially observable, continuous, stochastic, and adversarial !!!
 
where is the robot (irrespective of current velocity)  x,y and a heading direciton eg  theta.
 
dynamic state:  5...  2 velocities (forward velocites and YAW.... the bearing ; it also has a YAW rate..
how fast its turning)...
 
helicopter:  kinematic  ... x y z and yaw pitch rolll
 
dynamic state = 12 ...  each kinematic state and its velocity (eg  rate of roll or yaw or pitch)...
 
Localizaiton of a car....  probabilistic loacalization
(it cud use gps...  but that's a 5meter errr)  but we can get  10 cm  error by using localiaation techniques...
 
Monte Carlo localization...  x' = x + v* del t  *  cos theta  for x compenent of position
 
also has turning velocity omega  then  omega prime = theta +  del t * omega...
 
*
*
 
Robotics  Second Lecture:  Sebastian
 
Monte Carlo Localization... get apoint cloud with sl difft velocities  (world is noisy)
predicting the location and heading of the car...
 
particle filtering...  he uses example of a point car thatis either over a lane marker
in which case it senses  BRIGHT or  off of it  in which case it senses DARK.
 
prob (bright | on the lane marker) = 0.8.
prob (dark |  off the lane marker) = 0.9
 
weight of the answer depend on where the particles are predicted to be..
 
*
Resampling....  draw with replacement... new particles...
get a new posterior
 
look at measuremnt... compute weights  then sample then predict  and then repeat it...
 
PLANNING for robotics... what to do next...
 
easiest...  is city level...  going thru a maze...
 
goal is 14 steps away from the goal... so value of start = 86.
 
each step costs 1 dollar  (to traverse the maze)
 
Dynamic Programming...
 
he shows car that is blocked that backs up and goes around the blockage....
 
A* = A star ....world is continuous but A* is discreet....
 
cars parking in a parking lot...
 
planning using MDPs and A*
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 </p>