These are my raw notes on this video. Facts are not checked, some names could be wrong.
03:00 - Darpa - no funding -> contest - changed model from before
03:40 - 2007 GPS breadcrumbs
06:40 - a year roughly to put together a car - Stanford joined the race only in the 2005 version - timeline slide
07:50 - made a course DARPA Grand Challenge - no syllabus
08:00 - 40 people, 20 people stayed on - got in touch with Volkswagen - Touareg
08:20 - first thing they did - equipped the car with inertial guiding systems - GPS + inertial measurement unit that measures accelerations and rotational velocities => extrapolate between different GPS measurements
09:00 - low level steering control - make your front tires parallel to your reference trajectory - take image from here
- if you get off the trajectory, measure your error and steer in proportion to it
10:30 - they put lasers on the roof - image
- a laser beam goes into a rotating mirror that gets directed into the environment - the laser light goes out, it is being reflected by objects - received by the sensor - we can measure the "time of flight"
11:00 - online motion planning
12:54 - test - December 1 - California - ~off-road-desert - went further than CM went in 2004, but not much - the many obstacles pushed the car around a lot - too slow for racing
13:15 - replaced driving modules with new versions, based on the feedback in the dessert
14:03 - software architecture
14:15 - faced first big hurdle - DARPA had 195 submissions - they wanted to find 40 good ones - they had to submit a video - first time they drove the car without a person inside - 43 semi-finalists
18:00 - get out bugs - goal: hundreds of miles without intervention
18:50 - Stanley gone wild - all of a sudden - go crazy out of no apparent reason - drive maybe into a ditch, down a cliff etc - car had to berescued
19:25 - 3 different innovations we worked on
20:15 - Obstacle detection - simple logic - something vertical hit by laser - image
20:40 - Effect of pitching
19:50 - Probabilistic error model
23:00 - Had to populate the probabilistic model with human driving - discriminative learning - label flat terrain as flat an non-flat as non-flat - if the human driver drove oversomething, it was flat - they managed to basically eliminate all false positives without affecting the correct positive rate
25:25 - they wanted to drive really fast (35 mph goal), but the lasers don't cover more than a 20 meter distance - couldn't stop in time when the lasers caught an obstacle
25:40 - used a camera to find roads - how do you extract from a camera image where the road is - turns out it's not easy
26:13 - they started with the idea that maybe roads are brownish - doesn't work for paved roads
26:25 - maybe the road is the smoothest thing in the image - but the smoothest thing in the image is the sky :))
27:00 - Stanley's Adaptive Vision - extract drivable region with lasers (as far as lasers can go), and use the extracted pixels as training examples - mixture of Gaussian
27:50 - if the road changes it's color you just slow down and the lasers take care of it
29:00 - the car got too fast - they had to find a way to make it slow down when the terrain got bad - adaptive mechanisms for speed control - train with human driving - stretch in the mountains - built a controller that copied that. The speed controller considers: filtered vertical shock, terrain slope, road width
31:28 - 1000+ miles of testing
32:35 - funny videos from the Berkeley team + funny Carnegie Mellon picture
33:00 - Rainy day Mojave desert (July) - couldn't see - relied on the software to drive - robot was better than human driving
34:03 - The race took place at the Fontana speedway (national qualifications), and then in Prim in Nevada. Fontana - selected the 20 finalists - Stanford 2 cars
- gate
35:20 - the most difficult obstacle was a tunnel - emulated an underpass under a highway - no GPS coverage in a tunnel - have to drive locally - this can also kill you when you get out of the tunnel and realize you are somewhere else
37:00 - one of the contestants programmed the car to go full throttle when it lost GPS signal
~37:30 - funny videos
38:22 - 23 finalists - race on October 8th
38:40 - image with the race day timeline (5:30 - DARPA chases 12 cows off the race course)
38:58 - 6:30 - race began
41:23 - "It was a complete act of randomness that Stanley actually won. It was really a failure of Carnegie Mellon's engine that made us win, and no more and no less than that"
42:10 - Last obstacle - treacherous mountain pass - reached only by 5 vehicles - Stanley was the first car to clear it - 42:23 - image
43:50 - five teams finished - four within half an hour of each other
44:26 - description of Urban Challenge 2007
45:30 - The Big Picture
45:40 - Military perspective slide
45:50 - Social perspective - cars are deadly instruments - 42000 die every year in the US because of traffic accidents - 90% caused by human error
46:20 - Commuting - people spend 1.25 hours/day in average commuting - free that time => more productive / save money
46:56 - Aging population, people who can't drive (blind people, drunk people, children) / Increase highway throughput - most space on highways is not used
03:40 - 2007 GPS breadcrumbs
06:40 - a year roughly to put together a car - Stanford joined the race only in the 2005 version - timeline slide
07:50 - made a course DARPA Grand Challenge - no syllabus
08:00 - 40 people, 20 people stayed on - got in touch with Volkswagen - Touareg
08:20 - first thing they did - equipped the car with inertial guiding systems - GPS + inertial measurement unit that measures accelerations and rotational velocities => extrapolate between different GPS measurements
09:00 - low level steering control - make your front tires parallel to your reference trajectory - take image from here
- if you get off the trajectory, measure your error and steer in proportion to it
10:30 - they put lasers on the roof - image
- a laser beam goes into a rotating mirror that gets directed into the environment - the laser light goes out, it is being reflected by objects - received by the sensor - we can measure the "time of flight"
11:00 - online motion planning
12:54 - test - December 1 - California - ~off-road-desert - went further than CM went in 2004, but not much - the many obstacles pushed the car around a lot - too slow for racing
13:15 - replaced driving modules with new versions, based on the feedback in the dessert
14:03 - software architecture
14:15 - faced first big hurdle - DARPA had 195 submissions - they wanted to find 40 good ones - they had to submit a video - first time they drove the car without a person inside - 43 semi-finalists
18:00 - get out bugs - goal: hundreds of miles without intervention
18:50 - Stanley gone wild - all of a sudden - go crazy out of no apparent reason - drive maybe into a ditch, down a cliff etc - car had to berescued
19:25 - 3 different innovations we worked on
20:15 - Obstacle detection - simple logic - something vertical hit by laser - image
20:40 - Effect of pitching
19:50 - Probabilistic error model
23:00 - Had to populate the probabilistic model with human driving - discriminative learning - label flat terrain as flat an non-flat as non-flat - if the human driver drove oversomething, it was flat - they managed to basically eliminate all false positives without affecting the correct positive rate
25:25 - they wanted to drive really fast (35 mph goal), but the lasers don't cover more than a 20 meter distance - couldn't stop in time when the lasers caught an obstacle
25:40 - used a camera to find roads - how do you extract from a camera image where the road is - turns out it's not easy
26:13 - they started with the idea that maybe roads are brownish - doesn't work for paved roads
26:25 - maybe the road is the smoothest thing in the image - but the smoothest thing in the image is the sky :))
27:00 - Stanley's Adaptive Vision - extract drivable region with lasers (as far as lasers can go), and use the extracted pixels as training examples - mixture of Gaussian
27:50 - if the road changes it's color you just slow down and the lasers take care of it
29:00 - the car got too fast - they had to find a way to make it slow down when the terrain got bad - adaptive mechanisms for speed control - train with human driving - stretch in the mountains - built a controller that copied that. The speed controller considers: filtered vertical shock, terrain slope, road width
31:28 - 1000+ miles of testing
32:35 - funny videos from the Berkeley team + funny Carnegie Mellon picture
33:00 - Rainy day Mojave desert (July) - couldn't see - relied on the software to drive - robot was better than human driving
34:03 - The race took place at the Fontana speedway (national qualifications), and then in Prim in Nevada. Fontana - selected the 20 finalists - Stanford 2 cars
- gate
35:20 - the most difficult obstacle was a tunnel - emulated an underpass under a highway - no GPS coverage in a tunnel - have to drive locally - this can also kill you when you get out of the tunnel and realize you are somewhere else
37:00 - one of the contestants programmed the car to go full throttle when it lost GPS signal
~37:30 - funny videos
38:22 - 23 finalists - race on October 8th
38:40 - image with the race day timeline (5:30 - DARPA chases 12 cows off the race course)
38:58 - 6:30 - race began
41:23 - "It was a complete act of randomness that Stanley actually won. It was really a failure of Carnegie Mellon's engine that made us win, and no more and no less than that"
42:10 - Last obstacle - treacherous mountain pass - reached only by 5 vehicles - Stanley was the first car to clear it - 42:23 - image
43:50 - five teams finished - four within half an hour of each other
44:26 - description of Urban Challenge 2007
45:30 - The Big Picture
45:40 - Military perspective slide
45:50 - Social perspective - cars are deadly instruments - 42000 die every year in the US because of traffic accidents - 90% caused by human error
46:20 - Commuting - people spend 1.25 hours/day in average commuting - free that time => more productive / save money
46:56 - Aging population, people who can't drive (blind people, drunk people, children) / Increase highway throughput - most space on highways is not used
