Autonomous Car Progress

[Doggydogworld]

Brand new video of a real world test of Mercedes' L3 highway system on the EQS:

"Real world" on a closed track with other vehicles following a script

Is it just me or did it seem like the driver did not really have enough time to properly re-engage when the emergency vehicle zipped by at the 12:06 mark in the video?

IMHO the driver didn't need to respond before the emergency vehicle passed. The car asks the driver to engage in case the emergency vehicle doesn't simply pass by, but pulls the Mercedes over or moves to block traffic or something.

The re-engage chime at 13:58 bothers me more. What would the car have done if the driver had waited 9.5 seconds to take over? Come to a stop in the lane, I assume. That's "safe", but would further jam traffic.

If you think L3 systems will be able to predict a fault or condition 10 seconds ahead at a 100% accuracy rate, IMO you're dreaming.

Nobody thinks that, certainly not Mercedes. The car will slow for the condition, even stop if necessary. Bad for traffic flow, but beats plowing into a fire engine at full speed.

I propose throwing L3 in the trash,...

That's what Waymo did. And what a few OEMs said they would do. The OEMs may backslide, though.

 

[Tam]

The test for upcoming Mercedes level 3 with LIDARs by a reporter shows that it didn't colide with the partially obstructing car from the left lane and it just passed around it. Radar-acitvated Tesla has problems with colliding with partially obstructing cars like firetrucks, police cars... so someone needs to test this scenario with the current pure vision version.

 

[Daniel in SD]

so someone needs to test this scenario with the current pure vision version.

With thousands of different vehicle types and thousands of different lighting conditions. There are Tesla test vehicles around with LIDAR and people have speculated that they use them to test or train the vision system.

 

[diplomat33]

Elon offered an update on FSD during the Q2 earnings call just now:

We are making great progress on FSD. Progress is not easy to see because it is at a foundational software level. It ends up being two steps forward, one step back situation. But over time, if you do two steps forward and one step back, you move forward. So, I am highly confident that the cars will be capable of FSD, with the FSD computer and the cameras. I am confident they will be able to drive themselves with a safety level substantially greater than the average person.

 

[daktari]

My point is why "10 seconds?" Taking my mind or eyes off the road for 2-3 seconds still gives me the ability to pre-assess upcoming roadway and react accordingly if TACC/NoA freaks out.

Allowing up to 10 seconds for a driver to take over is the real recipe for disaster. "Oh, the car will catch it while I'm reading my newsfeed, and warn me in plenty of time." Till the case where all that pre-assessment (we, as humans should always be doing) is beyond the AV's ability, which is inevitable. This is why an L5 AV is going to be elusive, an L4 limited in scope, but something L2-3ish becoming the norm in driver assistance tools. That's really all I expect and would be happy with is competence similar to the NoA/FSD features on the highway for city. Still need to be paying attention, and still will be the driver.

2-3 seconds eyes off in a Level 3 system is not as good as 10 seconds IMO. I expect a Level 3 system to safely initiate a stop procedure after the wake-up-time, not disengage and crash. 3 seconds would be to fast. You need some time to close the kindle/computer etc and be ready to keep the flow.

But you have not been eyes off in a Level 3 system? None of us probably.
Hope you don't go eyes off with AP or NoA, cause that is a Level 2 system, but that is not what you are saying?

 

[stopcrazypp]

The test for upcoming Mercedes level 3 with LIDARs by a reporter shows that it didn't colide with the partially obstructing car from the left lane and it just passed around it. Radar-acitvated Tesla has problems with colliding with partially obstructing cars like firetrucks, police cars... so someone needs to test this scenario with the current pure vision version.

View attachment 688433

There are examples of even the older Tesla system avoiding partial lane objects (even stationary or slow ones, which are the biggest issue with radar), but being able to do so occasionally isn't really necessarily that telling.
For example, here's one of a construction cone just from a quick search (this was before the system was able to even visualize the cones, another example showing that things not showing on the visualization is not necessarily the same as the car not detecting it):
Watch Tesla Autopilot avoid object in the road flawlessly in construction zone
In video below it started avoidance maneuvers (and sounded a warning) for a box and a moving SUV intruding into the lane (there are many more examples in other videos where it avoids side swipes).

But in the latter part of the video around 7:40 it shows a failed example for the same type of cones as the first example.

When there are millions of miles of driving done by real users, that's when we'll likely see the real limitations. I think the question however is if the lawyers will really let L3 into the world in any significant volume such that we'll see the systems being used that often (Audi cancelled after a big media demo push, Honda is doing a very small limited 100 vehicle fleet). Also keep in mind other cars don't have built in dash cams, so the likelihood of any failures (especially partial failures where a crash didn't happen) being caught on camera is much lower.

 

[rxlawdude]

The test for upcoming Mercedes level 3 with LIDARs by a reporter shows that it didn't colide with the partially obstructing car from the left lane and it just passed around it. Radar-acitvated Tesla has problems with colliding with partially obstructing cars like firetrucks, police cars... so someone needs to test this scenario with the current pure vision version.

I think you have made a factual misstatement that is ridiculously biased. All of those collisions were with STATIONARY partially obstructed vehicles.

Why do people do this and omit the most salient fact? My bet is in this exact situation in the Mercedes video, Tesla FSD/EAP would handle it competently.

What Mercedes needs to do is test this with a stationary partial lane block. Get back to us when they do so that you compare apples to apples.

 

[Daniel in SD]

I think you have made a factual misstatement that is ridiculously biased. All of those collisions were with STATIONARY partially obstructed vehicles.

Why do people do this and omit the most salient fact? My bet is in this exact situation in the Mercedes video, Tesla FSD/EAP would handle it competently.

What Mercedes needs to do is test this with a stationary partial lane block. Get back to us when they do so that you compare apples to apples.

At 3:10 in the video they show stopped cars partially in the lane.
This sort of argument is silly though since obviously passing this test a couple times in a demo video (or on the street) tells us nothing about how reliable it is in real life over millions of miles. We'll only find that out if they release it.

 

[rxlawdude]

At 3:10 in the video they show stopped cars partially in the lane.
This sort of argument is silly though since obviously passing this test a couple times in a demo video (or on the street) tells us nothing about how reliable it is in real life over millions of miles. We'll only find that out if they release it.

And remember that the infamous Tesla collisions with emergency vehicles occurred at speeds much higher than 60kph (40mph).

 

[stopcrazypp]

At 3:10 in the video they show stopped cars partially in the lane.
This sort of argument is silly though since obviously passing this test a couple times in a demo video (or on the street) tells us nothing about how reliable it is in real life over millions of miles. We'll only find that out if they release it.

Right, and this is far more important for L3+ vehicles, as the "driver" is not expected to pay attention to the road when system is operating. For L2 you can tune the system to ignore such situations (to avoid phantom braking or unnecessary swerving) but you can't do that for L3+.

 

[Tam]

...All of those collisions were with STATIONARY partially obstructed vehicles....

Yes, indeed. We are talking about the same thing both in Tesla and Mercedes with stationary vehicles. That is the nature of automobile radar limitations with stationary obstacles. That's why Tesla is eliminating that fault by implementing radarless feature.

Back to the video, there were a partially obstructed on the right shoulder which was stationary below. The one captured on message above is also stationary.

...My bet is in this exact situation in the Mercedes video, Tesla FSD/EAP would handle it competently...

No doubt Tesla FSD will be competent in avoiding obstacles. The emphasis is the word "will" with no timeline in future. I would not bet "soon" like in months or several years.

 

[daniel]

My point is why "10 seconds?" Taking my mind or eyes off the road for 2-3 seconds still gives me the ability to pre-assess upcoming roadway and react accordingly if TACC/NoA freaks out.

Allowing up to 10 seconds for a driver to take over is the real recipe for disaster. "Oh, the car will catch it while I'm reading my newsfeed, and warn me in plenty of time." Till the case where all that pre-assessment (we, as humans should always be doing) is beyond the AV's ability, which is inevitable. This is why an L5 AV is going to be elusive, an L4 limited in scope, but something L2-3ish becoming the norm in driver assistance tools. That's really all I expect and would be happy with is competence similar to the NoA/FSD features on the highway for city. Still need to be paying attention, and still will be the driver.

The point of L3 is that the car is capable of driving itself, without human intervention, and at a level of safety better than a human driver, for at least ten seconds after it recognizes a situation it cannot handle. Ten seconds because that's a reasonable amount of time for an awake driver to recognize the "Take Over!" signal, assess what needs to be done, and respond.

If the car cannot do this safely, then it's not a Level 3 car. Whether this can be accomplished without going full-on Level 4 is a separate question. A lot can happen in ten seconds, and maybe a car that can do that is capable of L4 operation. I don't know. Level 3 by definition allows the driver to take her attention away from the road until alerted by the car of a situation that requires her to act.

A car that requires the driver to be constantly alert is Level 2 by definition, even if it allows you to take your hands off the wheel. If you only have 2 or 3 seconds to respond, you'd probably better have your hands on the wheel.

 

[rxlawdude]

The point of L3 is that the car is capable of driving itself, without human intervention, and at a level of safety better than a human driver, for at least ten seconds after it recognizes a situation it cannot handle. Ten seconds because that's a reasonable amount of time for an awake driver to recognize the "Take Over!" signal, assess what needs to be done, and respond.

If the car cannot do this safely, then it's not a Level 3 car. Whether this can be accomplished without going full-on Level 4 is a separate question. A lot can happen in ten seconds, and maybe a car that can do that is capable of L4 operation. I don't know. Level 3 by definition allows the driver to take her attention away from the road until alerted by the car of a situation that requires her to act.

A car that requires the driver to be constantly alert is Level 2 by definition, even if it allows you to take your hands off the wheel. If you only have 2 or 3 seconds to respond, you'd probably better have your hands on the wheel.

Precisely why Level 3 seems to be absurd to me. If you're talking about perceiving and reacting to a stationary object 10 or more seconds away, sure. But dynamically predicting any traffic scenario 10 seconds in advance? Hmm.

But this does give pause as to the computing power necessary to do L4 or L5!

 

[Daniel in SD]

Precisely why Level 3 seems to be absurd to me. If you're talking about perceiving and reacting to a stationary object 10 or more seconds away, sure. But dynamically predicting any traffic scenario 10 seconds in advance? Hmm.

But this does give pause as to the computing power necessary to do L4 or L5!

How much prediction is necessary to drive safely at 60kph on a restricted access highway? I’m not sure.
Practically speaking I think these systems will simply stop in their lane way before the 10 seconds has elapsed so the driver will be taking over a stopped car.

 

[Matias]

Elon offered an update on FSD during the Q2 earnings call just now:

Elon has been highly confident many times regarding FSD

 

[daktari]

This journalist has some more details on Mercedes Drive Pilot L3 and some objections. I don't share those, as they are not relevant for a traffic jam system.

The point is that the system is capable of stopping you safely after 10 seconds without any interventions. You don't need to take over within those 10, only if you wish to keep with the traffic flow. This is a huge difference from a L2 e.g. NoA or VW travel assist.

MB's demo ride was not really suited to show that difference, something I find strange. They demo it as something in between, or in a way many Tesla users with defeat devices use their system in a traffic jam.

Some videos show the additional hw needed; extra lidar in front, an extra rear camera and another GPS antenna or GPS unit on the roof.

Mercedes-Benz' Claims To Have Level 3 Automated Driving But There's Still A Huge Problem

 

[daniel]

Precisely why Level 3 seems to be absurd to me. If you're talking about perceiving and reacting to a stationary object 10 or more seconds away, sure. But dynamically predicting any traffic scenario 10 seconds in advance? Hmm.

But this does give pause as to the computing power necessary to do L4 or L5!

The more I think about it, the more I'm inclined to agree with people who say that if you solve L3, you've solved L4, or nearly so. If the car can do all the driving at better than human-level safety and give the driver a full ten seconds of warning when she's needed to take over, it could probably operate at L4. The only real difference is that an L4 car can find a place to pull off the road and stop safely, which an L3 car does not have to do. So much can happen in ten seconds that for driving tasks, L3 is very nearly L4.

Stopping in the middle of the driving lane is not always a safe option, so being able to pull over safely is the one big difference. If an L3 car gives the driver a full ten seconds warning, and an accident happens after those ten seconds, the car has not broken L3 operation. But in the full range of driving tasks, being able to pull over sounds much easier than everything else an L3 car needs to be able to do.

I am less and less hopeful about being able to buy an L3 or an L4 car in what remains of my lifetime. Maybe L3 highway, but not city. And if it's limited to divided highways, it would be of no use to me. There's just one rather short stretch of divided highway I ever drive on here, and that only about once a month, for maybe 5 minutes. (It's the only road here where autosteer will operate faster than 5 mph over the limit. And at 55 mph, the limit on that short section of road is fast enough that I don't want to go over.)

 

[diplomat33]

The more I think about it, the more I'm inclined to agree with people who say that if you solve L3, you've solved L4, or nearly so. If the car can do all the driving at better than human-level safety and give the driver a full ten seconds of warning when she's needed to take over, it could probably operate at L4. The only real difference is that an L4 car can find a place to pull off the road and stop safely, which an L3 car does not have to do. So much can happen in ten seconds that for driving tasks, L3 is very nearly L4.

I don't think that is necessarily true. I think it would depend on the type of L3 and ODD. For example, a "highway traffic jam" L3 would not be good enough to be city L4. So a "highway traffic jam" L3 would not mean that you've almost solved L4.

Also, L3 can get away with not solving a lot of FSD problems because you can just throw it back to the human driver if needed. For example, L3 does not need to do traffic light detection since you could have the car just request the driver take over every time you approach a traffic light. But you can't do that with L4. L4 needs to solve FSD problems like traffic lights because you can't just pull over to the side of the road for every traffic light.

 

[stopcrazypp]

The more I think about it, the more I'm inclined to agree with people who say that if you solve L3, you've solved L4, or nearly so. If the car can do all the driving at better than human-level safety and give the driver a full ten seconds of warning when she's needed to take over, it could probably operate at L4. The only real difference is that an L4 car can find a place to pull off the road and stop safely, which an L3 car does not have to do. So much can happen in ten seconds that for driving tasks, L3 is very nearly L4.

Stopping in the middle of the driving lane is not always a safe option, so being able to pull over safely is the one big difference. If an L3 car gives the driver a full ten seconds warning, and an accident happens after those ten seconds, the car has not broken L3 operation. But in the full range of driving tasks, being able to pull over sounds much easier than everything else an L3 car needs to be able to do.

I am less and less hopeful about being able to buy an L3 or an L4 car in what remains of my lifetime. Maybe L3 highway, but not city. And if it's limited to divided highways, it would be of no use to me. There's just one rather short stretch of divided highway I ever drive on here, and that only about once a month, for maybe 5 minutes. (It's the only road here where autosteer will operate faster than 5 mph over the limit. And at 55 mph, the limit on that short section of road is fast enough that I don't want to go over.)

Note L4 does not require the car to be able to pull to the side of the road, stopping in its current lane is an acceptable fall back.

The difference is for L4 there is no take over time. The car must be able to handle all situations it may come across and safely fall back even if the driver never responds.

Put in other words if that L3 car fails spectacularly after the buffer time (10 seconds in this case) that is still within its design constraints, but that's not the case for L4. Another way to look at it is that for L4, the request for take over is optional for the driver to respond, but not so for L3. A standard used by journalists is if the "driver" can sleep in the car: you can in L4, you can't in L3.

 

[Daniel in SD]

Put in other words if that L3 car fails spectacularly after the buffer time (10 seconds in this case) that is still within its design constraints

But realistically how could that ever happen unless the system is actually designed to fail?
I'm just curious if anyone can think of a scenario.

Note L4 does not require the car to be able to pull to the side of the road, stopping in its current lane is an acceptable fall back.

My interpretation is that it can stop in the middle of the lane if that's the minimal risk condition but that may not always be the case.

At levels 4 and 5, the ADS is capable of automatically achieving a minimal risk condition when necessary (i.e., due to ODD exit, if applicable, or due to a DDT performance-relevant system failure in the ADS or vehicle). The characteristics of automated achievement of a minimal risk condition at levels 4 and 5 will vary according to the type and extent of the system failure, the ODD (if any) for the ADS feature in question, and the particular operating conditions when the system failure or ODD exit occurs. It may entail automatically bringing the vehicle to a stop within its current travel path, or it may entail a more extensive maneuver designed to remove the vehicle from an active lane of traffic and/or to automatically return the vehicle to a dispatching facility.