diplomat33
Average guy who loves autonomous vehicles
I agree. That is why Waymo measures their AV performance against the ideal human driver that is always attentive, never impaired, always eyes on the road.
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Autonomous Car Progress
- Thread starter stopcrazypp
- Start date
-
- Tags
- Autonomous Vehicles
I agree. That is why Waymo measures their AV performance against the ideal human driver that is always attentive, never impaired, always eyes on the road.
I'm curious why the default position of most ADAS and AV systems is to just hard brake in the lane. In the video footage, there is no oncoming car. Couldn't the car have swerved to the left to avoid the peds and continued on driving? That's a natural, human behavior to swerve around a potential collision.I disagree. At the moment the kids start to move towards the car, the Cruise brakes. The reaction time was excellent, better than human.
Here is the moment right before the car starts to brake.
A fraction of a second later, we see the kid is still on the edge of the crosswalk but facing toward the car and the car is starting to brake (speed went from 18 to 16). The AV started to brake as soon as the kid started to run in front of the car.
Keep in mind that this was not a standard situation of pedestrians crossing the street normally. The kids wait and jump in front of the car on purpose.
But if your point is that the AV should have anticipated better and braked sooner, Kyle adresses why the Cruise did not brake even sooner:
Well, it was able to brake in time, which would be a reason. Swerving is more difficult, with a greater risk of error.
But a really good reason to brake is that it's almost always going to be sufficient to avoid fault.
Another, less cynical reason is that it's obvious and makes things easier for other parties involved. Nobody else is forced to maneuver. They just need to get stay out of the vehicle's line of travel.
(Human instinct differs. My instinct in such a scenario would be to brake and then decide.)
VW unveiled a fully autonomous EV concept that lets passengers sleep horizontally — take a look at its sharpest features
The Gen.Travel concept has been designed from the "inside out" to explore the experience of passengers traveling in a fully automated vehicle.
www.businessinsider.com
Nothing that's not completely obvious.
VW unveiled a fully autonomous EV concept that lets passengers sleep horizontally — take a look at its sharpest features
The Gen.Travel concept has been designed from the "inside out" to explore the experience of passengers traveling in a fully automated vehicle.
AlanSubie4Life
Efficiency Obsessed Member
As mentioned:
Swerving is actually hard, not a great idea unless required. Slowing down is easy.
It also potentially surprises other road users since you are not where expected and increases collision risk with them.
Right call here initially was to proceed with a lot of caution. Which arguably it did. But rather than an increase from 16mph to 18mph, I think a more natural good human reaction would be to gradually slow to more like 12-14mph, as the vehicle proceeded to the crosswalk.
Note there was no vehicle behind and no crossing vehicles, so Kyle’s response was kind of not relevant - it might not be if someone were following closely, but even then, slowing gradually before a potential sudden stop might reduce rear-end collision risk (a key thing to reduce, which is largely though not entirely under the control of the lead vehicle). There was no risk to taking a more cautious approach here in any case, which would have led to a smoother and less jarring stop when the kids jumped in front of the car.
diplomat33
Average guy who loves autonomous vehicles
Cruise expands to almost all of SF:
Last edited:
Doggydogworld
Active Member
The left side shows their original small service area. They expanded it earlier this year. This latest change adds some areas in the NE part of the city but deleted that big hole in the middle, which he calls Twin Peaks.
I don't know where the 400k miles comes from. Their CPUC report showed 1800 revenue miles total for June, July and August. Plus another 5k+ loitering and en route miles. They obviously do testing above and beyond that, but there's no real reason for testing to be driverless. Maybe their DMV report will shed more light.
diplomat33
Average guy who loves autonomous vehicles
Correct. Cruise did do some expansion earlier this year. The map on the right shows the current state of things as it stands now. The point is that Cruise is now operating driverless ride-hailing at night in almost all of SF.
The 400k driverless miles is likely the total driverless miles (both ride-hailing and testing). Yes, I think some of the testing is driverless. The DMV report should shed more light.
diplomat33
Average guy who loves autonomous vehicles
Speaking of expanding to SF, it looks like Waymo has also expanded their driverless in SF.
This is Waymo's ride-hailing area in SF:
Source:
This is Waymo's ride-hailing area in SF:
Source:
diplomat33
Average guy who loves autonomous vehicles
I highly recommend people read this article. It talks about what goes into the decision to pick a city for a robotaxi ride-hailing service.
Some key highlights:
Some key highlights:
- Waymo’s initial goal was to find a city that made it semi-challenging, but not impossible, to roll out a commercial service. And for those purposes, Phoenix had “the perfect sweet spot of difficulty”.
- Kampshoff said that some of the cities McKinsey has worked with tend to dislike the possibility of empty robotaxis clogging up roadways, so as a result, there may eventually be a “zombie tax” for empty robotaxis on a per-mile basis, Kampshoff added, or even a carpool-based ride-hailing service akin to Lyft Line or UberPool.
- Kampshoff also echoed the idea that complexity can make testing difficult but pointed out that it can also provide a strong opportunity for companies—making high-density cities the best case for profitability but also the most complicated case for safety.
- But whatever cities decide with respect to adopting AVs, Engel said the tech isn’t a silver bullet.
- This complex balancing act—weighing company goals and technical needs versus a city’s goals as well as its physical, regulatory, and social environment—may help explain why robotaxi companies are beginning to find and focus on goldilocks cities, per Kampshoff.
Why some robotaxi companies are looking for their Goldilocks cities
It’s all a financial, technical, social, geographical, and political balance.
www.emergingtechbrew.com
Doggydogworld
Active Member
Or they just mocked up an "aspirational" screenshot for this marketing video. I haven't seen reports or videos of driverless Waymos in the financial district. Absence of proof isn't proof of absence, of course, but it's still evidence.Speaking of expanding to SF, it looks like Waymo has also expanded their driverless in SF.
This is Waymo's ride-hailing area in SF:
Source:
Remember when Waymo launched the SF Trusted Tester program in September 2021? Who here predicted that more than a year later they still wouldn't even have a CPUC permit for non-paying driverless, much less permission to charge fares?
They should rename themselves SlowMo.
diplomat33
Average guy who loves autonomous vehicles
That's not entirely Waymo's fault though. Waymo applied for the permits. Blame the CA regulators for taking so long in granting the permits!
Waymo is not being slow. They are making faster progress on the tech than anyone and they are deploying AVs as safely as possible.
diplomat33
Average guy who loves autonomous vehicles
Curious how they overcame the ban they (or city authorities) made on unprotected left turns, after an accident
a few months ago.
Last edited:
diplomat33
Average guy who loves autonomous vehicles
Likely, the engineers analyzed the software during the accident and know exactly what the perception/prediction/planning stacks did. Based on that information, they likely retrained the software so the car can handle that situation better.
No doubt that's the generalized fix. Can't remember if it was a City-mandated ban or a voluntary act to suspend such
turn attempts. A quick fix would be to geofence that one location, a better one is to exclude all turns of that particular
type (only in San Francisco, at night, less than 35mph, etc). If it was a City-ordained ban though how do they convince regulators that all is OK now? A source code walkthrough? Invitation to simulator comparisons before/after showing the bug? DMV-style demo in the field? Sending over a slick gladhandler sales type to say it's all good now?
I was only positing that Cruise UPLs in SF might be OK now but upon further thought only saw a left turn in the video you linked to at a four-way stop near Golden Gate Park, not quite the same situation.
I know that this is all getting into the weeds, but Tesla can dodge these "open the kimono" bullets in other ways by working thru such while maintaining L2 hands-on-wheel responsibility.
They paused left turns until after a recall and a software update:
GM recalls Cruise autonomous vehicles after left-turn crash
One of its cars was involved in a not-at-fault collision making a left turn in June. The new software should prevent a repeat in the future.
www.autoblog.com
Seems like it is just correspondence based reviews and NHTSA largely trusts the automaker to do the right thing (just like other recalls). They don't do very deep investigations unless the situation really calls for it (applies to recalls in general). Zero indication of things like source code reviews or in person demos, seems like just writing a letter detailing things is what is done.
The city does not have regulatory power to ban the vehicles or things like it doing left turns. This was shown clearly when CPUC granted Cruise's license despite the city opposing.
Last edited:
I can't speak to automotive or California authorities, but for bugs in other regulated software areas (such as medical devices) the development team would:
1) replay the inputs via a simulator to demonstrate they can trigger the bug,
2) fix the bug, then run it through the simulator again to demonstrate it no longer occurs for that input,
3) then audit hardware/code for other errors of a similar style,
4) fuzz testing simulator runs: thousands of runs with values and timing adjusted by a small but random amount to show the fix was generic
The paperwork authorities require typically varies by severity of the issue, and whether the project is under authority of a P.Eng. or not.
Last edited:
Similar threads
- Article
