Where are these stats coming from? Quickly looking up this data for my city, it's more like 7% of all accidents involved left turns across oncoming traffic and that includes left turns at lights
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Does Tesla have a big math problem?
- Thread starter SanDiegoM3
- Start date
I didn’t read the pdf due to length but this seems to reach a similar conclusion on lefts.
Attachments
rxlawdude
Active Member
"Left-turning traffic constitutes between 10 to 15 percent of all approach traffic at an intersection yet they contribute approximately 45 percent of all traffic crashes. At signalized intersections, the typical left-turn accidents happen during the unprotected left-turn phase because of the left-turn vehicles that fail to yield right-of-way." (Emphasis added.) Page 1 of the report.
"Based on the literature review, the main factors that adversely affect safety of the left-turn traffic include incorrectly accepting a gap between the coming-through traffic, sight distance obstruction caused by opposing left-turning vehicles, left-turn driver’s distraction, and misunderstanding the intersection signal phase." ibid p. 24.
Note that these analyses are based on signalized intersections, not the type (stop sign for secondary road at divided highway intersection) that Chuck Cook uses to assess UPLT with FSD beta.
Daniel in SD
(supervised)
Either way it's not a very useful stat. What is the failure rate of humans drivers making various types of left turns? That's the metric.
What percentage of posts on this forum are "humans are bad drivers" statements with no useful statistics?
rxlawdude
Active Member
Not following your logic or what you're trying to say here. Contextually objective statistics were quoted in my post above.
This report also does not seem to cite a source, or even provide a table of the breakdown, and I have a feeling there's an issue with the wording in "contribute approximately 45 percent of all traffic crashes". 45% of all traffic crashes at intersections maybe, but 45% of all traffic crashes regardless of location? I don't think there's any way that is true.
Not able to find much on the net for breakdowns but I did find this from Canada
Source: https://open.alberta.ca/dataset/250...alberta-traffic-collision-statistics-2017.pdf
Daniel in SD
(supervised)
I'm saying that they tell us nothing about how reliably FSD has to perform left hand turns (which is the subject of this thread).
It really should have been a response to the original post about percentage of collisions at left hand turns.
The baseline FSD crash metric to test against isn’t human ability, it’s liability. If unprotected lefts are not only very difficult to execute and significantly more crash prone, whether FSD can take a 45% of all crashes to 35% of crashes isn’t helpful because now FSD is involved in significantly more crashes overall and I the consumer am taking personal liability for something a car knowingly performs poorly and my individual skill could be higher than the vehicle’s.
Unprotected lefts and rights are a big problem because I generally don’t have the right of way.
If the motor cycle is coming directly towards you, say at 60 mph, and you want to turn right into it’s path, I don’t think 4 pixels at 160m is close to being enough to get an accurate enough velocity vector to make a safe maneuver. The motorcycle is upon you in about 5 seconds
In fact all this math (as expected) turned out to be nonsense, since FSD beta is doing, right now, exactly what the math was trying to prove was not possible.
Uhhh yeah that’s not true. All you have to do is watch the YouTube guy who does all the unprotected lefts in the beta. He would literally get demolished in the latest version. In fact, it’s so unusable that Tesla should be thankful every day that it doesn’t have some drunk idiot using FSD to take himself home with unprotected lefts on the way.
Let me guess, it’s beta it’s going to get better. Extremely doubtful. The camera placement is not conducive to this type of maneuver and the camera simply doesn’t have the visual depth to know when it should go and when it shouldn’t and what to do when it makes a mistake and it has to bail out of the maneuver or rocket faster.
Can it do the maneuver? Yes. But in general it’s like 6-8 out of 10 success which is simply terrible. It was 1 of 4 this version, disgustingly dangerous.
Last edited:
EVNow
Well-Known Member
Waymo wouldn’t do those too. Unprotected turns when the speed is that high are a hazard - just poor job by city/county.
Massive regression. It’s possible chuck could’ve been killed had either he or the oncoming vehicles not intervened or made evasive maneuvers.
0:56 creeping into traffic
2:20 creeping into traffic
3:05 likely would’ve been hit by middle lane based on timid behavior
5:20 potential t-bone
5:55 no stop at stop sign stalled in high speed road
6:40 continued into road with oncoming vehicle
7:10 potential t-bone
0:56 creeping into traffic
2:20 creeping into traffic
3:05 likely would’ve been hit by middle lane based on timid behavior
5:20 potential t-bone
5:55 no stop at stop sign stalled in high speed road
6:40 continued into road with oncoming vehicle
7:10 potential t-bone
Last edited:
jabloomf1230
Minister of Silly Walks
Since this has been posted elsewhere:
teslamotorsclub.com
FSD Beta Videos (and questions for FSD Beta drivers)
Gets stuck behind an obviously stopped truck. Tesla shows the yellow lights as red lights and draws the truck pointing the wrong way It's awesome, it's like such a zen relaxing experience... all the stress is just washing away, 'cause I'm in total zen mode in my Tesla A minute later...
teslamotorsclub.com
Since this has been posted elsewhere:
FSD Beta Videos (and questions for FSD Beta drivers)
Gets stuck behind an obviously stopped truck. Tesla shows the yellow lights as red lights and draws the truck pointing the wrong way It's awesome, it's like such a zen relaxing experience... all the stress is just washing away, 'cause I'm in total zen mode in my Tesla A minute later...
I don’t think it’s final boss level at all. It’s an unprotected four lane highway where cars can likely go 50-55mph, a typical country road speed. The cameras can’t see far enough, the computer can’t process it fast enough, and the car can’t do the maneuver quick enough.
At 55mph, a vehicle reaches you in 3.25 seconds. But, the available time to maneuver is significantly less. You have to move forward to get above cars coming from the left, and you have to complete the turn and accelerate to not get smashed from cars coming behind you. Then, add detection time, calculation time, final decision making time, and changing conditions while in the maneuver and you get what we see. A human can clearly see better than the side cameras and we have two eyes, not a single lens.
I don’t think Tesla solves FSD in the current sensor suite. They’ll likely come out with new cameras and sensors and then deal with the likely lawsuits by offering free FSD on a new car or a transferable license.
I don't disagree with your general conclusion that the sensor suit is inadequate but be careful when making statements like "the computer cannot process it fast enough". What do you base this on?
Last edited:
Tesla having to continue to try new ricks to get the latency down? When you add more cameras you need more processing power. If you increase resolution you require more processing power. More frames per second equals more processing power.
In my view it mostly comes down to an ability to see and process long distance objects in a variety of weather and time of day patterns while learning how to handle them. That can be incredibly complex as what is obvious to a person may need to be specifically coded.
While I generally agree with you, I think it's also possible it's not so cut and dry in practice.
For example, maybe you have a high res camera, but you don't actually run any software on it unless a low res camera predicts "maybe that's a sign off in the distance there", and then you only do the additional processing to read text on the part of the high res image where the sign might be. Furthermore, if you know you're doing that processing to read sign text, then you know getting the answer within 1/36th of a second is probably not super important, so you can de-prioritize that task and spread it out to execute in any available idle time between the higher priority processing scans.
I could also imagine using existing low-res images to predict the geometry of the lane up ahead to as far as that camera can identify, and then you limit your high res image processing to the part of the image corresponding to where the low-res processing couldn't see it very well anymore.
Techniques like these could actually result in less processor utilization overall, by spending less effort trying to make something out of nearly nothing with low res images when it might be easier to do a quick pass on low-res images to guess where it's worth analyzing, and doing the analysis on higher-resolution images.
Btw - I'd bet money Tesla, Waymo, Cruise, and many others are already using techniques like these to avoid wasting precious processing resources analyzing image data they can already predict won't have relevant information.
Starting at the other side of this. What is the human reaction to "side view"
Regular humans have a reaction time of around .250 (1/4 second) but that doesn't include the time it takes to recognize the situation then hit the brake/go pedal etc.
Within that 1/4 second one camera at 36fps would have 9 whole frames to figure out what's going on and react, or 72 frames for all of them.
Add the fact the drivers spend most of their time looking forwards and many seem to need extra prodding to bother looking sideways or backwards.
On the face of it, it would seem the computer would react significantly faster than a human and spot oncoming cars from either direction faster than a human (who can't look in both directions at once)
Regular humans have a reaction time of around .250 (1/4 second) but that doesn't include the time it takes to recognize the situation then hit the brake/go pedal etc.
Within that 1/4 second one camera at 36fps would have 9 whole frames to figure out what's going on and react, or 72 frames for all of them.
Add the fact the drivers spend most of their time looking forwards and many seem to need extra prodding to bother looking sideways or backwards.
On the face of it, it would seem the computer would react significantly faster than a human and spot oncoming cars from either direction faster than a human (who can't look in both directions at once)
Humans have a huge instinctive advantage though. I just came back from a bike ride in a busy area with people, bikes, cars, pets everywhere. Lots of obstructed cross traffic barely visible through parked car windows. Bombing along at 30kmh with so many parked cars that the two-way street is really only wide enough for one car so you have to quickly determine what's happening and react accordingly.
I'm realizing how hyper aware we humans are, just a flick of the eyes in each direction to take on the constant changes. And yet we can do it, by sensing the sound queues of people and cars, feel the vibration of engines behind me and hidden in the cross-traffic. Seeing through car windows, shadows on the ground indicating objects, recognizing movement, and knowing instantly what each object is and what its potential for movement might be. We can determine closing speed very well, not in numerical terms but whether it is going to intersect with our intended pathway.
These are so instinctive that we don't think about them. Children don't have this experience so have to learn, it takes time. But we can take the whole scene in and make extremely rapid decisions.
AV's may have some advantage in sensor quantity, but we're way quicker in making it work.
Similar threads
