Reciprocity
Active Member
In a vehicle that is capable of Level 5 Autonomy, you wont need road markings to actually be that good. Remember that Level 5 requires the car to be able to drive in snow where no lane markings are visible and a person does not even have to be in the car. How the heck will that work you might ask? Its actually not as complicated as you might think, but its not easy either. The cars will download hi-def 3D map tiles that includes paths which the car can drive in. The car will use GPS to get its general location then it will use that to get the map tiles. Included in the map tiles are objects the system uses as landmarks. Things like signs, traffic lights, lane markings (when the car can see them) and other things that are at least mostly permanent and stationary. Once the system can recognize a landmark, it can then dial in its location in the 3D tile down to 5-10cm of accuracy. At that point, the car just follows the path and doesnt need to see lane markings at all. Now the vision system will still keep an eye on the road, because things change. Think of it as a backup to the path from the hi-def map tile. Radar and Vision is used to identify objects that are not included in the mapping tile. These objects are important to avoid if they are in way and they are physical objects and not just shadows or something that is not in the way. As you drive, the system will download tiles that you might need, in front of you and adjacent roads. This will keep map tiles small enough to download quickly. If you drive a certain path often, the tiles will be stored for longer and updated as they expire. Tiles will also be kept until you need the space as people tend the drive in the same areas. Tiles will be constantly updated as the many thousands of cars see new objects and new landmarks and as landmarks move or disappear.
People often think that Lidar is required and it might help. But Lidar could also cause confusion. If Linda, Vision and Radar all have a different idea of they see, which is right? Sometimes more sensors is not actually better. Actually better sensors are always better. I believe this is why Tesla decided to focus on Vision. It is going to be very hard to get Vision to work, but not so astronomically harder then using 2 lidars and many radars and a bunch of other stuff. Fully redundant systems is also a bit overrated because if someone fails you have to fix it regardless. If the entire vision system fails, the car can sill pull over the side of the road and turn on its hazards, no different then if you get a flat tire. This is not a military application where the car is going to be shot at and it has to function until the mission is complete at all costs. Anyone of the systems could fail and the car should still be able to safely pull over or at the very worst case, stop in place until it can be towed. And of course the instant it fails, Tesla will know and have a ranger in route.
I know this was a long winded response, but Autonomous cars need to be able to drive in all kinds of situations and road markings wont always be good as they are not always good for people. The advantage autonomous cars will have is that they will have driven every road tens of thousands of times when it might be your first time ever on the road. Machine learning is basically a brute force activity and that's why you need so much computing power in the car.
People often think that Lidar is required and it might help. But Lidar could also cause confusion. If Linda, Vision and Radar all have a different idea of they see, which is right? Sometimes more sensors is not actually better. Actually better sensors are always better. I believe this is why Tesla decided to focus on Vision. It is going to be very hard to get Vision to work, but not so astronomically harder then using 2 lidars and many radars and a bunch of other stuff. Fully redundant systems is also a bit overrated because if someone fails you have to fix it regardless. If the entire vision system fails, the car can sill pull over the side of the road and turn on its hazards, no different then if you get a flat tire. This is not a military application where the car is going to be shot at and it has to function until the mission is complete at all costs. Anyone of the systems could fail and the car should still be able to safely pull over or at the very worst case, stop in place until it can be towed. And of course the instant it fails, Tesla will know and have a ranger in route.
I know this was a long winded response, but Autonomous cars need to be able to drive in all kinds of situations and road markings wont always be good as they are not always good for people. The advantage autonomous cars will have is that they will have driven every road tens of thousands of times when it might be your first time ever on the road. Machine learning is basically a brute force activity and that's why you need so much computing power in the car.