This thread is just a stub to continue discussion on autonomous car progress by other manufacturers and avoid the continual hijacking of other useful threads to discuss this subject.
-
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
Just some recent news from this month to get things going.
Voyage is starting their self driving car trial:
Voyage’s first self-driving car deployment – Voyage
Details of recent Cruise accidents last month:
GM's self-driving cars involved in six accidents in September
New Ford CEO comments on how he views self driving cars (in same article it mentions new partnership with Lyft and also a recent $1 billion investment Argo AI, a new startup that just launched February):
https://jalopnik.com/new-ford-ceo-i-wouldnt-get-into-a-fully-autonomous-car-1819058015
Voyage is starting their self driving car trial:
Voyage’s first self-driving car deployment – Voyage
Details of recent Cruise accidents last month:
GM's self-driving cars involved in six accidents in September
New Ford CEO comments on how he views self driving cars (in same article it mentions new partnership with Lyft and also a recent $1 billion investment Argo AI, a new startup that just launched February):
https://jalopnik.com/new-ford-ceo-i-wouldnt-get-into-a-fully-autonomous-car-1819058015
Last edited:
lunitiks
Cool James & Black Teacher
on this one point I agree.
Here's a picture of the zFAS architecture. You are correct that sensor fusion occurs on other chips, not on EyeQ3:
No at-fault Bolt AV incidents. Waymo and Uber moved most their testing to AZ where the roads are simpler and you don't have to report accidents or disengagements. GM increased their SF fleet.
6 accidents in the same month does seem like a lot though. I don't remember there being that many for other companies (even ones operating in the Bay Area). However, you are right that they increased their fleets drastically, so that probably has to do with it. I just saw another one yesterday.
If you read the accident reports, they are nearly all minor rear end collisions, some with no damage at all. These are very common incidents in tight urban areas and congested freeways. Perhaps hundreds a day occur in the Bay Area.
If I recall correctly it a little on the high side statistically speaking.
What might account for that is the autonomous car is either stopping or slowing down where human drivers don't expect it to. So even if the autonomous car isn't at fault it still potentially could have induced the human driver into crashing.
Like my moms driving most definitely induces people to doing crazy things to get around her. So I love the idea of getting her into an autonomous car, but it's not going to do any good if it's as slow and annoying as she is.
Hopefully human drivers and autonomous cars will learn to coexist, but I'm worried that the human will just end up bullying the autonomous car. Where the autonomous car is going to find itself trapped as human drivers realize that it won't disobey its programming. It's not like a human driver that will get restless and angry.
Maybe we'll have anti-bullying driving laws.
It's hard to tell. I've had 6 rear hits that I can remember that caused no damage to my car/truck, so they were not reported. I don't drive like an obstacle, so it was just inattention that caused it. However, this is more than 1,000,000 miles. Way more.
In any case none of the documented AV incidents by any company were AV-at-fault. 2? were Human at Fault while driving an AV in normal mode.
An AV driving in Normal in traffic, and the driver was inattentive and hit the car in front of them?
An AV driving in AV mode was going to hit debris. The driver took over, and accidently hit another car?
Surprised there hasn't been more discussion on the Semi.
Exclusive: Tesla developing self-driving tech for semi-truck, wants to test in Nevada
Will we see Autonomy demo'd at the Semi launch? Will FSD for MS & MX be the "one more thing"?
Unlikely, I know, but it is pretty certain that Tesla wouldn't apply for a permit to test something if they didn't have something to test.
Exclusive: Tesla developing self-driving tech for semi-truck, wants to test in Nevada
Will we see Autonomy demo'd at the Semi launch? Will FSD for MS & MX be the "one more thing"?
Unlikely, I know, but it is pretty certain that Tesla wouldn't apply for a permit to test something if they didn't have something to test.
There's no way that Tesla will just one day 'come out with' FSD; it'll be a slow and incremental trickle of features over the coming years. It'll have to be. I'm fairly convinced you can do it with just cameras and a map, at least from a technical perspective (but not from a regulatory perspective).
There hasn't actually been a _new_ feature for autopilot for a couple of years now. Since 2016, they've just been re-releasing features they originally had working in 2015, just on a new platform (and not quite as good).
I just hope that we start to see some more features for EAP start to trickle down soon. Personally, I think any new semi-autonomous EAP or ADAS features are entirely dependent on Tesla's fixed-point landmark "Atlas" map maturing and being deployed.
Thinking it through: Tesla started on their Atlas system back in 2015, or so they claimed. I would imagine they had to throw away all of that original harvested data. Since it was generated on Mobileye's vision platform, it's probably only works properly when paired with Mobileye's vision platform to do the same landmark detection required for localisation. AP2 "sees" the world differently, and thus would require its own map to localise within.
We have to remember that Tesla's goal is to create a pure vision + maps system only. The "FSD" video we all watched last year was done using only "passive optical and GPS". No radar, no lidar etc. It's highly likely it was using Atlas to navigate those roads - it would be relatively easy to build rails for the testing zone only; you could even do that manually, without the automation and huge datacenter tools you'd need to have a continuously updated crowd-sourced "roadbook", which Mobileye have been actively working on for years. It wouldn't be that crazy to map a few cross-country routes to enable the cross country demo.
I think Intel actually got an amazing deal; if Mobileye hold the keys to a realtime map of the world within 3cm accuracy, then $5bn was a bargain.
There hasn't actually been a _new_ feature for autopilot for a couple of years now. Since 2016, they've just been re-releasing features they originally had working in 2015, just on a new platform (and not quite as good).
I just hope that we start to see some more features for EAP start to trickle down soon. Personally, I think any new semi-autonomous EAP or ADAS features are entirely dependent on Tesla's fixed-point landmark "Atlas" map maturing and being deployed.
Thinking it through: Tesla started on their Atlas system back in 2015, or so they claimed. I would imagine they had to throw away all of that original harvested data. Since it was generated on Mobileye's vision platform, it's probably only works properly when paired with Mobileye's vision platform to do the same landmark detection required for localisation. AP2 "sees" the world differently, and thus would require its own map to localise within.
We have to remember that Tesla's goal is to create a pure vision + maps system only. The "FSD" video we all watched last year was done using only "passive optical and GPS". No radar, no lidar etc. It's highly likely it was using Atlas to navigate those roads - it would be relatively easy to build rails for the testing zone only; you could even do that manually, without the automation and huge datacenter tools you'd need to have a continuously updated crowd-sourced "roadbook", which Mobileye have been actively working on for years. It wouldn't be that crazy to map a few cross-country routes to enable the cross country demo.
I think Intel actually got an amazing deal; if Mobileye hold the keys to a realtime map of the world within 3cm accuracy, then $5bn was a bargain.
Actually Tesla claimed their lane-by-lane mapping was GPS generated - not visual. So theoretically they should still be able to use it - if they ever did. To summarize - in theory you could take samples of GPS position reports, average out the reports (to account for the sampling errors) and arrive at a lane by lane map. Again - we do not know if they ever used it - but I used AP1 from the very beginning until now and it has significantly improved its accuracy and reliability on poorly marked freeway lanes in Southern California during the 2 years since it was released in Oct 2015. Whether it's mapping or something else they used - they did something. You'd think if it was GPS they could use it again - but I believe they are going for even higher precision maps this time - perhaps made using the cameras - mobileye is building a crowd-sourced vision based map using landmarks like billboards as the markers to triangulate position.
To add to Waymo:
Waymo deploys 500 self-driving Pacifica hybrid minivans in Phoenix for rides open to public
Waymo’s self-driving minivans are now offering rides to real people in Arizona
application at Early Rider Program – Waymo
To go back in time,
Testing of Autonomous Vehicles has a list of companies/entities issued Autonomous Vehicle Testing Permits by California DMV.
(CA) disengagement reports so far at:
Autonomous Vehicle Disengagement Reports 2015
Autonomous Vehicle Disengagement Reports 2016
When I was at an EVent (2018 Nissan Leaf preview) at Nissan's Sunnyvale research office on Sept 14, 2017, while we were all waiting in line to get into the building, I saw at least 2 or 3 Baidu test vehicles (they had lidar on top) go by. Baidu has an office right across the street from Nissan research: Google Maps. Both were very odd experiences: waiting in a long line outside a Nissan corporate building to get in and watching that many self-driving car test vehicles pass by in such a short time.
Last edited:
lunitiks
Cool James & Black Teacher
Maybe...
What's even more likely though, is that they're developing an inertial navigation system (source: Tesla job listing), which uses vairous algos to calculate the overall most likely lat/long based on accumulated readings from GPS, odometer and IMU. Let's not forget they have the uBlox's M8L installed in every single AP2 car out there. (I'm sure the former LEA-6R-0-001 came in handy before AP2 too.)
BTW @verygreen do you have a pic of the "Localization" settings?
GM on the move, proving yet again why they are one of the front-runners in self driving tech.
How we’re solving the LIDAR problem – Kyle Vogt – Medium
How we’re solving the LIDAR problem – Kyle Vogt – Medium
Last report concerning Waymo's Pacificas indicates 50 vehicles are in use. There are plans to expand to 600, but as of Sept, they are still just plans. Fully driverless cars could be months away
GM/Cruise ramped up their program in California in 2017. They are also operating in AZ and Michigan, but those numbers are confidential. There are 100 in California from what I understand. GM's Cruise Automation Doubles Test Fleet, Reports More Collisions - HybridCars.com
That is an excellent blog post just published last week by GM’s Cruise Automation that includes interesting data tables showing the frequency of different challenging situations that Cruise’s autonomous Bolt EVs encounter when testing in urban SF streets versus boring suburban Phoenix streets.
Also: fun videos showing autonomous navigation through a 6-way intersection with broken (all directions flashing) traffic lights and also through a construction zone with humans directing traffic.
I actually just stumbled over it myself and was about to post it to this thread but McRat beat me to it by a few minutes.
Similar threads
