Autonomous Car Progress

[JHCCAZ]

I explained what the definition was in context in at least two posts.

But your point is taken. Let's call it an "unprotected crossing with an unprotected left turn" for the Chuck Cook example.

I'm still waiting for that Waymo video.

Unprotected left turn into high-speed two-way cross traffic. Well prior to any AV considerations, this was a dubious maneuver that may be best avoided by a right and then a (protected or onprotected) U-turn, or some other routing solution as available. Not as an embarrassing surrender, but as a prudent and more predictable alternative.

Everyone should be perfectly happy if Tesla, Waymo, and humans would all be trained this way. Don't mark it as a failure or turn up the ridicule if it's simply avoided by the AV Nav. A significant part of the problem is that it's often not possible to change your mind and legally/safely bail out to the right, once the car is positioned for the attempted left. The mounting pressure to move along then encourages further dangerous (though very common) actions, like stopping at the median and/or turning into a gap with high-speed traffic stil coming in the next lane over.

 

[Dan D.]

There's E Warner Rd
https://goo.gl/maps/fhdxCTCtgsdfhX5o6

S Rural Rd
https://goo.gl/maps/k39KpfqejRANkN9F9

I don't think those are in the service area. But plenty others are. The famous conegate video shows one. I've only seen Waymo turn right in such cases.

Waymo does perform unprotected lefts all the time in AZ, just not the particular type of unprotected left that @rxlawdude wants to see. And he's wrong about the left turn in the SanFran video, which was unprotected by definition.

My marked points are in their area. Are you referencing a different service area than Waymo is using?

 

[Doggydogworld]

My marked points are in their area. Are you referencing a different service area than Waymo is using?

View attachment 699388

You're right. I was reading the map wrong.

Getting back to the definition wars, unprotected left covers all types. You can specify unprotected against oncoming traffic, unprotected against cross-traffic, etc. if you want to be more specific.

 

[rxlawdude]

Unprotected left turn into high-speed two-way cross traffic. Well prior to any AV considerations, this was a dubious maneuver that may be best avoided by a right and then a (protected or onprotected) U-turn, or some other routing solution as available. Not as an embarrassing surrender, but as a prudent and more predictable alternative.

Everyone should be perfectly happy if Tesla, Waymo, and humans would all be trained this way. Don't mark it as a failure or turn up the ridicule if it's simply avoided by the AV Nav. A significant part of the problem is that it's often not possible to change your mind and legally/safely bail out to the right, once the car is positioned for the attempted left. The mounting pressure to move along then encourages further dangerous (though very common) actions, like stopping at the median and/or turning into a gap with high-speed traffic stil coming in the next lane over.

Yes, but currently that throws FSD for a loop.

 

[diplomat33]

Mobileye (ME) is adding a second test car and plans to increase their AV test fleet in NYC to 7 cars by the end of the year. As a result, NYC DOT will have a hearing on Sept 1 about requiring Mobileye to get a city permit to test their AVs in addition to the State permit ME already has. There would be a $5,000/day fine if Mobileye does not get the permit. The NYC permit would require Mobileye to self-certify that the AV is safer than humans, submit a "safety plan", take out $5M accident insurance, provide information about any collisions and must allow DOT to ride in the AVs:

A tech firm that has been quietly testing a single self-driving car on the streets of New York City — which prompted the Department of Transportation to initiate a process to further regulate the testing of such driverless vehicles — is about to deploy a second “look-ma-no-hands” car in Gotham this month, with plans for five more by the end of the year, Streetsblog has learned.

But the prospect of having seven inhuman machines driving around town led the DOT to announce a new regime for testing — one that will kick off with a rules hearing at 10 a.m. on Sept. 1 [Zoom link here, password: 043304]. The new city permit will require companies to self-certify “that their autonomous vehicles will operate more safely than human drivers in New York City,” submit “a safety plan,” and carry $5 million in insurance to “indemnify the city against legal liabilities”; and provide information about any collisions. The DOT must also be allowed to take a ride in the autonomous car as a “demonstration of the autonomous vehicle technology.”

Failure to get the permit will result in $5,000-per-day fines — a pittance to a company like Intel. (At that rate, a year of unpermitted testing would cost just $1.8 million for Intel, which has a market cap of $216 billion.)

The city has said it wanted the additional permit beyond the state permission because New York City streets need to be regulated by New York City’s transportation officials.

“New York City has the right to ensure the safety of its streets — that includes regulating autonomous vehicles that operate here,” said City Hall spokesman Mitchell Schwartz. “We won’t hesitate to take action against companies who could put pedestrians, cyclists, or other drivers in harm’s way.”

Self-Driving Car Company to Test a Second Autonomous Vehicle in NYC - Streetsblog New York City

Wait a minute — there's going to be another one of those things out there? And then five more?!

nyc.streetsblog.org

My thoughts:

First, I think 7 cars is a ridiculously small test fleet.

Second, I do find it interesting that NYC is requiring Mobileye to get an additional test permit for just 7 test cars. The city officials seem to be reacting really strongly about AV testing.

The $5M insurance is not a big deal for Mobileye or Intel. Certainly, the fine is nothing. Mobileye could probably just test their cars without the permit and just pay the fine every day.

 

[diplomat33]

"This is a video showcasing how AutoX fully driverless RoboTaxi deals with the ultimate autonomous driving challenge of road conditions in China’s urban villages. Unique to China, urban villages are ultra high-density residential neighborhoods in major Chinese cities. The public roads in urban villages are extremely busy, chaotic and overcrowded, especially around dinner hours. In the video, the AutoX fully driverless RoboTaxi needed to nudge its way out of dense crowds, bikers, scooters, and pets surrounding the self-driving vehicle."

 

[Doggydogworld]

"This is a video showcasing how AutoX fully driverless RoboTaxi deals with the ultimate autonomous driving challenge of road conditions in China’s urban villages. Unique to China, urban villages are ultra high-density residential neighborhoods in major Chinese cities. The public roads in urban villages are extremely busy, chaotic and overcrowded, especially around dinner hours. In the video, the AutoX fully driverless RoboTaxi needed to nudge its way out of dense crowds, bikers, scooters, and pets surrounding the self-driving vehicle."

Hmm. I didn't watch it all, but it's going very slowly. Also stopping a lot, both with and without reason. I can see it getting rear-ended, as Waymo did early on. Looks like they have a flashing red on back of the vehicle?

Slow the video to 1/4th speed (real time) to get a feel. At 3:10 it goes almost completely into the other lane, looks almost like it's intentionally trying to hit the dog. At 3:43 the oncoming car signals AutoX to move onto the sidewalk so he can pass. AutoX refuses and the other driver has to back up. This whole scene takes two minutes in real time, and only worked because there were no other cars behind him. Another two minute stoppage at 4:40, this time AutoX backs up some and seems to signal the other driver. Then a slightly shorter stoppage immediately afterward. These may be "phone home" situations.

At 9:14 AutoX yields to a pedestrian that was trying to walk behind. This is dangerous as the car in the right lane can't easily see the pedestrian.

It's a very challenging run, and I give full credit for being driverless. But safety isn't that hard if you just inch forward all the time. Think early Waymo in a Costco lot. Hard to judge where AutoX really stands from this. Except that Tesla is ten years ahead, of course

 

[diplomat33]

At 3:10 it goes almost completely into the other lane, looks almost like it's intentionally trying to hit the dog.

Yeah, I saw that. I suspect the AV saw the dog in the right lane and decided to move over to the left lane to go around the dog. But then the dog decided to go into the left lane, causing the AV to change its path again to avoid the dog. It looked like that classic case when two people cross in a hallway and don't know who is going right and who is going left and they both go left and almost bump into each other.

I agree the video is a challenging route because of all the road users. It's why Prediction is so important and so challenging. You want the AV to drive smoothly and avoid jerky "stop and go" movements but that can be hard to do when there are a lot of road users all doing whatever they want. The path might look clear so the AV starts to go but then a pedestrian decides to cross the street, so the AV has to brake again. You also don't want the AV to get paralyzed with indecisiveness. It can be hard to find that middle ground IMO.

Hard to judge where AutoX really stands from this.

Well, based on the 2020, CA DMV report, AutoX had 1 disengagement per 20,367 miles, the 3rd best, behind Waymo and Cruise respectively. But they only did 40,734 miles. They did not do a lot of miles in CA. I suspect most of their miles are in China.

 

[Doggydogworld]

Yeah, I saw that. I suspect the AV saw the dog in the right lane and decided to move over to the left lane to go around the dog. But then the dog decided to go into the left lane, causing the AV to change its path again to avoid the dog.

But the dog made his move two seconds (real time) before the car. Dog was almost to the center line and clearly headed to the other lane when AutoX started to swerve. The proper move was to stay in the lane and slow slightly in case the dog darted back in front of AutoX. You can't be two seconds behind the scene like that. I can't imagine this is a processor speed issue that will magically go away with a faster chip. It feels more fundamental, like they have really bad prediction and added a bunch of lags to prevent excessive stopping/starting and left/right ping-ponging.

Well, based on the 2020, CA DMV report, AutoX had 1 disengagement per 20,367 miles, the 3rd best, behind Waymo and Cruise respectively. But they only did 40,734 miles. They did not do a lot of miles in CA. I suspect most of their miles are in China.

CA really needs to re-do the whole disengagement report concept. Maybe have independent safety drivers test and rate the cars. Self-grading is pretty useless.

 

[Dan D.]

"This is a video showcasing how AutoX fully driverless RoboTaxi deals with the ultimate autonomous driving challenge of road conditions in China’s urban villages. Unique to China, urban villages are ultra high-density residential neighborhoods in major Chinese cities. The public roads in urban villages are extremely busy, chaotic and overcrowded, especially around dinner hours. In the video, the AutoX fully driverless RoboTaxi needed to nudge its way out of dense crowds, bikers, scooters, and pets surrounding the self-driving vehicle."

Some of that driving is a special use case for North America, I would call it "driving after the fireworks finish". Lots of different users everywhere, each somewhat loose on following the rules of the road. I saw how many of the scooters and pedestrians were leaving a car-sized hole ahead of the AutoX, and sometimes the car took the hint and sometimes not, so some bold users would then reclaim the space, and the cycle repeated. The AutoX made its way gradually. It might be more challenging once the users figure out that this type of car will be easy to bully, but AutoX could increase its forcefulness to match.

I liked how even they have to buckle the driver's seatbelt to fool some nanny system, or maybe it's just a look to show "we meant to drive with nobody in the car". They do seem to have the interior lights turned on in the car.

3:50 was fun. I wondered if it was programmed to drive on the sidewalk. It didn't take the oncoming car driver's hand gesture to "go around me". So, no, it wasn't going to drive on the sidewalk to get around. Nor, unfortunately would it reverse to help the driver out - who is trying to overcome the robotaxi's unwillingness to be creative. AutoX continues to nudge forward eventually pushing the oncoming car back down the narrow way. Yes, it solved the problem but if there were other cars behind the oncoming car there might have been an unsolvable stalemate. Gets a B+ for effort, C- for creativity.

At 4:50 it actually does reverse in a similar situation. I wonder if that was a remote intervention? The hazards appear to come on, there is a time edit and the other users appear a bit annoyed, flashing their lights and eventually giving up and driving up high on the sidewalk.

It does need to have very fast reflexes to not hit things, and accomplishes that part well. It's definitely the kind of place where you can't have an entitlement ego problem, there's a lot of give and take.

 

[diplomat33]

But the dog made his move two seconds (real time) before the car. Dog was almost to the center line and clearly headed to the other lane when AutoX started to swerve. The proper move was to stay in the lane and slow slightly in case the dog darted back in front of AutoX. You can't be two seconds behind the scene like that. I can't imagine this is a processor speed issue that will magically go away with a faster chip. It feels more fundamental, like they have really bad prediction and added a bunch of lags to prevent excessive stopping/starting and left/right ping-ponging.

Fair point. It was definitely an odd move by the AV. I agree that AutoX's prediction might not be very good. Hopefully, testing on these congested roads will give them lots of data to improve their prediction and planning.

CA really needs to re-do the whole disengagement report concept. Maybe have independent safety drivers test and rate the cars. Self-grading is pretty useless.

Yeah, I would be in favor of CA rethinking the disengagement report. And, I would be in favor of independent safety drivers that follow a standard reporting mechanism. This would allow us to compare the different companies better. I know that right now regulators rely on AV makers self-certifying their FSD, but I personally think that it would be better if an independent group certified all AVs, according to a standard system, before they are allowed to be deployed on public roads. That way, the public could know that all AVs on the road passed the same safety standard. IMO, that would increase public trust in autonomous driving vehicles.

As I see it, the main problem with the CA DMV report is that the disengagements are not standardized. Companies may use different metrics for when to report a disengagement. For example, we know Waymo only reports disengagements that they deem to be safety related (based on simulation results). Other AV companies maybe just report all disengagements. Furthermore, other than "city" or "highway", we don't know the precise conditions of the disengagement so we don't really know if it was a difficult driving situation or not.

 

[Dan D.]

"This is a video showcasing how AutoX fully driverless RoboTaxi deals with the ultimate autonomous driving challenge of road conditions in China’s urban villages. Unique to China, urban villages are ultra high-density residential neighborhoods in major Chinese cities. The public roads in urban villages are extremely busy, chaotic and overcrowded, especially around dinner hours. In the video, the AutoX fully driverless RoboTaxi needed to nudge its way out of dense crowds, bikers, scooters, and pets surrounding the self-driving vehicle."

It's surprising it didn't stop at the crosswalk. Could its "rules" require that any user on the crosswalk be actually moving and actively going to be in AutoX's way? This would seem to fit with many of the other interactions, not yielding unless necessary to avoid hitting things.

 

[diplomat33]

It's surprising it didn't stop at the crosswalk. Could its "rules" require that any user on the crosswalk be actually moving and actively going to be in AutoX's way? This would seem to fit with many of the other interactions, not yielding unless necessary to avoid hitting things.


View attachment 700434

It's possible.

 

[Dan D.]

Interesting, it does stop at this crosswalk well before the user on the left is anywhere near the car. However this crosswalk has some kind of blue/red flashing light, perhaps you have to stop for those ones? (though nobody else did)

 

[Doggydogworld]

It's surprising it didn't stop at the crosswalk. Could its "rules" require that any user on the crosswalk be actually moving and actively going to be in AutoX's way? This would seem to fit with many of the other interactions, not yielding unless necessary to avoid hitting things.


View attachment 700434

Good catch. The scooter was in fact moving, though partially occluded, as AutoX approached. Scooter stopped when it saw AutoX was not slowing. Car in the other lane was stopped, a strong clue for AutoX to expect someone in the crosswalk, too. I'd call this a fail.

 

[diplomat33]

Interesting, it does stop at this crosswalk well before the user on the left is anywhere near the car. However this crosswalk has some kind of blue/red flashing light, perhaps you have to stop for those ones? (though nobody else did)

View attachment 700447

The pedestrian on the left does start walking across before the car stops. I think the AV just made a determination to yield to the pedestrian to avoid any risk of a collision.

It is also interesting that if we look just a moment earlier, there was a car that passed in front of the pedestrian, momentarily occluding the pedestrian from the AV. And the pedestrian is wearing dark clothes at night. The pedestrian is kind of hard to see in this shot:

Kudos to AutoX's sensors for detecting the pedestrian so well in those conditions and yielding correctly.

 

[Dan D.]

Good catch. The scooter was in fact moving, though partially occluded, as AutoX approached. Scooter stopped when it saw AutoX was not slowing. Car in the other lane was stopped, a strong clue for AutoX to expect someone in the crosswalk, too. I'd call this a fail.

In North America yes it is a fail. I'm supposing (without any evidence to back this up) that this car has been set to bend rules and go with "don't hit things, get there quickly, take some advantage, obey traffic lights". Not sure. It does seem a bit of a Wild West town, people are riding scooters at speed on the sidewalk for example.

For some reason I didn't feel as anxious as I do when I watch FSD Beta videos. It feels like the car can handle itself, and I was more focused on behaviour than would it run into the street furniture.

 

[diplomat33]

Waymo is finally expanding to SF. They have launched their Trusted Testers program, allowing the public to ride in SF:

Our San Francisco Trusted Testers can hail autonomous rides for their everyday needs anywhere they want to go in our initial service area, whether it’s their favorite bakery in the Sunset, or a special picnic spot in Golden Gate park. All rides in the program will have an autonomous specialist on board for now, and our Trusted Testers will also share feedback with us on their ride experience. From using the Waymo One app, to pickup and drop-offs, to the ride itself, we receive valuable feedback from our riders that allows us to refine our product offering as we advance our service. We kicked off this program last week with a select few and are now expanding the program to all interested San Franciscans. We’ll begin with an initial group and welcome more riders in the weeks to come.

https://twitter.com/x/status/1430199174776201216

 

[diplomat33]

Here is the map of Waymo geofenced area in SF:

More info here:

Waymo robo-taxi starts picking up passengers in San Francisco

Robo-taxis aren't just for Arizona anymore.

mashable.com

 

[ZeApelido]

Here is the map of Waymo geofenced area in SF:

More info here:

Waymo robo-taxi starts picking up passengers in San Francisco

Robo-taxis aren't just for Arizona anymore.

mashable.com

Not even usable in the most dense part of the city. Hope that changes soon for them.