Got a YouTube account?

New: enable viewer-created translations and captions on your YouTube channel!

English subtitles

← 19-02 Autonomous Vehicle Intro 2

Get Embed Code
2 Languages

Showing Revision 1 created 11/28/2012 by Amara Bot.

  1. Here is imagery of our laser system mapping out the terrain ahead.
  2. We talked a little bit about lasers and range finders in this class.
  3. Here you can see all these systems work together on building 3D maps of the environment
  4. that our car, Stanley, uses to assess the driving situation.
  5. This shows work on machine learning autonomous driving,
  6. where we used the laser to identify driveable terrain at a short range
  7. and then extrapolate this out into the long range using a machine-learning technique
  8. applied to computer vision.
  9. What you see here is a coloring, which is the output of a machine learning algorithml
  10. that identifies driveable terrain in the desert.
  11. So very briefly to tell you about the race, one with a lot of fame and $2 million.
  12. This race started early in the morning. The sun was basically still gone and was just rising.
  13. Our car was able to drive itself followed by a human-driven change vehicle and did quite well.
  14. It did so well that it actually passed the front-seated and first-running vehicle by Carnegie Mellon University.
  15. It had to navigate complicated and dangerous mountain trails where destruction lured on both sides of the car.
  16. On the left there was a cliff. On the right side there was a mountain.
  17. It is here followed by a human-driven chase vehicle.
  18. Our car very carefully ascended this route.
  19. You can see it here close before the finishing line,
  20. and after just about 7 hours it managed to do what no robot had every done before.
  21. It managed to really finish DARPA Grand Challenge, do this race, and won Stanford $2 million.
  22. We were insanely proud on this day.
  23. From this we moved onto build Junior, which competed in the DARPA Urban Challenge.
  24. Here you can see Junior's laser pursuing obstacles and being able to detect those,
  25. using basically range vision.
  26. We will talk today of localization.
  27. Junior was able to localize itself using particle filters
  28. relative to a given map of the environment, which is essential for navigating safely in traffic.
  29. It was able to detect other cars using particle filters
  30. and estimate not just where they are and how far they are moving but also what size they are, how big they are.
  31. You can see on the left the detected cars.
  32. On the right side, you see our camera view of the same situation.
  33. Here again, you can see it detect cars.
  34. Here is how it looked like from an external observation point.
  35. You can see Junior, our vehicle, driving in a fairly busy city street with lots of cars passing.
  36. It has to wait for a gap to take a left turn.
  37. When the gap finally occurs, it confidently takes the turns and drives.
  38. Today in today's class I teach you how to basically program a car just like that.
  39. So this is footage from our Google self-driving car, which you might have heard about.
  40. This car was able to drive at speeds as high as a Prius can go.
  41. It drives seamlessly in traffic.
  42. In fact, we drove over 100,000 miles without anybody noticing
  43. that there were self-driving cars in our experiments.
  44. This is near Stanford University on University Street in Palo Alto.
  45. You can see how the vehicle yields by itself for pedestrians.
  46. Of course, there's also a human driver on board just for safety,
  47. but this car, you can take my word for it, is really driving itself in traffic.
  48. This is image footage from the car itself as it goes onto a highway.
  49. This is sped up, I should say.
  50. Driving through a toll booth, and driving in Los Angeles.
  51. You can see a lot of palm trees here. It's a beautiful environment to drive in.
  52. Here you can see some of the inner workings,
  53. where you can see a corridor that the vehicle attempts to go.
  54. We can see obstacles being flagged using machine-learning techniques,
  55. range vision, laser radar, and so on.
  56. You can see it is colored by its relation to our car and its nature,
  57. and you can see it drives fairly confidently.
  58. This is an attempt to drive down Lombard Street in San Francisco--the famous crooked street.
  59. It's very curvy, and while this is sped up it gives you a sense of the complexity
  60. that is involved in building cars like these.
  61. It's actually quiet amazing how far technology has come in such a short amount of time.
  62. Here is an experiment that my Stanford students did on south parking using machine learning,
  63. reinforcement learning for control,
  64. and you can see how agile and how capable these methods are.
  65. So today I really want to enable you to write software like this based on lots of what we learned before.
  66. We talked a little bit about machine learning, a lot about particle filters,
  67. and some about motion planning, which relates to the planning class
  68. that Peter taught you quite a while back.