Apparently a video from Tesla on what the AP system sees and interprets on the road.
Last edited by a moderator:
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
This is what the autopilot sees...
- Thread starter mspohr
- Start date
travwill
Active Member
Yeah, that is from MobilEye directly done as a presentation on what their v3 chips can do, etc with the right software. Tesla however doesn't likely use them to that extent with the single camera hardware and software. Tesla seems to use a lot less data or detects a lot less; e.g. only moving cars somewhat directly ahead (not wide), etc.
I'm pretty sure Tesla tracks cars in adjacent lanes. Yesterday, i was driving on a two lane freeway with tacc set to 75. There was a car in front of me doing about 65, and another passing us on the left going only slightly faster (maybe 70). I used auto lane change to move over to the left, and was a bit surprised that my car didn't attempt to speed up while changing lanes. I expected it to speed up while changing lanes, then slow down again once it locked on to the car in the new lane, but it seemed to already be aware of the car in the left lane before even starting the maneuver.
MichFin
Member
This is the MobilEye technology and while Tesla doesn't use the stop lights and signs yet. It's cool to see how they are percieved and likely used in the future.
GlmnAlyAirCar
Active Member
Perhaps traffic lights are used to some (unadvertised) degree. I was second in line at a red light the other day with autopilot engaged. The car in front of me turned right on red. To see what AP would do, I left it engaged. It started to pull forward to the intersection then I received the "Take over immediately" warning, which I did.
Did AP know the light was red and, noting the car I was following disappeared, needed me to make a decision? Possibly. Of course there could be other explanations but I would be interested to hear others' experiences in this same situation.
garygid
Member
More likely, I think, is that the AP just fails to aquire lane lines through the intersection,
so it wants you to manually control the steering wheel. Under TACC alone, or if it felt
that it recognized sufficient lane markings, it would just attempt to accelerate right
into the intersection, in spite of the red light. This is a situation to avoid.
This is the main reason that AP or TACC should not be used at intersections.
AP is currently designed for limited-access (no intersections) highways only.
so it wants you to manually control the steering wheel. Under TACC alone, or if it felt
that it recognized sufficient lane markings, it would just attempt to accelerate right
into the intersection, in spite of the red light. This is a situation to avoid.
This is the main reason that AP or TACC should not be used at intersections.
AP is currently designed for limited-access (no intersections) highways only.
GlmnAlyAirCar
Active Member
Yes, your assessment makes sense, and that is another possibility I considered, but I'm not yet sold on it.
On the other hand - I don't agree with your "designed for limited-access" statement. There is a difference between what it is designed to do and what use is supported and endorsed. If the system were not designed to work on surface streets, it wouldn't work on surface streets. It would either disengage because it didn't understand its surroundings or simply try to keep driving as if it was on the highway. We know it doesn't do the former, and it doesn't do the latter because as I'm sure we have all seen it can make pretty tight turns on clover-leafs and other windy roads at slow speeds. So, it IS designed to work on surface streets.
However, Tesla does not approve or endorse its use. There are unpublished limitations on its use, which we are discovering here. I imagine testing is not as thorough either. But clearly the works in many conditions other than limited access highways and this is not an accident.
GasKilla
No Gas Know Peace
Here is a link to a presentation from the founder of MobilEye Brains, Minds Machines Seminar Series: Computer Vision that is changing our lives - YouTube it is long but worth a viewing to really understand what is going on with the tech. The OP's clip is taken from this presentation. The founder starts his presentation about 4:00 into the video. Very interesting and worth a viewing.
garygid
Member
dkelly,
Your statement
"If the system were not designed to work on surface streets, it wouldn't work on surface streets."
is incorrect logically.
If the system was designed to not work on surface streets, it wouldn't work on surface streets.
would be more nearly correct, assuming that the design worked as intended.
The AP was designed for some conditions usually found on limited/restricted access roads.
It was not designed to handle signal lights, stop signs, or "follow" a car.
It was not designed to detect all the conditions where it would have trouble.
It was designed to operate reasonably well (beta version, may have bugs) in
a sufficiently well detected specified set of circumstances, mainly where it
can visualize lane boundries, and not need to deal with all the various kinds of
speed/stop sign/signal/control issues.
The fact that stretches of surface streets happen to meet these criteria does not
mean that this AP was designed to reject, or even detect, the intersections, cross
traffic, sihnal lights, etc.
Please be careful when testing AP under ANY conditions.
In my personal experience with 2.7.56 version of AP, there
are still hidden interactions that appear to make the steering
misbehave, even on the freeway.
Your statement
"If the system were not designed to work on surface streets, it wouldn't work on surface streets."
is incorrect logically.
If the system was designed to not work on surface streets, it wouldn't work on surface streets.
would be more nearly correct, assuming that the design worked as intended.
The AP was designed for some conditions usually found on limited/restricted access roads.
It was not designed to handle signal lights, stop signs, or "follow" a car.
It was not designed to detect all the conditions where it would have trouble.
It was designed to operate reasonably well (beta version, may have bugs) in
a sufficiently well detected specified set of circumstances, mainly where it
can visualize lane boundries, and not need to deal with all the various kinds of
speed/stop sign/signal/control issues.
The fact that stretches of surface streets happen to meet these criteria does not
mean that this AP was designed to reject, or even detect, the intersections, cross
traffic, sihnal lights, etc.
Please be careful when testing AP under ANY conditions.
In my personal experience with 2.7.56 version of AP, there
are still hidden interactions that appear to make the steering
misbehave, even on the freeway.
GlmnAlyAirCar
Active Member
Please read all of my post. There are behaviors, such as following a vehicle around a sharp turn (eg exit ramp in traffic) that are inconsistent with highway only driving. This behavior cannot be an accident. If it didn't work on surface streets, you would get the "take over immediately" warning whenever it entered that mode.
However, we know there are significant limitations using it off highway. Center lane lines but nothing at the shoulder causes the car to get precipitously close to the edge. So don't use it under this condition. There are many other situations where it doesn't work so well, along with many others where it does. Tesla obviously has no intention to train all drivers on the specifics; as a result the general "highway use only" restriction.
Regarding your "be careful" admonishment, one would be crazy to not have both hands on the wheel under any of these conditions. If the car does something unanticipated, take over. Of course, I nearly always have at least one hand on the wheel on the highway, especially when approaching turned or near other traffic. I'm not going to let the car so anything I don't want it to do. It's great technology but certainly something that needs to be monitored closely, just as with an aircraft.
Oh yes, 8500+ hours flight time, and I now design and maintain steering control systems, including autopilot software, for some very big (90000 ton) ships, so I do feel qualified to weigh on here.
Similar threads
