Autonomous Car Progress

[MP3Mike]

Just really thick white paint in strips caused enough discomfort that you have to slow down or your filings will pop out of your mouth.

I think you need to find a better dentist.

 

[diplomat33]

Mobileye Keynote by Prof. Shai Shalev-Shwartz at AEK´23:

Some target KPI:s regarding reliability:
View attachment 955520

I watched the keynote several times.

I thought this was very interesting. Shai showed an example of their next gen crowdsourced maps. He said with one single camera, they can generate this 3D map that is cm level accurate.

 

[kabin]

Pretty interesting. Sounds like he is possibly focusing this warning to a large player with level 3+ plans for one integrated unit and not subdividing the design for independence and increased MTBF.

"It is really really insane in my opinion if there is a broad movement in the industry who believe" .... they can integrate and integrate onto a single chip.

 

[stopcrazypp]

Pretty interesting. Sounds like he is possibly focusing this warning to a large player with level 3+ plans for one integrated unit and not subdividing the design for independence and increased MTBF.

"It is really really insane in my opinion if there is a broad movement in the industry who believe" .... they can integrate and integrate onto a single chip.

Which player is he talking about? All the major ones use two chips (Tesla FSD board, Nvidia Drive Hyperion, Qualcolmm Snapdragon Ride). There's no one using only one chip.

 

[kabin]

Which player is he talking about? All the major ones use two chips (Tesla FSD board, Nvidia Drive Hyperion, Qualcolmm Snapdragon Ride). There's no one using only one chip.

I dunno. Maybe it was just a warning to the industry?

 

[diplomat33]

Pretty interesting. Sounds like he is possibly focusing this warning to a large player with level 3+ plans for one integrated unit and not subdividing the design for independence and increased MTBF.

I took it as a criticism of the Tesla FSD approach, specifically Tesla's attempt to achieve "eyes off" with no redundancy. He feels that you need redundancy to get to the target MTBF of 10M hours.

And if you look at this slide in the talk, Shai talks about the challenges of building an "eyes off" capable system. He says it is unrealistic that ML alone can achieve 99.999999%, especially as the data sets get bigger and bigger. He does not think that one single ML only stack for the entire driving task can achieve 99.999999% accuracy. So, I think it was a criticism of Elon's tweet about V12 going to E2E. My interpretation is also based on the fact that Shai immediately touted the Mobileye approach which is diametrically opposed to vison-only E2E. The Mobileye approach separates planning and perception, deliberately does not use ML for planning, and uses a modular and super redundant design in both hardware and software. Shai's view is that you need separate modules and redundancy to achieve the target MTBF of 10M hours.

The reference to "moonshot" could be a dig at Waymo since Shai mentions that they don't believe in going straight to driverless. Waymo of course is going straight to driverless and even referred to the Google Self-Driving Project as a moonshot. I know in other places, Mobileye has taken shots at Waymo, saying the robotaxi first approach won't scale to mass market autonomy. But it could also be a dig at Elon for claiming Tesla will go straight to driverless with their approach. Shai touts the Mobileye approach of being more incremental, doing "hands off" first, then limited "eyes off" and then driverless.

I also thought it was interesting that Shai mentioned the different methods they are using for object detection within their perception stack. He made a point to say that it is not just redundancy in sensors, it is also redundancy inside the perception software. That could also be a dig at vision-only E2E approach.

So the point is that they believe in a modular approach and deep redundancy in both hardware but also software in order to achieve the target MTBF of 10M hours.

 

[diplomat33]

NHTSA will be proposing a new program this Fall called AV STEP that will allow deployment of AVs with no driving controls:

A top federal safety official outlined a new national program Wednesday for greenlighting the deployment of autonomous vehicles that do not have traditional controls like steering wheels or brake pedals.

Companies could potentially deploy large numbers of self-driving vehicles under the proposed "ADS-Equipped Vehicle Safety, Transparency and Evaluation Program," known as AV STEP.

"This is a new and exciting opportunity for all of us," Ann Carlson, the acting administrator of the National Highway Traffic Safety Administration, told an audience of industry executives, transportation officials and academics gathered at the Automated Road Transportation Symposium.

She said NHTSA expects to publish a notice of proposed rulemaking on AV STEP this fall.

While the exact number of vehicles allowed remains to be determined, industry experts expect it to be substantially more than the 2,500-vehicle maximum currently permitted through a process that allows companies to request an exemption from federal motor-vehicle safety standards.

Stuck in slow lane, NHTSA proposes new regulatory path for self-driving vehicles

A top federal safety regulator outlined a new regulatory pathway called AV STEP that could speed the deployment of autonomous vehicles. A notice of proposed rule making is expected this fall.

www.autonews.com

 

[Bladerskb]

This is why I have NO respect for Tesla fans who are pushing this none-sense. They will say or agree with anything to prop up Tesla. Its absolutely disgusting. Substitute Waymo and Cruise for Tesla and they would be in an uproar claiming "Big Oil and Big Auto".

 

[stopcrazypp]

This is why I have NO respect for Tesla fans who are pushing this none-sense. They will say or agree with anything to prop up Tesla. Its absolutely disgusting. Substitute Waymo and Cruise for Tesla and they would be in an uproar claiming "Big Oil and Big Auto".

The video mentions nothing about Tesla, so not sure what you are going on about. Skipping through it, the end mentions GrowSF, which is a group of Tech workers trying to influence politics in SF and having a grudge with certain supervisors (and unsurprisingly they are using this as a chance to trash certain supervisors even though they have little to do with the concerns brought up by SFMTA). Not that hard to research and it's obvious they are heavily politically biased (just as they claim the opposing side to be).

Dean Preston vs. GrowSF is San Francisco's spiciest political feud

In conversations with SFGATE, Supervisor Dean Preston and GrowSF co-founder Steven Buss...

www.sfgate.com

Looking through it more, the only stat they analyze on the actual SFMTA letter is the misleading one on the collisions, but fail to acknowledge the primary problem brought up (namely halting incidents increasing drastically after the switch to driverless).

They mention a tweet by a supervisor about incident count being 18 and focus on one incident counted that was not applicable (forgetting of course other incident stats also include incidents that are not the fault of the subject vehicle!), but that neglects that supervisor tweets are irrelevant to the permit approval (CPUC does not consider them). Also does not acknowledge there are plenty of other incidents that are applicable. Part of the request is more data so that it is it is easier to filter out irrelevant incidents.

The later part of the video is all political stuff and promotion of their group.

 

[stopcrazypp]

Saw Cruise filming a commercial today in SF, overheard them talking about how safe it is etc. I guess they are going to have a media blitz to try to fix their reputation a bit ahead of the CPUC vote.

 

[diplomat33]

Saw Cruise filming a commercial today in SF, overheard them talking about how safe it is etc. I guess they are going to have a media blitz to try to fix their reputation a bit ahead of the CPUC vote.

The media blitz has already started. Cruise ran this full page ad yesterday in the NT Times and other newspapers:

https://twitter.com/x/status/1679517290847694848

 

[spacecoin]

ICRA'23 presentations from Cruise, Waymo, Wayve and others. Very interesting stuff:

 

[diplomat33]

ICRA'23 presentations from Cruise, Waymo, Wayve and others. Very interesting stuff:

Thank you!! I will definitely watch. I always love these academic talks to see all the new stuff companies are doing to try to solve various autonomous driving problems. And it is great to see all the different approaches.

 

[ThomasD]

That orange car in the simulation can't predict if the white car will increase it's speed as the orange car tries to get in front of it. So lets assume the white car needs to get into the right lane and it speeds up at the same time the orange car speeds up to get in front of it. Both cars hit as they try to occupy the same lane who is at fault?

 

[diplomat33]

CPUC will be having a public meeting on July 31 at 10AM PDT to address the incidents of unexpected stops by Waymo and Cruise robotaxis.

You can watch the meeting online:

Join link:
https://cpuc.webex.com/cpuc/j.php?MTID=m7060ecc16079c2e8060695c2
7c0ab29c
Webinar password: 2023 (2023 from phones and video systems)
Webinar number: 2494 018 1373
Join by phone: 1-855-282-6330 United States Toll Free
+1-415-655-0002 United States Toll
Access code: 249 401 81373

Source: cpuc_meeting

Waymo and Cruise reps will be asked the following questions during the meeting:

Driverless AVs Stopping Unexpectedly and First Responder Training
1. How many times has a Cruise or Waymo driverless AV
come to an unexpected stop in San Francisco?
2. What were the reasons for a Cruise or Waymo driverless
AV to come to an unexpected stop in San Francisco?
3. Describe how remote operators interact with AVs in
emergency situations, including actions remote operators
are able to use to provide navigation aid to vehicles to
move vehicles to locations that do not block traffic. If
remote operators do not take control of the vehicle and
perform the dynamic driving task in these situations,
please describe why not, including technical and liability
issues associated with remote control of AVs.14
4. How many of the Cruise or Waymo driverless AV
unexpected stops have impeded a San Francisco first
responder from executing their duties, if any?
5. How were these unexpected stopped driverless AV
situations resolved and how long did they take to be
resolved?
6. Describe the testing protocol used by Cruise or Waymo to
test that its AVs recognize an emergency situation.
7. How many times, if any, has Cruise or Waymo conducted
training sessions for San Francisco first responders in
dealing with unexpected stopped driverless AVs?
8. How many San Francisco first responders, if any, has
Cruise or Waymo trained in dealing with unexpected
stopped driverless AVs?
9. How long does each training last?
10. Does Cruise or Waymo update its training materials for
first responders, and if so, how often?
11. How many training sessions, if any, does Cruise or Waymo
have planned in the future?
12. How do first responders learn about the Cruise or Waymo
training regarding dealing with unexpected stopped
driverless AVs?

Improving Manner and Speed in Resolving Unexpected Stopped Driverless AVs
1. In a circumstance where a first responder is responding to
an emergency (police, fire, medical) and an AV is blocking
the way, how should this situation be resolved and how
quickly?

Updating Passenger Safety Plans
1. The Commission has heard from first responders that from
their perspective the Passenger Safety Plans (PSPs) filed by
the AV companies have gaps when dealing with
emergency and catastrophic situations. For first
responders, please describe specific gaps in AV companies’
Passenger Safety Plans with regard to protecting
passengers and the public during emergency situations.
2. How quickly do first responders need to be able to
communicate with AVs in such emergency and
catastrophic situations?
3. For Cruise and Waymo, please respond to the comments
that we’ve received from first responders on this issue.
4. To the parties, please suggest solutions that can be
implemented to address identified gaps in the PSPs.
5. The Los Angeles Department of Transportation has
recommended in comments that AV companies in the city
of Los Angeles use the Mobility Data Specification now
required for scooters and soon to be required for taxi cabs
to immediately communicate to AV companies real-time
and fluid traffic data such as special events and road
closures. Are tools like this an option for enhancing safety
and improving the ability of AV passenger services to
identify safety hazards to protect passengers?

 

[diplomat33]

I am hoping this whole mess with SF and "stalls" can have a happy ending. If the CPUC meeting is productive and they get the data they need, and SF officials are satisfied their concerns are being addressed, and it spurs Waymo and Cruise to make their robotaxis more reliable, then it could end up being a win-win for everybody. And it occurred to me that these unexpected stops could be one of the last big tech challenges before scaling (there can be non-tech challenges too). Waymo and Cruise have focused a lot on driving safely which was critical to start with. So Waymo and Cruise can drive really well and safely in typical driving conditions. But dense cities do have cases like first responders blocking the road, unexpected road closures, constructions, busy streets with pedestrians, etc... all cases that are causing unexpected stops because they change the drivable space that the robotaxi expects. So once Waymo and Cruise can get super reliable with these cases, they will be more ready to handle dense cities. So it could end up making AVs more intelligent, the very thing they need to scale.

 

[diplomat33]

Cruise software release notes (June):

The latest software release has rolled out to the driverless fleet, containing improvements in these areas:

• Shipped STA v25 that improves prediction of cut-ins by other vehicles by another 73%.
• Shipped TSEL v16 with up to 70% improvements in AVs ability to quickly accelerate, reach, and maintain legal speed limits, especially on higher speed roads.
• Shipped MIR v2 that increases assertiveness and confidence while changing lanes by 16%, and increases comfort during lane changes by 10%.
• Shipped AV Seeds v5 that simplifies vehicle trajectory generation resulting in overall smoother trajectories, especially around sharp turns.
• Improved the vehicle’s ability to accelerate faster on higher speed roads from a stop improving overall safety and comfort as the AV merges into high speed traffic quickly.
• Released a new ML-based model for recognizing the boundaries of construction zones which improves detection by 83%.
• Shipped DLA v5 and KSE v15, which improve speed estimation of high-speed vehicles by up to 90%.
• Shipped RadarRPN v1, which improves detection of long distance & high-speed objects ahead of the AV by +17%.
• Shipped Rigel v1.6, which improves tracking of long distance & high-speed objects in the AV’s lateral field of view by ~7%; for better turns onto busy high-speed roads.
• Improved sensor calibration diagnostics by 60% resulting in fewer trip interruptions.
• Improved reliability of Remote Operators tools when assisting the vehicle reducing the resolution times. Also improved our ability for RA to assist at high speeds.
• Shipped TREX v4, which improves passenger pick up and drop offs and smoother AV turns resulting in better rider comfort.
• As part of LPC v8, improved how the AV considers nearby obstacles, resulting in smoother acceleration and braking at low speeds.
• Enabled right turns onto high-speed roads, significantly improving trip times in suburban areas.
• Enabled right turns from yield signs unlocking additional additional routes and faster travel times.
• Released Android App to allow android users access to the service.

Statistics measuring the percentage of improvement are based on internal benchmarks and metrics.

Source: 06.05.2023 | Software release

The multiple references to "high speed roads" seems to indicate that Cruise is planning to expand their ODD beyond the 20 mph limit, likely 45 mph roads and maybe highways later.

 

[Doggydogworld]

• Improved sensor calibration diagnostics by 60% resulting in fewer trip interruptions.

At first this sounds like a way to reduce stalls, but the stalls people complain about seem to be complex situations, not sensor calibration.

• Improved reliability of Remote Operators tools when assisting the vehicle reducing the resolution times. Also improved our ability for RA to assist at high speeds.

The first part should resolve stalls faster. The second part doesn't make sense to me. If the car is confused to the point that it needs RA help, why the hell is it traveling at high speed? Sounds like guaranteed disaster.

• Enabled right turns onto high-speed roads, significantly improving trip times in suburban areas.
• Enabled right turns from yield signs unlocking additional additional routes and faster travel times.

These limits may explain their occasional bizarre routing.

• Released Android App to allow android users access to the service.

This was always just weird.

The multiple references to "high speed roads" seems to indicate that Cruise is planning to expand their ODD beyond the 20 mph limit and maybe highways soon.

Many videos show 24-25 mph. I've heard some claim 30 mph, but not seen it myself. A leap to highway speeds would be reckless, IMHO. I suspect high speed roads means 40 or 45 mph. Similar to Waymo. That's sufficient for San Francisco, Cruise's main focus. The token services in tiny parts of other cities are just to meet funding milestones, IMHO, not serious commercial efforts.

 

[diplomat33]

At first this sounds like a way to reduce stalls, but the stalls people complain about seem to be complex situations, not sensor calibration.

It likely refers to some sort of sensor issue that is causing the car to stop the trip. So if the sensor calibration is off, the car knows that the perception will be less reliable so the car pulls over and cancels the trip. So this fix is likely designed to improve sensor calibration so that there will be fewer cases of trips being interrupted because the sensors are not calibrated correctly. If I am right, that does not really bode well, if Cruise is having sensor calibration issues that cause trips to be cancelled.

The second part doesn't make sense to me. If the car is confused to the point that it needs RA help, why the hell is it traveling at high speed? Sounds like guaranteed disaster.

I think it is more about unlocking those 45 mph roads. Cruise wants to do the 45 mph roads and they want to be able to rely on RA to help on the 45 mph roads if needed. But I see your point.

Many videos show 24-25 mph. I've heard some claim 30 mph, but not seen it myself. A leap to highway speeds would be reckless, IMHO. I suspect high speed roads means 40 or 45 mph. Similar to Waymo. That's sufficient for San Francisco, Cruise's main focus. The token services in tiny parts of other cities are just to meet funding milestones, IMHO, not serious commercial efforts.

Yeah. I edited my comment. It is likely a reference to 45 mph roads. But highways could come later.

 

[diplomat33]

https://twitter.com/x/status/1681695475836289025