Here is one I don't think has been posted where they did lots of testing of obstructions in the road:
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
FSD rewrite will go out on Oct 20 to limited beta
- Thread starter diplomat33
- Start date
Here is one I don't think has been posted where they did lots of testing of obstructions in the road:
It would be interesting if people also provided the Teslacam footage as well especially when it's dark or bright out. He does comment how the visualization shows his wife very early even though it gets darker, and she's wearing all black. From my recordings, the side repeaters can see things at night even without a direct light source, and the front camera picks up traffic light colors fine as long as the sun isn't directly behind the signal.
Similarly, this video up a mountain road at night looks extremely dark from the recorded perspective, but Teslacam probably has a much better view:
Ludalicious
Active Member
Don't know if he mentioned it or not in the video. But why did he pick 20 MPH as a testing point? Great amount of information in this video. Very impressive
powertoold
Active Member
I can see why people think Tesla uses HD maps. The intersection and road predictions are magical.
Yeah, at least for me, it's quite interesting to see when the visualization updates from a wrong curving road prediction to the correct T-intersection as the car gets closer and shines more light in less than a second:
Buckminster
Well-Known Member
MadOverlord
Member
I would not be surprised if one of the uses for the data flood from the fleet, in particular once the new FSD architecture goes wide, is to generate maps with centimeter resolution where they are helpful, and keep them updated. You don't need this level of mapping for every meter of roadway, but there are going to be locations (particularly, complicated intersections) where knowing what should be there will help process the sensor inputs, in particular in challenging environmental conditions.
A particular car might only have the enhanced map data for routes it commonly takes, plus the general map that has flag bits that mean "this location is difficult, try to get the enhanced data before you arrive".
It may well be the case that getting to 10x human performance on the edge cases of the driving task within a reasonable power/computation budget will require this kind of assistance; effectively it's precomputing some aspects of the problem.
There is a dedicated video thread in the autonomous section so I don't see the point in cluttering this thread up with them. It is a good place to discuss one that has interesting behaviors though.
HenryFarkas
Google "Henry Farkas Tesla"
The trouble with HD maps as a way for the car to know what's out there even in dense fog or other or visibility conditions is that even with excellent signals from the GPS network, there's still a few meters of uncertainty as to the actual location of the car. With poor signal quality, there's even more uncertainty. And that doesn't even cover cars parked in no-parking zones or other new stationary objects. If the car can't locate itself precisely, it can't depend on the HD map for precise driving.
Daniel in SD
(supervised)
Companies using HD maps don't use GPS for localization. As you said GPS is nowhere near accurate enough and is not reliable. They localize using LIDAR (Waymo/Cruise) or cameras (Mobileye). A car driving in fog so thick that no landmarks could be sensed would probably be driving extremely slowly.
Waypoint - The official Waymo blog: Building maps for a self-driving car
The Why and How of Making HD Maps for Automated Vehicles | Intel Newsroom
Buckminster
Well-Known Member
Another reason why HD maps is not a good way forward:
If you look at FSD videos, a remaining issue is the computer making it's mind up - left or right if a cone is in the way. The last thing it needs is being told the road continues straight ahead when their are cones ahead.
If you look at FSD videos, a remaining issue is the computer making it's mind up - left or right if a cone is in the way. The last thing it needs is being told the road continues straight ahead when their are cones ahead.
powertoold
Active Member
That's not particularly true either. HD maps don't dictate the car's path. The HD map defines drivable space, and the driving policy can be coded to move around the cones within the drivable space. So no, this wouldn't be an issue with HD maps.
I hope they do this, but I haven't seen any evidence of this happening aside from peoples perceptions.
I'd love to see it taken a step further where it's not only continuously updates the maps, but expected driving speed/position that can be individualized.
As example during my drive to work I slow down before the corner before a traffic light because I know vehicles can get backed up to the corner so I need to slow down before going around the corner. The current version of AP doesn't slow for the corner, and continues going at my set speed. So there is about a 10+ MPH differential from what I'd do, and what it does.
Another example is I don't take the way the navigation tells me to go because that would force me to take a left onto a busy highway without the assistance of a stop light. The only reason the nav thinks its the right way is its trying to stay on the same street name.
The last example is on the way home I put myself into the right side of the road in one section for smoothness since the left side has a bunch of bumps.
I don't want it to duplicate all my behaviors. On one stretch of road I drive like a speed demon just because its suited to it. I'm perfectly okay with it doing normal driving there.
So ideally there would be a save, and replay drive buttons.
In Tesla's patent for training FSD they talk about automatically training the neural net based on sensor data from vehicles, then deploying that updated model, and testing if it increases the accuracy.
GENERATING GROUND TRUTH FOR MACHINE LEARNING FROM TIME SERIES ELEMENTS - Tesla, Inc.
"In various embodiments, a portion of the training data is generated automatically from data captured from vehicles, greatly reducing the effort and time required to build a robust training data set. In some embodiments, the format of the data is compatible with a machine learning model used on a deployed deep learning application. In various embodiments, the training data includes validation data for testing the accuracy of the trained model."
bd7349
Member
He usually posts the videos at 1x speed as well. Here’s the link to that version of this drive:
I prefer these as well, makes it easier to judge and see where it made weird/wrong decisions.
Looks like there's new UI for S/X and 3/Y with 2020.44.10.2. For S/X, time and temperature are moved to the bottom right where it can be temporarily covered up by a message, and the speed limit is no longer hanging out in the middle of the screen and grouped with a smaller speedometer and autopilot icons:
https://twitter.com/teslaownerssv/status/1327538428196528128
And for 3/Y, the left side expands when activating Autopilot:
https://twitter.com/teslaownerssv/status/1327538428196528128
And for 3/Y, the left side expands when activating Autopilot:
Last edited:
Making full use of the wide display on the X to visualize a sharp 15mph curve at night:
Good thing the cameras can see enough in darkness and neural network can predict confidently even when the headlights aren't aimed at where you need to go.
Good thing the cameras can see enough in darkness and neural network can predict confidently even when the headlights aren't aimed at where you need to go.
Similar threads
- Question
