5 ms·

With all respect I think "easy-to-access" is a bit of an understatement when you are comparing a 3D game to real-world sensor data. They are barely comparable.

by drcross 10y ago

wrsh07 10y ago
The fact that you can even consider training an NN for driving after under an hour of setup is absurd. Did you see the "self driving mario kart"? [1] That's a lot of complexity just to try out some algorithms for driving. Is GTA V the same as setting up a rig which controls I/O for a real car? No. Of course not. You're not being fed data from LIDAR, which Waymo [Google X driving] seems quite invested in. [2] However, much of Waymo's self-driving training is simulated [of course it is, how else would you get enough data?]. This is filling out a gap that is currently otherwise only filled by proprietary solutions. That's a big deal. [1] http://kevinhughes.ca/blog/tensor-kart http://kevinhughes.ca/blog/tensor-kart [2] https://news.ycombinator.com/item?id=13352385 https://news.ycombinator.com/item?id=13352385
- cr0sh 10y ago
  > Is GTA V the same as setting up a rig which controls I/O for a real car? No. Of course not. You're not being fed data from LIDAR, which Waymo [Google X driving] seems quite invested in. I've been coming around to the conclusion that LIDAR isn't going to be the way forward, unless prices come down greatly and mechanical scanning is done away with. Perhaps something like Flash LIDAR, if it could be made 360 degree FOV and cost under $500.00 in quantity - that might be worthwhile for manufacturing a vehicle with, but currently even the cheapest 3D LIDAR is still way too expensive. I honestly think we'll see camera-based solutions becoming the norm; high-speed, high-resolution machine-vision cameras are already an established technology (though mainly black-and-white; pair those with lower-res color cameras and fuse the output as needed). Set up multiple stereo-camera pairs for overlapping 360 coverage - it would likely be a fairly cheap solution. At that point - the frame-buffer view data from GTA (or some other simulator) could be considered close to "real world" for data gathering and model training.
  - wrsh07 10y ago
    I think LIDAR is essential - at least for any cars to achieve L5 in the next 4 years. [which I also think will happen] I think it will become cheaper - probably to ~$1-2k in the next two years. And even if it doesn't, we aren't expecting self-driving cars to replace the fleet of cars on the road, we're expecting to increase utilization % of cars. So we won't need as many vehicles, and the cost per mile will still be much cheaper. But GTA V is scriptable. We'll be able to simulate LIDAR in-game within a few months, if desired. So... all-told, I'm pretty excited about the possibilities here.
hughes 10y ago
If it is highly scriptable, we may see people simulating sensor data within the GTA V environment. Data streams from cameras, lidar, and other types of sensors could be simulated with as much noise as you want. This could be used as input to train an AI using data that looks much closer to what you'd see in the real world.
- argonaut 10y ago
  This is way easier said than done. Even with added noise, the sensor data from the real world is nothing like simulation.

With all respect I think "easy-to-access" is a bit of an understatement when you are comparing a 3D game to real-world sensor data. They are barely comparable.

by drcross 10y ago

With all respect I think "easy-to-access" is a bit of an understatement when you are comparing a 3D game to real-world sensor data. They are barely comparable.

wrsh07 10y ago

The fact that you can even consider training an NN for driving after under an hour of setup is absurd. Did you see the "self driving mario kart"? [1] That's a lot of complexity just to try out some algorithms for driving. Is GTA V the same as setting up a rig which controls I/O for a real car? No. Of course not. You're not being fed data from LIDAR, which Waymo [Google X driving] seems quite invested in. [2] However, much of Waymo's self-driving training is simulated [of course it is, how else would you get enough data?]. This is filling out a gap that is currently otherwise only filled by proprietary solutions. That's a big deal. [1] http://kevinhughes.ca/blog/tensor-kart http://kevinhughes.ca/blog/tensor-kart [2] https://news.ycombinator.com/item?id=13352385 https://news.ycombinator.com/item?id=13352385

cr0sh 10y ago
> Is GTA V the same as setting up a rig which controls I/O for a real car? No. Of course not. You're not being fed data from LIDAR, which Waymo [Google X driving] seems quite invested in. I've been coming around to the conclusion that LIDAR isn't going to be the way forward, unless prices come down greatly and mechanical scanning is done away with. Perhaps something like Flash LIDAR, if it could be made 360 degree FOV and cost under $500.00 in quantity - that might be worthwhile for manufacturing a vehicle with, but currently even the cheapest 3D LIDAR is still way too expensive. I honestly think we'll see camera-based solutions becoming the norm; high-speed, high-resolution machine-vision cameras are already an established technology (though mainly black-and-white; pair those with lower-res color cameras and fuse the output as needed). Set up multiple stereo-camera pairs for overlapping 360 coverage - it would likely be a fairly cheap solution. At that point - the frame-buffer view data from GTA (or some other simulator) could be considered close to "real world" for data gathering and model training.
- wrsh07 10y ago
  I think LIDAR is essential - at least for any cars to achieve L5 in the next 4 years. [which I also think will happen] I think it will become cheaper - probably to ~$1-2k in the next two years. And even if it doesn't, we aren't expecting self-driving cars to replace the fleet of cars on the road, we're expecting to increase utilization % of cars. So we won't need as many vehicles, and the cost per mile will still be much cheaper. But GTA V is scriptable. We'll be able to simulate LIDAR in-game within a few months, if desired. So... all-told, I'm pretty excited about the possibilities here.

hughes 10y ago

If it is highly scriptable, we may see people simulating sensor data within the GTA V environment. Data streams from cameras, lidar, and other types of sensors could be simulated with as much noise as you want. This could be used as input to train an AI using data that looks much closer to what you'd see in the real world.

argonaut 10y ago
This is way easier said than done. Even with added noise, the sensor data from the real world is nothing like simulation.