6 ms·

billions of miles of real-world driving data from a million and growing fleet of vehicles. Tesla has built their own chip optimized fully for self-driving. al

billions of miles of real-world driving data from a million and growing fleet of vehicles. Tesla has built their own chip optimized fully for self-driving. all-in on vision based full self-driving, which seems to be the only way to solve all the edge cases. Even if google accomplishes self driving with lidar, lidar will never be able to compensate for all edge cases. Human vision already compensates for all edge cases, thus the case for vision based self driving.

jcranmer 6y ago
It's been such a long time since I've heard someone trot out the "Tesla has the best raw data" canard that I thought it had finally died off for real. For starters, Tesla still seems to have this annoying habit of not detecting large, white trucks that are stopped in the middle of the highway, and will happily drive into them at full speed. This isn't exactly an "edge case" issue here.
- OldHand2018 6y ago
  > For starters, Tesla still seems to have this annoying habit of not detecting large, white trucks that are stopped in the middle of the highway, and will happily drive into them at full speed. This isn't exactly an "edge case" issue here. <joke>To be fair, it was an edge case until Tesla started selling cars.</joke>
- sixQuarks 6y ago
  I didn't say best raw data. Lidar obviously produces better raw data on an individual basis. Tesla has more real-world driving data overall. Nevertheless, it doesn't matter. Tesla is the only one going all-in on vision based self driving. That's the key thing that's going to get us to full self driving. Everyone will eventually need to do this.
  - jcranmer 6y ago
    > I didn't say best raw data. Lidar obviously produces better raw data on an individual basis. Tesla has more real-world driving data overall. The "real-world driving data" is what I'm calling the "raw data" here, as it's the raw data from the standpoint of the (perceived) AI algorithms driving self-driving cars. > Tesla is the only one going all-in on vision based self driving. That's the key thing that's going to get us to full self driving. Why is it the key thing? Why should worsening the quality of your input data improve the results?
    - sixQuarks 6y ago
      Because the important part is the software that interprets that data. Our eyes get worse data than lidar, yet our brains are able to process this data more effectively. Also, lidar is super expensive. It will take much longer to get affordable lidar into everyday vehicles.
      - jcranmer 6y ago
        The hardware necessary to emulate our brain capabilities with neural nets is several orders of magnitude more expensive than lidar. Running with a more responsible hardware stack means the software has to be more cleverly designed, and that design is mostly not in the just-throw-data-at-a-neural-net stage. Building your visible-light-only driving system is going to cause you to struggle a lot harder at getting a good subroutine for "how far away is that object?" while providing you no benefit to solving the harder problem of "what will that object react to my coming nearer to it?"
        sixQuarks 6y ago
        I was using an analogy with the brain. You don't have to replicate a brain for full self driving.
bhupy 6y ago
Waymo also appears to have real-world driving data from their fleet of cars, and also appears to have an operational ridesharing pilot with no driver in the front-seat. Another way of asking this question: how many "millions" of miles is sufficient, and at what point are there diminishing returns?
- nickik 6y ago
  They have a partially driverless service in one small city where conditions are mostly perfect and traffic is light. The problem is slogging threw millions and millions of corner cases in the real world. A trillion miles on the highway are quite pointless, a trillion miles in one Arizona city will also not get you there. You need millions of miles in all driving conditions and train the car to always figure out 100s of details. Look at the Tesla presentation where just recognizing a 'Stop' is incredibly difficult. Working threw all these problems requires having 10000s of cars with real sensors sourcing data from you. The advantage Tesla has is that they can do this with zero capital, while Waymo needs to have a expensive cars (100 to 1000s) with a payed driver, go threw every location and create a HD maps.

jcranmer 6y ago

It's been such a long time since I've heard someone trot out the "Tesla has the best raw data" canard that I thought it had finally died off for real. For starters, Tesla still seems to have this annoying habit of not detecting large, white trucks that are stopped in the middle of the highway, and will happily drive into them at full speed. This isn't exactly an "edge case" issue here.

OldHand2018 6y ago
> For starters, Tesla still seems to have this annoying habit of not detecting large, white trucks that are stopped in the middle of the highway, and will happily drive into them at full speed. This isn't exactly an "edge case" issue here. <joke>To be fair, it was an edge case until Tesla started selling cars.</joke>
sixQuarks 6y ago
I didn't say best raw data. Lidar obviously produces better raw data on an individual basis. Tesla has more real-world driving data overall. Nevertheless, it doesn't matter. Tesla is the only one going all-in on vision based self driving. That's the key thing that's going to get us to full self driving. Everyone will eventually need to do this.
- jcranmer 6y ago
  > I didn't say best raw data. Lidar obviously produces better raw data on an individual basis. Tesla has more real-world driving data overall. The "real-world driving data" is what I'm calling the "raw data" here, as it's the raw data from the standpoint of the (perceived) AI algorithms driving self-driving cars. > Tesla is the only one going all-in on vision based self driving. That's the key thing that's going to get us to full self driving. Why is it the key thing? Why should worsening the quality of your input data improve the results?
  - sixQuarks 6y ago
    Because the important part is the software that interprets that data. Our eyes get worse data than lidar, yet our brains are able to process this data more effectively. Also, lidar is super expensive. It will take much longer to get affordable lidar into everyday vehicles.
    - jcranmer 6y ago
      The hardware necessary to emulate our brain capabilities with neural nets is several orders of magnitude more expensive than lidar. Running with a more responsible hardware stack means the software has to be more cleverly designed, and that design is mostly not in the just-throw-data-at-a-neural-net stage. Building your visible-light-only driving system is going to cause you to struggle a lot harder at getting a good subroutine for "how far away is that object?" while providing you no benefit to solving the harder problem of "what will that object react to my coming nearer to it?"
      - sixQuarks 6y ago
        I was using an analogy with the brain. You don't have to replicate a brain for full self driving.

bhupy 6y ago

Waymo also appears to have real-world driving data from their fleet of cars, and also appears to have an operational ridesharing pilot with no driver in the front-seat. Another way of asking this question: how many "millions" of miles is sufficient, and at what point are there diminishing returns?

nickik 6y ago
They have a partially driverless service in one small city where conditions are mostly perfect and traffic is light. The problem is slogging threw millions and millions of corner cases in the real world. A trillion miles on the highway are quite pointless, a trillion miles in one Arizona city will also not get you there. You need millions of miles in all driving conditions and train the car to always figure out 100s of details. Look at the Tesla presentation where just recognizing a 'Stop' is incredibly difficult. Working threw all these problems requires having 10000s of cars with real sensors sourcing data from you. The advantage Tesla has is that they can do this with zero capital, while Waymo needs to have a expensive cars (100 to 1000s) with a payed driver, go threw every location and create a HD maps.