> 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.
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.
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.