I think this is the edgiest case I've seen while using Autosteer. How do you train a neural net to recognize that the bright lines on the road are sunlight?
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Real Autosteer Edge Case
- Thread starter Daniel in SD
- Start date
EVRider-FL
Active Member
diplomat33
Average guy who loves autonomous vehicles
I imagine it would be very difficult. That's a good example of why other AV companies don't do vision-only. They know that it is much harder to handle these types of edge cases with vision-only. That is why they use cameras, radar, lidar and HD maps. Radar and lidar would not be confused by the bright lines and the HD map would also tell you where the true road lines are. So using cameras, radar, lidar and HD maps makes it easier to handle these and other edge cases more reliably.
Haha. I normally don't look at the visualization while driving at 70mph. The car pulled hard right and I resisted which disengaged Autosteer. It would be cool if dash cam overlaid all the perception outputs and controls.
Now I'm thinking it would have been fun to see if I could find a spot to record all the Teslas driving through.
TresLA
Member
I think I understand the reasoning here, but how does such a reconciling of conflicting inputs work in this case? For example, how does the car know to ignore vision input here (presumably listening to HD maps of lane markings), but know to listen to vision input and ignore HD maps of lane markings when construction repaints lane lines to shift them left or right for a temporary time?
diplomat33
Average guy who loves autonomous vehicles
That would be done by the sensor fusion algorithms which would tell the car when to listen to what. In most AVs like Waymo, they fuse the camera, radar, lidar and HD map data early. So the car takes all the data and builds a single 3D view of the world based on all the data. In the case of construction zones, the car could take camera and lidar cues from cones, construction signs etc and the paths other cars are taking to know to follow the temporary lines and ignore the HD map. Also, the first car that encounters the construction zone could also update the HD map for subsequent cars. So the next cars that encounter that same construction could follow the updated HD map. The cars would be told that the HD map is updated.
TresLA
Member
Do cars (and associated data carrier plans) have the bandwidth for realtime HD maps?
It seems the question of improving sensor fusion and solving with vision only (which currently still uses real-time non-HD maps) isn’t answered. I’m not convinced which (if either) is the best path to solving self-driving. I hope parties pursue (at least) both until either seems definitively not worth whatever cost/time/sacrifice in light of working solutions.
diplomat33
Average guy who loves autonomous vehicles
The maps are stored offline in the cars. So, the cars don't need a cell or internet connection to drive.
TresLA
Member
I meant they have to be downloaded sometime, and if we’re talking about real-time updates then bandwidth matters. Even now, “normal” map downloads/updates are huge and take awhile.
Mobileye claims their AV maps (which used to marketed as HD maps) are 10kb/km so you could easily store a map of all the roads in the world in the car.
Mobileye REM™ | Road Experience Management
Mobileye’s Road Experience Management™ (REM) is the basis of a highly accurate, constantly refreshed HD map built for AVs that can benefit driver-assist tech and road maintenance.
The more I think about this case though the harder it seems. It's a construction zone so the car would know it couldn't trust the HD map. LIDAR can see lane lines so maybe that would be more reliable than cameras? In the end it seems like collision avoidance logic would prevent anything really bad from happening. Leaving your lane without signaling isn't that big an error as long as no one is there. It also seems that the system could observe the actions of all the other vehicles when determining where the lanes are (after all humans can drive on snow covered roads with no visible lane markings).
diplomat33
Average guy who loves autonomous vehicles
Some ideas:
1) They could use compression techniques to help with bandwidth.
2) They can abstract features on the map. For example, Waymo's vectorNet uses simple geometric shapes like lines and vectors to represent features like lanes. These geometric shapes would not take up a lot of bandwidth.
3) If there is a change in the map, they might be able to just download the specific change and not download the entire map. For example, if they need to add a new stop sign to the HD map, they can just update the map with the new stop sign, not download the entire map.
4) Robotaxis like Waymo and Cruise operate in geofenced areas. That probably helps with the bandwidth since they only need to download the HD map for a relatively small area, like 10 sq mi, not a big map.
I know Mobileye uses "AV maps" which are like HD maps but less detailed than what Waymo uses. Mobileye says that it only takes 10 kb per km of data to create the maps. That is not a lot of bandwidth. Cars would have no issue handling that. That is probably one reason why Mobileye claims their maps are more scalable than the HD maps Waymo or others use since Mobileye is able to do autonomous driving with less detailed maps that don't require a lot of bandwidth.
Are the maps preloaded into the car or does the car have to download maps every time it operates? Hopefully soon these cars will be able to cross state lines. Like NYC to NJ. or LA to LV.
diplomat33
Average guy who loves autonomous vehicles
They are preloaded into the cars. Of course, they would get periodic OTA updates with map changes but they would not need to download the maps every time they operate. That's silly. Does your Tesla have to download the software every time you go for a drive? No. And there is nothing stopping the cars from crossing state lines now.
Similar threads
