Unless the developers work as salesmen as well, they should be able to push forward at a steady pace while selling cars. Selling FSD is another matter.
Next decade is 1 year away so here's hoping for an insane jump from previous performance
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Firmware 9 in August will start rolling out full self-driving features!!!
- Thread starter diplomat33
- Start date
-
- Tags
- 9.0 Tesla Inc.
Unless the developers work as salesmen as well, they should be able to push forward at a steady pace while selling cars. Selling FSD is another matter.
Next decade is 1 year away so here's hoping for an insane jump from previous performance
Interesting enough today on my way to work i just ran into a edge case. I usually encounter road debris that i have to dodge every other month on the freeway (tire, etc). Today I encountered two. The first being a big bucket.
I initially noticed it about 350-400 meters away.
The bucket was in the left lane of a 4 lane divided freeway.
There was a left shoulder lane aswell.
The moment I saw it, I knew instantly that it was a bucket. (seeing 350-400m is though for any sensors let alone understanding what you see that far away.)
I was going 75+ MPH and I instantly started planning in my mind what i would do.
There were two cars far ahead of me. The bucket was 25% away from the right line of the lane.
The first car successfully dodged it By moving left and using 25% of the left shoulder lane.
After seeing that I thought, that's it, I'm gonna use the shoulder lane (SDC have to be able to see what others do, see if its successful, so they can copy it/ move it higher in their list of options)
The problem was as the first car went by, its wind effect made the bucket to roll to the left. Now the bucket is now in the center of the lane. The second car which was close behind the first car now had to make a quick decision by swerving into the left shoulder, using 75% of the shoulder to avoid the rolling bucket.
At this point I'm like Oh *sugar* ****!
I still had acouple more seconds of planning because i had seen the bucket from afar. But I had to change my plans. As I made my approach (75MPH). The bucket has now moved because of the second car wind and was now 25% to the left line of the lane.
Using the shoulder was no longer an option. A better option was to go into the next right line.
So I had to do a quick check before using 25% of the lane next to me to avoid the bucket.
What I just described is the Achilles heels of self driving. Seeing an object far away and being able to classify what the object is and predict how it will move. While also planning, canceling your plan, re-planing and negotiating while still going 75MPH.
This is the first time i actually encountered a moving obstacle that could do real damage to my car (atleast that i remember, my memory may be failing me). Usually they are all static although just as potent, sometimes even more potent. I didn't over-react, nor did i under-react.
It reminds me of this video:
Interesting enough today on my way to work i just ran into a edge case. I usually encounter road debris that i have to dodge every other month on the freeway (tire, etc). Today I encountered two. The first being a big bucket.
I initially noticed it about 350-400 meters away.
The bucket was in the left lane of a 4 lane divided freeway.
There was a left shoulder lane aswell.
The moment I saw it, I knew instantly that it was a bucket. (seeing 350-400m is though for any sensors let alone understanding what you see that far away.)
I was going 75+ MPH and I instantly started planning in my mind what i would do.
There were two cars far ahead of me. The bucket was 25% away from the right line of the lane.
The first car successfully dodged it By moving left and using 25% of the left shoulder lane.
After seeing that I thought, that's it, I'm gonna use the shoulder lane (SDC have to be able to see what others do, see if its successful, so they can copy it/ move it higher in their list of options)
The problem was as the first car went by, its wind effect made the bucket to roll to the left. Now the bucket is now in the center of the lane. The second car which was close behind the first car now had to make a quick decision by swerving into the left shoulder, using 75% of the shoulder to avoid the rolling bucket.
At this point I'm like Oh *sugar* ****!
I still had acouple more seconds of planning because i had seen the bucket from afar. But I had to change my plans. As I made my approach (75MPH). The bucket has now moved because of the second car wind and was now 25% to the left line of the lane.
Using the shoulder was no longer an option. A better option was to go into the next right line.
So I had to do a quick check before using 25% of the lane next to me to avoid the bucket.
What I just described is the Achilles heels of self driving. Seeing an object far away and being able to classify what the object is and predict how it will move. While also planning, canceling your plan, re-planing and negotiating while still going 75MPH.
This is the first time i actually encountered a moving obstacle that could do real damage to my car (atleast that i remember, my memory may be failing me). Usually they are all static although just as potent, sometimes even more potent. I didn't over-react, nor did i under-react.
It reminds me of this video:
Alternative strategy:
1. hit the hazard warning and slow the hell down when seeing cars swerving around obstacles up ahead.
2. pull over when safe and notify authorities to remove hazard.
P.S.: Am having severe difficulty believing you at 75mph+ could even see [much less identify] a bucket at 350m through two leading vehicles.
Last edited:
That is what first Level 3 vehicles will probably do in this scenario except the notification to authorities is notifying the driver. Lidar can see the bucket even if it is not identified. Though it is not hard to see at all that this will not be an especially good solution. But at Level 3 it probably would suffice.
P.S.: It is not hard to see a road scenario where curvature clearly shows the road ahead of leading cars.
I find that taking human descriptions of events literally and trying to find illogicalities in literal interpretations is road only to misunderstanding and unnecessary doubt. It is probably much more fruitful to try and see what @Bladerskb meant rather than trying to poke holes into it.
Distances and exact description are not the point in my view. Those a human (even one that is always right) can make mistakes on after all they are just their visual take and how they remember things and what they found worthy of mention when recounting it. The overall event here is much more interesting I think.
A bucket on a highway? Completely plausible. Moving because of air movement caused by other cars? Again plausible. Thinking of bypass strategies instead of stopping dead on a highway? Normal.
For me this one isn’t a particularly hard one to imagine or believe at all. All it takes is a winding road, uphill or @Bladerskb in a much taller car than was in front and seeing beyond leading cars is trivial. Trash on a highway is common too.
The road certainly wasn't completely flat, it had a tiny tiny dip and very small curve radius. Plus my gap between the front cars were like 10+ cars length so i could see the cars and what was ahead of them. i had atleast 250m gap. But yes my speed was constantly above 75mph and i saw and identified the bucket 350-400 meters. I know its 350m+ because i had been doing some speed/distance test in my car this past week to see how far 150m is and how much distance relative to the driver 5 seconds is while going 65 mph.
Slowing down in situations like this can actually escalate the problem, especially if you don't see the obstacle 350m+ away like i did.
Excessive deceleration would lead to a rear end.
I'm pretty sure that's often the fault of the RADAR. It can't always tell the difference between a sudden rise in the road and an actual obstacle, because it isn't precise enough. When that happens to coincide with deep shadows from road signs or bridges, suddenly both the RADAR and the computer vision say something is happening, and it panics.
This should get better as the NNs get better at recognizing depth clues.
You left out the third (and IMO most reasonable) option: Do nothing.
The worst case for evading is smashing into another car, which is much worse than hitting a bucket, and the worst case for slowing down too quickly is getting rear-ended, which is also much worse than hitting a bucket. The worst case for hitting the bucket is probably some paint damage and minor dings, which is what insurance is for.
Once you determine that an object is neither a person nor something large enough to cause life-threatening damage, the best choice is often to just keep driving and hit the obstacle.
You don't need to be on the millimeter either. Or else we couldn't drive cars ourselves. If we humans worked like a LIDAR, our heads would spin above the sunroof 360 degrees every second
[evil voice]Get back! The road is mine![/evil voice]
“Do nothing” is what they currently err on the side of doing and it’s frankly not working. At least two people have died within 3 years of time on two different hardware designs due to the “do nothing” strategy of dealing with unrecognized obstacles.
AP1 doesn't do lane changes on local roads anymore? Sorry I didn't know they stopped that.
Your sarcasm wasn't required. AP2 does lane changes in local roads now I'm sorry to inform you. It would seem the AP1 worshippers think that AP2 is stuck in its original pathetic incarnation forever, which is simply not true.
Last edited:
I think @conman @BigD0g just was not aware AP2 now does local lane changes. It wasn’t that well publicized probably given that Tesla sticks to their use on divided roads only recommendation so these things happen in the shadows.
Sorry to hear about your accident @Bladerskb. Hope you and your possible passengers are fully okay.
Sorry to hear about your accident @Bladerskb. Hope you and your possible passengers are fully okay.
This is the cause of a lot of the phantom braking, but it's not new in V9. I think it's slightly worse in V9 but not significantly. What is new in V9 is that they are now paying more attention to cars to the sides and slightly ahead. I think they're trying to handle merging vehicles better, and that requires them to rely more on the cameras to determine if those nearby cars are coming into your lane or not -- this requires very precise 3D positioning and measurement of the full extents of the vehicle. Do do this with cameras rather than lidar requires accurate 2D bounding boxes (or better yet masks) combined with accurate depth information. And they simply don't have either reliably, which leads to false positives that neighboring cars are in your lane, or coming into your lane.
I just spent more words than I really ought to have describing why they are inherently handicapped in getting depth information from their cameras. They have some room for improvement but they're fundamentally limited and whatever they do, even if it is highly accurate in 99% of cases, will be wrong a significant portion of the time. Enough to make the false positives (and false negatives) very problematic (and potentially deadly).
This is exactly what I did years ago when I was in a similar situation. A ladder fell off a truck in front of me on the highway. It was nighttime which made everything even more dangerous. I had only a fraction of a second to decide. There was no conscious thought in that moment, only the conditioning of years of driving. In my case, that conditioning (not my conscious thought) decided that the risk of entering an adjacent lane without being able to check first if there was a car there was higher risk than the ladder. I just moved slightly to center the car over the ladder so my tires didn't hit it, and drove right over it. (It was still bouncing and sliding on the pavement.) I had some damage to the undercarriage but it was repairable, and all occupants (which at the time included by newborn daughter) were safe.
A "software v1.0" hand-coded motion planner could do exactly the same sort of reasoning in this situation, except it wouldn't need to since it would already have a 360-degree view of the cars around it and would know for sure whether it could safely swerve.
A pure deep learning, end-to-end NN approach, unless it had very similar scenarios in its training data, would do something random and unpredictable. You would need nighttime highway scenarios of a ladder falling off a truck (or something ladder-like and truck-like) with various combinations of neighboring vehicles of all sorts to cover different cases of swerving left being safe, swerving right being safe, and neither being safe. You'd need to simulate all of these combinations at various speeds, traffic densities, lighting conditions, weather/road conditions, different vehicle types, obstacle types, road curving left, road curving right, going uphill/downhill, etc.
You would need a set of scenarios like that for all sorts of things that could happen suddenly on the road, like a tree falling or an animal running into the road. This is why end-to-end "software v2.0" is not going to be ready for another decade. Sure, it may be ready eventually, but Tesla will be irrelevant by then if they stick to that plan.
Similar threads
