PDA

View Full Version : Autonomous Vehicles (new job ?)



samoht
August 3rd, 2017, 02:24 PM
I've been looking to move jobs, and also potentially move out of London. Atsuko and I have always been fond of Cambridge, so I looked at jobs there.

Anyway, I had an interview on Tues at five.ai (http://www.five.ai/), and today they've offered me a job (albeit details such as salary to be discussed).

They're a startup founded last year, and are developing a software and sensor package that can fit any* car, making it autonomous. They're aiming to have autonomous taxis in South London by the end of 2019, as part of a government initiative.

It's not 100%, but I'm hoping everything will work out and I'll be moving to Cambridge shortly. I'm excited, as it's such a cool challenge programming a car to drive itself, but also a bit nervous, as it's way different to what I've done so far, and they will be expecting me to bring in some 'professional software development' practices. Also, the interview was tough, I thought I hadn't done that well afterwards.

Anyway, I'll keep y'all informed of how things work out. Questions and scepticism welcome!

XHawkeye
August 3rd, 2017, 02:52 PM
Good luck.

You don't want to move to Bristol? :p

balki
August 3rd, 2017, 05:24 PM
What do they mean by any car?
What are the actual prerequisites? AT, DBW, ...

MR2 Fan
August 4th, 2017, 06:30 AM
sure, I want to see an autonomous driving yugo

Crazed_Insanity
August 4th, 2017, 08:26 AM
Wow! Sound like a fun and challenging job! Autonomous driving seems to be the next big thing...

On a more negative side, you probably should also give them a tough interview. Try to figure out if you really believe in this company... Because there seems to be a lot of folks currently working on autonomous cars. Why would this one have a higher chance of success? You might take a few months or years on the job to figure that one out... ;)

Anyway, speaking purely regarding autonomous driving tech, I do wonder how things will play out. Probably will be like search engines that only 1 end up monopolizing everything as Google is now... However, if this company is more into after market mods to existing vehicles, then maybe there can be room for multiple players in this field...

Yw-slayer
August 4th, 2017, 09:08 AM
Whatever you decide to do, I only wish the best for you mate! :up:

Of course I would like to say to people in a few years "Oh, that groundbreaking autonomous car? Yeah, my mate works for that company." :D

thesameguy
August 4th, 2017, 09:08 AM
Way to sell out the human race to skynet. Hope it pays well.

samoht
August 4th, 2017, 12:46 PM
'Any' car was badly phrased. Basically, what they said is that they spoke to the major carmakers about integrating, but discovered they'd have a five-year lead time. As they have a 2.5yr goal, they decided to deliver a package of software + add-on hardware that they can fit to ordinary cars. I imagine they'll do a deal with some manufacturer nearer the time.

The 2.5 yr goal is a partially govt-funded initiative, they are aiming to launch a pilot scheme of autonomous taxis, basically, on-demand from your phone, in a limited area of South London.

On the question of why five.ai would be the one to succeed, I agree that there are plenty of other players and success is far from guaranteed. I don't think I know enough to really judge. Their theory is that self-driving in Europe is different from in North America, what with our roundabouts, ancient road network, and different driving habits. Thus although Tesla, Google etc are likely to succeed locally, the race to launch a safe self-driving car in Europe is still wide open. I don't know how true this will turn out to be, but it's a theory.

Apart from that, the head guy Stan Boland has been successful with tech startups before, and they have some top AI boffins from Cambridge, Oxford, etc helping out. They also seem to be in with the government (albeit a government with a potentially short shelf-life).

At the least, they have funding for two years, so if they don't succeed I'll hopefully gain experience that will put me in a good position.

I'm not sure that search is quite the right analogy for autonomous driving; I could imagine multiple companies creating workable systems within a few years, I don't think it naturally becomes a monopoly. Like how there are a handful of mobile phone makers; just because Samsung is #1, I can still buy an HTC.

Yeah, my wife seems concerned I'm hastening the end of humanity with this career move! I'm curious to get an inside view of this AI... I'll let you know if we're in danger ;)


Whatever you decide to do, I only wish the best for you mate! :up:

Cheers, it's a challenge but it feels like a good bet to make.

Kchrpm
August 4th, 2017, 12:51 PM
Like how there are a handful of mobile phone makers; just because Samsung is #1, I can still buy an HTC.
While I agree that, whether the company is an overwhelming success or otherwise, the experience for you will be invaluable, this comparison in particular is an interesting one for various reasons relating to HTC dying in the market.

Anyway, the manufacturers will most likely be making the decisions on vehicle automation supplier choices, and purchasing in bulk on specific deals, so it's probably not that useful to compare to consumer products.

thesameguy
August 4th, 2017, 12:59 PM
All it takes is one substantial tech innovation to guarantee a piece of the market for a long time. Often times these startups end up patenting some key piece of the puzzle, or get a significant edge or lead on a piece and get gobbled up by a larger player. Inventing the secret sauce that lets Apple or Google apply their tech to legacy vehicles would be a massive financial victory. Sometimes you get hosed not being the biggest fish in the pond, some times it's ULTIMATE VICTORY. What's life if not a roll of the dice?

Dicknose
August 4th, 2017, 09:32 PM
Sounds cool and hope the move out of London goes well.

Crazed_Insanity
August 5th, 2017, 10:08 AM
2yrs of experience in that field should still be very valuable and things might turn out better! If wife is on board, might as well go for it!

Anyway, not sure how autonomous driving will play out in the future..., besides tech, govt regulation can be a killer too... So it's good to have that connection...

And using you mobile phone analogy, I'd say traditional carmakers would become like the carriers or maybe the phones too... And although there's no monopoly on the OS yet, but it's pretty much between apple and android, right? If it weren't for Google, Jobs could've cornered the entire market already... So I'd imagine Skynet will eventually become a monopoly! but of course ai cars will probably take a while to develope regionally and merge to form a more formidable streetnet first... Before Skynet... ;)

Mr Wonder
August 6th, 2017, 02:41 AM
Congratulations mate! I'm sure it'll be an interesting challenge, and Cambridge might be a nice change of pace from London. I wish you all the best with it. Though if this ends up being skynet - I told you so!

IMOA
August 6th, 2017, 01:09 PM
Sounds like a brilliant move to me. Chances are they won't make it, well, not in their existing guise anyway but there's a chance that some variant of the company will and even if they don't you'll be getting experience on the cutting edge which will leave you in a very good place professionally. And, I suspect it will be an incredibly rewarding experience being at the pointy end of the next big thing. Awesome stuff, congrats.

Crazed_Insanity
August 7th, 2017, 08:31 AM
Yep, can't argue with IMOA's advise. :up:

thesameguy
October 11th, 2017, 02:37 PM
Here's something I've not yet heard talked about:

http://www.msn.com/en-us/money/markets/driverless-cars-are-giving-engineers-a-fuel-economy-headache/ar-AAthUVo?li=AA4Zjn&ocid=ientp


Some of today’s prototypes for fully autonomous systems consume two to four kilowatts of electricity.

WOW.

Crazed_Insanity
October 11th, 2017, 09:16 PM
Wow indeed. I don't ever recall Elon Musk hinting autonomous driving can significantly impact range significantly like that. I always thought those roof mounted sensor look ugly and create lots of drag..., but the power requirement for the whole system sucking up this much juice is really surprising. Tesla vehicles have no range extenders... wonder if they are having similar issues as claimed by the article.

thesameguy
October 12th, 2017, 08:24 AM
I don't even pretend to understand the actual technology that goes into autonomous vehicles - and I wonder what the balance between hardware and software is. Maybe there is some magic missing algorithm or maybe there is some specialized chip missing from the equation. Still, I am blown away by those power requirements - just the computational power alone described by "several kilowatts" is off the charts even by, say, 2014 or 2015 standards.

neanderthal
October 12th, 2017, 08:51 AM
A lot of continuous real time processing. Lot's of sensors and inputs.

Question is can AI learn and start to drive without doing the language thing where they decide the rules don't matter, efficiency at all costs.

samoht
October 15th, 2017, 02:56 AM
A couple of weeks ago this was being discussed in the office, and I asked which was going to use more battery power, the car's drive motor or the computers. Answer, the latter (!) Albeit this is on our prototype, not a productionised system. Part of the issue is that the drive motor is only used when you're moving, whereas the computers are running full tilt even while you sit at a red light, at least for now.

Neural networks are kinda like lots of huge matrices, so there's lots of parallel computation going on. We have something like a rack of high-end NVidia graphics cards for this, which I think is where most of the power goes. Autonomous Vehicles turn out to have quite a bit in common with videogames.

I'd be fairly confident this won't be a huge issue in the medium term, we'll see some fairly dramatic improvements in efficiency, both in software and in hardware - like how games consoles tend to use less power than a gaming PC with similar performance. It's still quite early days.

Godson
October 15th, 2017, 10:40 AM
The big thing that needs to happen is have an algorithm to allow certain systems turn off when certain parameters are active. Think of VW cheating the emissions game.

Crazed_Insanity
October 15th, 2017, 01:52 PM
Hopefully moore's law will shrink these computers down to cellphone size in a few years...

Dicknose
October 15th, 2017, 03:20 PM
I don’t see this as a big issue.

For commuting there is probably plenty of stops, when you can power down systems, things like “radar”. Don’t need that while stopped, turn it on again just before moving to check it’s clear.
Things will improve, as Billi said, Moore’s law (although it often turns into, same price, more power and features!)

But let’s consider who would want these vehicles and what it would cost them.
I see the main early adopters being richer, busy people. They commute in peak times and want more free time (to relax or maybe work during the trip). They won’t care if the running costs are double a self drive electric car. The benefit to them in time is way more than the extra cost. Especially if they are jumping to their first electric vehicle, they probably still save running costs over what they have now.

Where it will hurt most is the reduced range, if that is what stops people buying the car in the first place. This is likely to be just a perception of a problem, rather than an actual issue for these people.

But in the battle of self-drive-electric v autonomous-electric, I don’t think the extra power requirements of the autonomous systems will make a huge difference, the benefits will be so significant (and radical, game changing) that it would dominate an increase in costs or decrease in range.

Crazed_Insanity
October 15th, 2017, 05:06 PM
Energy cost is not much, just surprised to hear these systems can impact EV ranges that significantly...

I wonder how teslas system sucks up power now...

Colleague of mine still doesn't feel safe letting his model x drive by itself.

My Honda clarity's semi-automnomous system(adaptive cruise and lane keeping) sucks too. Every so often it'd warn me that it couldn't 'see' the lane lines and need me to take over! Adaptive cruise also doesn't inspire confidence but better than lane keeping. If these features were not included in the clarity ev, they're definitely not worth paying extra for. Luckily as far as I can tell, these simple systems didn't cost me extra energy... AC/heat impacts range more significantly.

IMOA
October 15th, 2017, 05:48 PM
Well, since those cars are level 2 only you do need to be paying attention and be ready to take control at a moments notice. Something which some Tesla drivers don't seem to understand.

One comment about Moores law, in many areas of automated vehicles it doesn't apply because there is a significant amount of mechanical parts involved in the automation of the driver. Radar, lidar, pretty much any of the sensors etc are the expensive bits and are not subject to moores law. Sure they'll get cheaper because of increased investment and mass production but we won't see the gains in those spaces at anywhere near the rate we saw it in computing power.

Crazed_Insanity
October 16th, 2017, 08:38 AM
Original article... and also according to samoht, it's not just the financial cost, but the computational power requirements being a huge drag! Computers making calculations, not just sensors, using up more power than the car's engine! That seems pretty crazy... I always thought those crazy roof top sensor will cause a huge aero drag for those cars, but I guess, aero drag is probably nothing compare to computational drag!!!

But supposedly human brain is also the most energy hungry organ compared to any other body parts...

So I guess we're on track for AI.

Dicknose
October 16th, 2017, 01:54 PM
The human brain does use up most of your energy, if you are “at rest”.
Once you are moving, exercising, you muscles easily use up way more energy.

I wouldn’t be surprised if the cars are the same, especially the more powerful vehicles.
While stopped, no power to engines but the computers are still needed.
Less powerful engine commuter style, more stops - overall ration of computer to engines would go up.

Dicknose
March 19th, 2018, 12:01 PM
Seems there has been a fatal incident with a pedestrian and an autonomous Uber.
https://gizmodo.com/uber-self-driving-car-killed-arizona-woman-while-in-au-1823891032

Not a lot of details, it was at night at an intersection. Will be interesting to find out details.
And I’m not trying to be ignoring that this is the death of a person, but for people who don’t know the victim the story is about the car, technology and how this might affect the industry.

Crazed_Insanity
March 19th, 2018, 01:38 PM
Should be an easy investigation with so many sensors/cameras onboard...

Also, what would be the point of having a human driver in these tests? Auto-ubers have ran a red light before. Now killed somebody... all with a human driver sitting inside!

Uber tried to blame the red light running on driver error, turned out it was a computer error.

Now, I'm sure they'll probably also try to blame the driver in order to save their tech?

Uber really should just stick to ferrying people using 'proven' tech. I think they're in over their heads.

In any event, when shit hits the fan, who should ultimately be responsible?

Owner of the vehicle or manufacturer of the vehicle? For a full autonomous vehicle, it only makes sense for the manufacturer to be responsible.

But if somebody converts a Ford or Honda into an autonomous vehicle, then the "converter" should become responsible, not Ford or Honda.

Anyway, whatever legal issues that need to be figured out, it's safe to say that every autonomous accident will be recorded with precision... making investigations much easier.

IMOA
March 19th, 2018, 04:42 PM
The cars have a human driver because they're only level 2 and so the driver needs to be there to be able to take control at a moments notice.

In terms of the laws they're being written atm however Volvo has come out and said that if an accident occurs with one of their vehicles in Autonomous mode that they will take full legal responsibility. That obviously doesn't mean that it's their fault (if the person walks out from behind a bus and there's no chance to stop it wouldn't be their fault for instance), just that they will take responsibility if it is.

Freude am Fahren
March 19th, 2018, 04:45 PM
For all we know right now she stepped right out in front of it, and no human would have even been able to save here.

I've seen some reports she was not in a crosswalk, but also that it happened at an intersection, so who knows what happened until officials release more info.

dodint
March 19th, 2018, 05:51 PM
Keith says NBD, so carry on...

Kchrpm
March 19th, 2018, 06:13 PM
The loss of an innocent life is a big deal.

Except to the people who see the potential to make billions of dollars. Uber and their partners will figure out what they can say caused the accident, "fix" it, they and their competitors will point to the accident and fatality rates of human drivers, and the autonomous market will continue on its path.

Everyone had to know this was going to happen eventually, and they accepted it as part of being in this business.

IMOA
March 19th, 2018, 06:19 PM
It's also a part of increasing safety. A good example I read last week was that at the beginning air bags were set to inflate too quickly and that did cause some deaths. As a result of that we didn't chuck airbags out, we changed the rate in which they inflated, and the net results was tens of thousands of lives same. Interestingly the article itself was about what would happen the first time an autonomous car killed someone and the importance that we didn't abandon the approach but instead learnt from it and used that to make the systems safer.

Kchrpm
March 19th, 2018, 06:53 PM
Yes, but I presume Uber doesn't really care about that. They just want to cut out the need to pay drivers and deal with labor-related lawsuits. I do not think highly of Uber, considering their track record.

IMOA
March 19th, 2018, 06:55 PM
Uber is the customer, it's Volvo that are making all the systems.

Jason
March 20th, 2018, 02:06 AM
It's amazing how much attention this is getting compared to pedestrian deaths at the hands of people driven vehicles

IMOA
March 20th, 2018, 03:15 AM
It's amazing how much attention this is getting compared to pedestrian deaths at the hands of people driven vehicles

Thats because people driven vehicles have been killing pedestrians for 100 years whereas this is the first time a computer driven vehicle has managed that

balki
March 20th, 2018, 04:00 AM
Yes, but I presume Uber doesn't really care about that. They just want to cut out the need to pay drivers and deal with labor-related lawsuits. I do not think highly of Uber, considering their track record.
x2
So much so that they easily overwhelm all the trust/respect I have for Volvo

Crazed_Insanity
March 20th, 2018, 10:24 AM
Pictures from media led me to believe they have Ford Fusion and Chrysler vans in their fleet... so it's not just Volvo, right? Or the pics are just from other developers?

Anyway, I still think Uber's in way over its head. It's like a taxi company all of sudden decide to build their own cars. Why bother? Should focus on their core business first. If you're hugely profitable like google, then perhaps you can try it out...

Further, partnership with Volvo also should've been much lower key... so that when accidents happen, let the media report that it was a Volvo car rather than a uber car!

Anyway, the partnership is definitely good for Volvo because I don't see any media talking about them... they can develope and learn at Uber's expense...

Godson
March 20th, 2018, 12:29 PM
Current story seams the woman may have stepped out in front of the vehicle.

:/

Jason
March 20th, 2018, 02:12 PM
Doesn't matter, media outrage machine (or at least headlines) have already gotten their adclicks

Godson
March 20th, 2018, 02:44 PM
Only matters if you don't correct those mis-informed.

Crazed_Insanity
March 22nd, 2018, 07:53 AM
Release video showed multiple fuck ups.

Vehicle's LiDAR should've picked up the pedestrian even in no light conditions.
Safety driver also wasn't paying attention until moments before the impact.
Pedestrian didn't just step out, but really shouldn't have j-walked in such low light conditions assuming drivers can see her.

In all honesty, even if I were fully alert and in control, I'm not sure if I could avoid hitting that woman. I'd probably brake a bit more(the uber car did not decelerate at all), and perhaps I'd swerve a bit... and hopefully won't lose control and kill myself...

Anyway, this incident is the perfect storm for uber I guess.

Under normal traffic rules, I understand that if I hit a pedestrian in the US, no matter what, it's my fault.

I guess uber could get out of this by blaming it on volvo and the driver who's not paying attention...

Tom Servo
March 22nd, 2018, 01:18 PM
Keep in mind too that the video looks darker than it really was. I am familiar with this road, it's not that dark. If any attempt at all had been made to slow down, the person may well still have been hit by the hit would be less likely to be fatal.

There was a big kerfuffle when CA kept going after Uber for testing their autonomous cars on the streets without any oversight, and after the CA DMV revoked the registrations of Uber's cars when they kept neglecting to comply with the regulations, the AZ governor welcomed Uber with open arms, telling them they could test there all they wanted.

Now someone's dead. If the person had been in a crosswalk, do we think the outcome would have been any different?

IMOA
March 22nd, 2018, 02:24 PM
Video looked very bad to me. I think human eye would have seen the pedestrian much earlier and I can’t understand why the lidar didn’t.

Crazed_Insanity
March 22nd, 2018, 03:04 PM
That's the thing... both LiDAR and driver were not paying attention for one reason or another.

The pedestrian also wasn't paying attention.

Multiple failures resulted in this tragedy.

Wonder what will Uber/Volvo tell us about what the car saw with its LiDAR.

If they conceded that their LiDAR saw nothing or they're not sure why their LiDAR saw nothing, then it's time to shut them down or at least stop them from testing on public roads.

Unless they know what went wrong and have a solution for the fix, it'd be crazy to allow them to continue.

Anyway, maybe Skynet is slowly becoming self aware. Just recently, another guy was killed in the movie theater by the automatic lounge chair. http://www.newsweek.com/man-dies-freak-theater-chair-accident-853827

Machines are beginning to see what they can get away with...

Tom Servo
March 23rd, 2018, 07:36 AM
To the point of the Uber video looking *much* darker than it really is, here's video from the same section of road, the point of the incident is at about 33 seconds in.


https://www.youtube.com/watch?v=CRW0q8i3u6E

Crazed_Insanity
March 23rd, 2018, 08:42 AM
Digital photos/videos often do not truly reflect reality..., to me some of them can appear amazingly bright and clear at night while others can appear super dark and fuzzy lacking sufficient contrast... Safe to say that we can't really rely on them to give us true representation of reality.

We know the driver just wasn't paying attention. So how bright or dark it was doesn't matter there.

We also know the LiDAR didn't see her..., so assuming it was bright enough for the camera as it is shown above, computer still failed to see it with regular camera.

The system is suppose to help us avoid situations just like this one, but it failed for some reason. It'll be nice if they provide us with more info with regard to exactly what the computer saw... or why it failed to see the pedestrian.

Tom Servo
March 23rd, 2018, 09:53 AM
They may still be trying to figure that out, as it's entirely possible that the LIDAR system saw nothing, so there's nothing to show. I just want to kill off that whole idea that "she came out of nowhere" and that there was no way a human driver doing what they were supposed to be doing could be expected to have done any better. This was a failure mostly on Uber's/Volvo's part, including a failure it seems to do an adequate job of choosing and policing the drivers that are supposed to be the backup system in case this exact situation happened.

Crazed_Insanity
March 23rd, 2018, 10:19 AM
I think it's clear that their hiring process is questionable when reporters dug up info that the uber driver had a checkered past...

It's also obvious uber/volve/driver are mostly to blame even when the pedestrian j-walked, possibly in the dark...

Still, I'm not sure which video reflected the actual situation better. Original uber video looks way too dark and your youtube video looks way too bright... almost like we can clearly see even without turning on head lights!

Anyway, like I said, regardless of which video is more realistic, uber/volve shares most of the blame. I do hope they figure out and learn something out of this incident rather than just get shut down...

Tom Servo
March 23rd, 2018, 10:27 AM
I'm willing to bet you can clearly see without headlights. It's a well-lit urban street. People drive around here without their headlights on all the time.

Like I said, I'm pretty familiar with that area, I have family in the Tempe/Scottsdale area. It's very well lit.

Crazed_Insanity
March 23rd, 2018, 10:40 AM
Ok. I'll take your word!

dodint
March 23rd, 2018, 03:32 PM
Having finally watched the video I'll say that walking directly into oncoming traffic with your back to the traffic (she was walking on the traffic side of the bicycle she was pushing, causing her back to face the headlights of the Volvo) at night is not a choice I would have made.

I jaywalk every weekday, in traffic around these exact vehicles. You can't assume anyone is going to stop for you under any circumstance. I wouldn't walk that blindly through a crosswalk let alone on a dedicated two+ lane roadway. My presumption is that if that was a human driver and if they weren't drunk or distracted and if they reacted as a reasonable driver would, she would've been struck but likely survived as the panic stop would've lessened the impact.

You only get one life and it's yours to safeguard, leaving it up to someone else to look out for it is foolhardy.

balki
March 23rd, 2018, 05:54 PM
Keep in mind too that the video looks darker than it really was.
The accident video and yours are like night and day. The pedestrian seemingly popped out of nowhere while your video is like you said; could drive without headlights.

Tom Servo
March 24th, 2018, 03:31 PM
Waymo's cars average 5,600 miles traveled before a human safety driver has to "intervene". Uber's were struggling to get to 13 miles. CA requires that data like that be reported to regulators, AZ does not. Uber also recently started having their safety drivers go solo rather than in pairs.

https://www.nytimes.com/2018/03/23/technology/uber-self-driving-cars-arizona.html

This is exactly what everyone warned about.

Crazed_Insanity
March 26th, 2018, 10:03 AM
13 miles isn't so bad. My Honda Clarity EV is struggling to stay within the lane for 3 miles on the freeway! Thankfully Honda probably didn't spend as much money as uber developing its lane keeping assit... :p

It's weird, Honda also has lane departure warning..., I wonder why it could not warn itself that it's departing from the lane...

Anyway, suffice to say autonomous driving is not easy...

Anyway2, I wonder how is the (new job)? Any updates?

Freude am Fahren
March 26th, 2018, 06:54 PM
Arizona Governor suspends Uber test indefinitely. (https://jalopnik.com/arizonas-governor-orders-uber-to-suspend-self-driving-c-1824097133)

Crazed_Insanity
March 27th, 2018, 11:54 AM
Uber has already suspended testing indefinitely. Good call Governor! ;)

BTW, is this autonomous feature developed by Uber or Volvo? If anything, it should be a joint partnership, right?

Wonder why most of the blame seems to go to Uber. Volve has no say on whether or not to have 1 or 2 humans watching over their cars? Since the cars were struggling to go 13 miles, why Volvo didn't insist on testing more on private roads 1st?

Or perhaps it wasn't a partnership at all? Volvo just sold them the cars?

IMOA
March 27th, 2018, 02:39 PM
I thought it was joint but its all uber. In fact the belief is that uber turned off the standard Volvo stuff which would have prevented this accident so their self driving stuff had clean testing.

Freude am Fahren
March 27th, 2018, 05:07 PM
You wouldn't want Volvo's equipment contaminating your testing.

I thought I read Volvo's stuff would have interviened 1 sec. before impact. So probably still a bad accident, maybe still death.

Dicknose
March 28th, 2018, 08:16 PM
1s can loose a decent amount of speed.
Braking at 1g (which is very hard braking for a decent car with decent tyres) for 1s is about 35km/h (22mph)
That can be enough to make a huge difference to an impact.

Can’t say if it would have saved the person. I’d guess having a bicycle beside you could worsen the impact (put forces thru a smaller area)
Certainly something went seriously wrong for this not to be detected. Even if the system mistook her for something else (not a person) an impact was going to happen that should be avoided.

Freude am Fahren
March 28th, 2018, 08:54 PM
Yeah, I was thinking about that when posting, and came to the conclusion an impact would still happen (haven't seen the speed posted, but figured 35mph minimum). Doesn't take much speed to kill a pedestrian with an SUV, especially if it results in running over them. So maybe still death. Probably bad injuries.


If it were another vehicle, it'd be very different though.

IMOA
March 29th, 2018, 01:28 AM
Volvo stuff would detect way more than than 1 second from impact however I’m told that on that era model that the city breaking is disabled over 50kmh.

balki
March 29th, 2018, 05:27 AM
intervened 1s before impact probably means the stock system would have 'seen' the pedestrian before that and at t-1s it would have done something. 40mph vs 20mph would have made a big difference.
Not sure about Volvo, but the Germans have systems that swerve; 1s is plenty of time to miss the pedestrian (wonder if the bike's rear tire would have registered in the system)

Crazed_Insanity
March 29th, 2018, 10:28 AM
What if the system brakes hard or swerve or whatever evasive action to avoid running over a dog and then resulted in a vehicle crash that end up killing the occupants of the vehicle or caused some other accidents?

That would suck.

I'm not particularly happy with Honda's automatic safety features... while in traffic, I'm approaching the car ahead who is stopped or slowing..., I checked and saw that I'm safe to change lane so that I can just keep going, but if I change my lane too late(get too close to the vehicle in front of me), my Honda Clarity would actually end up trying to brake when I end up in the other lane! It just doesn't know that I'm steering away from the vehicle ahead and thought I'm about to crash! Now that's dangerous because vehicle behind me in the other lane may rear end me!

Anyway, so far this safety feature has caused me more trouble than saved me.

My brother in law has a new Model S and his car has often stopped itself on the road for no apparent reasons. Not sure what the sensors saw...

I suppose it's better to be safe than sorry. Still, until the day we can produce quality computers and software that'll never crash, how can ever we expect to be able to build autonomous computer based cars that'll never crash? Makes me wonder...

balki
March 29th, 2018, 05:43 PM
there was an article in a car magazine (Motor Trend or Road & Track most likely) where the car violently just swerved itself and the guy was at a loss like your friend in the Tesla
Turns out the MT/R&T guy had someone following him and said something to the effect of 'how'd you see that guy and react so fast', so sometimes the systems work as designed when the person in the drivers seat curses out their hi-tech wizardry

Crazed_Insanity
March 29th, 2018, 05:55 PM
Interesting..., it quite possible my bro in law wasn’t very vigilant....

If that’s the case, system should give explanations to the driver what happened when it took over control. Model S does have bunch of cameras so should be able to display what the car avoided on that giant flat screen for the driver.

MR2 Fan
July 26th, 2018, 08:51 PM
I had a thought recently. When driving through certain cities up the east coast and mid-west, I've seen certain places where they have, basically separate express lanes, either for, I believe, special paying passengers, or to add more lanes during rush hours (one way in the morning, then open opposite in the evening?)

Anyway, my thought was, why don't they have autonomous modes for one lane of traffic, just for electric vehicles that's designed to go much faster, and then have the driver take control once they reach their destination. It's much easier for cars that only have a single lane choice to go faster and follow in front/behind.

I think they tested this concept several years ago with a line of Ford Crown Victorias in some kind of setup, but that wasn't exactly the same thing.

If the car reaches the destination area and the driver doesn't respond, then it will pull over to a side area until the driver wakes, up etc.


Yes, there's a lot of reasons why this wouldn't work that well, but just wanted to put it out there.

Crazed_Insanity
July 27th, 2018, 11:00 AM
Yeah, in theory, autonomous highway driving should be simple, but look at that exploding Model X that hit the center divider...

I really don't understand why cars like Teslas can hit the center divider at a fork road. If you don't know which way to turn, then at least slow down and stop at the fork... or just pick one and go. Why go in the middle? Sensor didn't sense that you're about to hit something?

There are also Teslas that drivers accidentally drive into a GYM... confusing the gas pedal with brakes really shouldn't happen with Teslas. If sensors and cameras are already built-in, the car should be able to make a judgement call that the driver is mentally impaired to command the vehicle to drive into buildings!

I can understand it's hard to spot pedestrians or bikers on the road, but hitting center dividers and crash into buildings? Not sure these are oversights or tech challenges...

I think before we go full autonomous, I'd like to see a autonomous car race of some sort. Ideally let these systems drive an Indycar or F-1 car... or be part of the Formula E car or something...

I'd feel better about autonomous cars if I can see them reliably finishing races and perhaps even beat out human drivers.

Prove it on the track 1st before you sell these things...

Dicknose
July 29th, 2018, 02:23 PM
It cost of infrastructure. An extra lane is expensive or takes away from existing lanes.
Especially while these vehicles are the minority it doesn’t make much sense.

But long term, yeah it could work. Faster and more efficient if they slipstream.

Kchrpm
August 6th, 2018, 03:26 AM
The skills you need to drive a race car at the limit are very different (and on a very different scale) from what th average person uses to drive around daily.

Crazed_Insanity
August 6th, 2018, 08:23 AM
On a track or oval, there’d be minimal challenge in terms of navigating changing terrain, so main challenge is stay on track while avoiding collision with walls and other cars while trying to push to the limit in various weather conditions. I think these skill sets can definitely still benefit future AVs.

As to test AV navigation skills, we’ll have to let them do rally races.

When AVs can do all these races better than professional racers, then I’d say they’re somewhat ready. ;)

On the road, there are way more variables such as bikes, and pedestrians, crazy drunk drivers...

Anyway, my point is do you guys think the latest and greatest google ‘driver’ can win any races?

I kinda doubt it. Plus, even if they could theoretically win with their algorithms, these cars will be in such aero disadvantage to be able to win...

Kchrpm
August 6th, 2018, 11:57 AM
1) You're right, navigating the track would be much easier than navigating complicated and variable routes.

2) Avoiding collisions with other drivers on a race track is incredibly difficult. The percentage of humans that can do this effectively and repeatedly is tiny.

3) Driving a race car at the limit is incredibly difficult. The percentage of humans that can do this effectively and repeatedly is tiny.

4) The skills needed to do 2 and 3, while useful in some circumstances, are, for the most part, completely unnecessary in daily driving.

5) Autonomous driving AI development must be focused on the problems it wants to solve, it's not going to learn things you aren't teaching it, so if you're simply teaching it how to best replicate the actions of a race car driver, it is not going to learn daily driving lessons.

6) Suggesting that an AI *must* learn to be able to beat a tiny segment of humans at a tasks almost entirely unnecessary for the task you actually want it to do is a waste of everyone's time and money. Can they try to do it? Sure. Make it a requirement before they're allowed on the road? That would be more dense than saying that every 16 year old has to be able to win a race against professional drivers before they can get their license.

Dicknose
August 6th, 2018, 10:34 PM
The AI has an advantage that it should be able to “10/10ths” without straining its brain and making mistakes. Humans find it hard to be repeatable when we are at our limits.
It should also have faster reactions.

I think it will take a while for the smarts to get there. But once it’s there it will be better because it won’t make human mistakes.

Crazed_Insanity
August 7th, 2018, 08:04 AM
Oh surely we can strain anything... even computers! Further, they could avoid human mistakes, but can they fix computer mistakes? When traveling on a highway, if the computer crashes, we don't have time to stop the car and open the window and then close the window waiting for it to reboot! :D

Driving is incredibly difficult. A lot of seemingly simple human activities can be incredibly difficult for computers. This is why it takes a lot of computing 'POWER' to power these systems at this stage.

In order for AVs to share roads with human drivers, they have to be able to not only track other cars, but be able to predict how other cars might move. Likewise with bikes and pedestrians.

Knowing how much road space you have... and anticipate whether if other 'things' might get in your way... if so, do you have enough space to swerve to avoid contact? Or perhaps you need to slow down in order to have sufficient stopping distance... and in order to safely do those sudden maneuvers, you NEED to know the dynamic limits of the vehicle under various weather/road conditions.

Best way to learn these things should be on the track, not in Arizona. Maybe after they've mastered these things on track, then we can start testing in Arizona...

Now, if all vehicles are AVs and they communicate with one another, then it should be a lot simpler. However, if AVs have to deal with erratic human drivers, they need to learn how to drive more defensively... and they need to know the physical limits of itself when doing emergency avoidance maneuvers.

AV needs to be much better driver than average humans. It wouldn't hurt to make them championship winning caliber. ;)

Kchrpm
August 7th, 2018, 08:16 AM
They have, and will continue, to test AI for autonomous vehicles on track. They're focusing far more on navigating safely than quicky, though. I don't want the AI learning that it can remain full throttle during a lane change at 150 mph, or that it can outbrake a competitor and force them off line to make a pass. I want it learning defensive driving techniques, how to deal with multiple complex inputs coming at it at the same time (rain/snow, an animal crossing the highway while cars of varying speeds surround you, avoiding the blind spots of a large truck, etc). If it can clip an apex but doesn't know what to do with a jaywalking senior citizen, we're screwed.

Crazed_Insanity
August 7th, 2018, 10:03 AM
True and I agree.

Of course I'm not asking AVs to be driven at the edge of tire adhesion at all times, but it needs to know these varying limits and be able to adapt.. and make sure it operates well within these limits... or in case of some odd events such as spotty icy roads or blown tires... it needs to be able to compensate and stabilize the vehicle safely.

Anyway, I just think it'd be nice if they can eventually create a computer version of the Stig... the same set of algorithm, capable of adjusting itself in different vehicles/conditions rather than depending on programmer's input. I think 'he' would also be a very valuable test driver in racing teams! ;)

Kchrpm
August 7th, 2018, 11:02 AM
That would be great! It would also be, by far, the most advanced AI ever created for anything by anyone.

Godson
August 7th, 2018, 11:09 AM
Odd, something like that already exists...

Kchrpm
August 8th, 2018, 08:18 AM
I've never heard of an AI system that can be put into any car on any track and figure out how to drive both to their limits completely on their own.

Godson
August 8th, 2018, 09:13 PM
I can see Keith replied, but when I click the link, nothing shows up.

My comment was going to be along the lines of a blonde joke, but whatevs.

Godson
August 8th, 2018, 09:34 PM
And now it shows up... I dunno

JoshInKC
August 9th, 2018, 06:13 AM
And now it shows up... I dunno

Same thing was happening to me and I assumed it was just my computer. Weird.

Crazed_Insanity
August 9th, 2018, 08:19 AM
Yeah, I saw Kchrpm made a new post, but it wouldn't show up once I clicked on the thread. I thought perhaps Kchrpm deleted the post or something.

Skynet is interfering with our discussions. Probably just want us all dead rather than having to learn to drive.

samoht
September 17th, 2018, 05:48 AM
Tomorrow will make a year I've been at my current company, so I really owe you guys an update :-)

I'm on a team making the software that we use to annotate videos. Basically neural networks need large amounts of training data, which means where you give it both the raw input data and the 'right answers' for that data. It then adjusts the weights so that its outputs best match the right answers, and if you're lucky it'll then generalise to new situations. Think of it like making a line of best fit on a graph, but with lots more dimensions. So the more good training data you have, the better your NN will be - in fact adding more data is often more helpful than anything you can do to the network itself.

We take dashcam video and automatically recreate the 3d shape of the road, then human annotators use the 3d model to align virtual lanes on top of the real ones in our software. Then boxes around other vehicles. The system we have for doing this is quite sophisticated compared to others around, it can be quite productive in cases where it works well. We have applied for a couple of patents on this, and are about to apply for another for some extension stuff I did.

In May my team lead went off sick with stress and subsequently opted to return onto a different group, so I'm now team lead (albeit there are only three of us).

I also had four days' advanced driving course, including a day around Millbrook test track going on the high speed bowl and the skidpan, which was fun. This is so that when I'm driving one of our Mondeos with all the cameras and logos, I don't end up doing something silly and making the company look bad. Or alternatively if the worst happens, at least the company has a good story to tell in terms of training of their safety drivers!

I'm not directly working on the car software, which is in some ways a shame but in other ways we have a lot of leeway to make decisions and get on with things, without the pressure of being on the critical path for the car, so it's quite a good situation.

In March we realised the car stuff wasn't going that well, so we re-focused and cut out a lot of scope and complexity in order to get something working to demo to investors this year. The demos will be early next month, we're asking all sorts of big names to invest in us. If they bite, we'll be in a good place - should find out by next Jan/Feb I think. The demo car seems to be coming together and is showing encouraging signs of starting to work, and the 'investment climate' seems very favourable for big investments into what we're doing - I just hope that it works 'on the day' and that we're seen to have done 'enough' to be investible.

Even if things don't work out, I've definitely enjoyed my year in Cambridge and have no regrets. At the same time, the freedom and autonomy I'm enjoying and the excitement of being part of bringing something like this together means I'm really hoping we will get to continue :)

speedpimp
September 17th, 2018, 02:41 PM
Congrats.

Dicknose
September 17th, 2018, 11:06 PM
Sounds cool!

Crazed_Insanity
September 18th, 2018, 09:49 PM
Good luck man!

Thanks for the update!

Gee, it’s been one year already?

IMOA
September 19th, 2018, 12:05 PM
Awesome stuff. Over the next 5 years our concept of transport will fundamentally change and you’re going to be in the thick of making that happen. Savour every moment :)

stephenb
September 19th, 2018, 03:38 PM
Good to hear things are going well. Fingers crossed for the investment you need to carry on.

stephenb
October 8th, 2018, 03:06 AM
Should help boost the company's profile.
https://www.bbc.co.uk/news/business-45706509

Crazed_Insanity
October 8th, 2018, 09:36 AM
Such a scary and exciting industry!

Anyway, yeah, in order to help the public not fear it's taxi service, I still think it'd be nice to be able to see them in action on race tracks... to have all these competing AI driving the same vehicles... to not only 'line follow', but to be able to take proper evasive actions and adapt to different weather conditions.

Formula AI would be an awesome series! Or some sort of NAISCAR series... :p

samoht
October 8th, 2018, 01:30 PM
Yeah, we certainly seem to have a good PR department if nothing else!

JoshInKC
October 8th, 2018, 05:35 PM
Out of curiosity, if you're allowed to say - what company's LiDAR sensors do you guys use? Or is it a built in-house thing?

samoht
October 8th, 2018, 11:03 PM
Velodyne. I know it's ok to answer, because we had this photo on the title slide of a talk a colleague and I gave recently:
https://pbs.twimg.com/media/DoH8E_NXoAIgmhi.jpg:large
(I'm on the left)

We buy all the sensors, the computers and NVidia graphics cards etc, and buy the cars. What we do in-house is a) assemble all the parts into our platform, and b) everything downstream of the sensors, i.e. interpreting the data, sensor fusion (integrating readings from the different sensors), stereo depth, perception (machine learning), planning, prediction and control. In due course we'll also need to make an app and a platform for dispatching the taxis, unsure if we'll contract that out or make it ourselves.

JoshInKC
October 9th, 2018, 03:32 AM
I was wondering if it was them- I know Velodyne makes real good stuff, and I've heard they do a lot of car-based sensors.
I know them mostly from their UAV-based units, which are really nice.

Crazed_Insanity
October 9th, 2018, 09:48 AM
Tesla's autonomous system has no LiDars..., would you feel safe riding in Teslas in autonomous mode, samoht? ;)

samoht
October 10th, 2018, 02:52 PM
As I understand it Lidar is in some ways a short-cut; it gives you accurate distances to objects within about 50m, so you don't need to rely on the AI to judge the distances of those objects visually; and of course rate of change of distance tells you about closing speeds. Obviously, humans can drive ok without Lidar, so in principle there's nothing impossible about a safe autonomous car without Lidar.

Why are we using Lidar? Because going for full autonomy in a taxi, where there is no driver, only a passenger, is hard, and anything we can do to make it safer / easier is beneficial; and because, unlike Tesla, we're not trying to sell cars but taxi journeys, it's much easier to accommodate the cost. Better a slightly more expensive system that hits the road a year earlier, than a camera-only one that is cheaper but is later to market.

I have no problem with Tesla Autopilot, but it's not truly 'autonomous', rather a driver aid. I imagine if they keep developing they would eventually arrive at a safe cameras-and-radar-only system that is fully autonomous, but it's hard to say how long that will take.

Crazed_Insanity
October 10th, 2018, 05:27 PM
That modelX which crashed into the center divider probably could easily avoid that with a lidar..., but I suppose you’re right that it should be possible to eventually developed the software better...

BTW, does your company’s system behave truly like AI and capable of self learning?

Sorry for being so nosey..., just curious. Please don’t answer anything proprietary! :p

MR2 Fan
October 18th, 2018, 09:45 AM
I was watching a Jvlogger (Japan vlogger) I'm friends with and he was talking about self-driving cars and how he expects that this transition will be like the transition between the horse and buggy to the automobile...meaning in a few years everything will go toward self-driving cars very rapidly.

Again, this is all based on the speculation that we'd be going to fully self-driving vehicles in let's say a decade. I seriously doubt that would happen, but just to play along...

While that's a fascinating concept I think it would be met with a LOT of issues and questions that I figured I'd also discuss here:

- Would self-driving cars promote ride-sharing services more where people just won't need or want to own vehicles?
- Would non self-driving cars be considered a hazard by some governments who want to force older cars off the road and in that case...
- Would more cars be retro-fitted with this technology, or are we going to have an overwhelming need to recycle/destroy old cars?
- Would we need less lanes of traffic and would we start shutting down/re-purposing lanes of traffic that are no longer needed?
- How radically would "car culture" be affected by these changes? Would people no longer consider the car as a status symbol the same way?


As I mentioned before, once more cars are self-driving, I think we'll start seeing much lighter, smaller vehicles that won't require as much safety/luxury built in and I'm still not sure why people disagree with this concept. Once we have self-driving cars and a networked grid (theoretically), our commute times will probably be down to 1/3 of what they are now. If all cars are lighter and on a networked grid, there's much less need for heavy safety equipment, and it's a multiplying factor..the lighter the vehicle, the less damage done if it did hit an object.

Yes, larger vehicles would still be needed for hauling things, but I predict for a lot of people they'll be rented for a short time basis.

Dicknose
October 18th, 2018, 04:48 PM
I agree that self driving will be a fundamental change
I think the two major things
- road trauma decreases
- move to "uber" style, dont own a car (or family has 1), you can just get a ride anywhere for anyone - kids, drunk, never had a license

I dont think it will be 10 years. The cars arent ready yet, it could be 10 years before they start to sell mainstream.
Maybe another 10 years before they are the dominant new vehicle and existing cars start to go. Actually I could see old cars quickly becoming worthless - a 5 year old bmw 3 series - worth nothing. Why? Well its not exciting enough for a drive-myself enthusiast. But too expensive to own and run compared to getting a ride service.

As you mention above - the ride service would mean ownership is not a big deal. Sure you might still have some cool driving car (sportscar or maybe a "classic"). But most people wont be impressed.
Another thing that will happen - look at a shopping mall - how much space is parking? Well that can mostly go. I think we will end up with a set of smaller parking areas where cars wait till needed. Here we also still have issues with surface roads that are "main roads" but have shops on them and parking on the road. We dont have as many strip malls - no enough space. The parking is crazy. Its a road that should be for driving, 3 lanes each way. One lane is parking. Then someone reverse parks and its now 1 lane for a minute. Usually big politics to get the parking removed - the shops and locals want to keep it. But without car owners needing to park, all that is needed is some much smaller drop-off bays.
So I dont think we will lose lanes - maybe just how they are used. I can see "autodrive only" lanes.

There will be some hassles while we are in transition, especially when infrastructure starts to change.

Yeah I think its going to be a huge change to ownership, infrastructure and lifestyle.
Ill probably still be driving when it finally hits. Ill be that old guy still wanting to drive his own car (which will be a manual)

Crazed_Insanity
October 19th, 2018, 09:51 AM
- Would self-driving cars promote ride-sharing services more where people just won't need or want to own vehicles?
Not sure why would anyone want to actually 'buy' a fully AV unless he or she is germaphobic. AV will probably become like planes, trains, buses... be pay per use. Unless you're really against sharing, I don't see why anyone would want to own an AV.

- Would non self-driving cars be considered a hazard by some governments who want to force older cars off the road and in that case...
Probably. However, thankfully cars do tend to break down over the years by themselves. If governments can't ban them, they'll probably at least give incentives to get rid of it... or perhaps just make it not economical to keep...

- Would more cars be retro-fitted with this technology, or are we going to have an overwhelming need to recycle/destroy old cars?
Yeah, this is probably the most likely option. If human error accidents start to cause more and more traffic issues, I'm sure regular cars will either be banned or retrofitted... or people may need to pass much stricter driving tests in order to keep our super licenses! ;)

- Would we need less lanes of traffic and would we start shutting down/re-purposing lanes of traffic that are no longer needed?
Kinda doubt we'll ever make roads narrower... probably can repurpose them to be parking spots for unneeded AVs.

- How radically would "car culture" be affected by these changes? Would people no longer consider the car as a status symbol the same way?
I think that has already shifted. I remembered I can't wait to get my license before turning 16. Nowadays, my nephew has dragged his feet and didn't get his license until later in college!!! I think kids today are not into cars as much already. When cars become fully automated, surely people will care even less. But I suppose status symbol wise... the rich will still be able to hail a ride in a Rolls Royce and the regular folks can only afford to hail rides in a Honda still...

I wonder what will happen to car companies like Ferrari and Porsche. Will consumers like DN be enough to continue to support them... or will they eventually go out of business. Later is more likely I think.

Yw-slayer
October 19th, 2018, 05:12 PM
There should be priority for frequent brand hailers. For example, if I regularly hail a Subaru then when there's a typhoon I'd better get mine as a matter of priority. It'll be like airlines and frequent flier points.

IMOA
October 19th, 2018, 05:21 PM
Pretty much everyone expects car ownership to drop off but interestingly ownership rates in the OECD countries is still increasing (slightly). With the rise of car sharing, ride sharing, urban living etc it's something which no-one can actually explain. That said, I'm quite confident that quite soon car ownership will start to drop away, the only question is how quick that will happen.

Parking is an interesting one and one that I think could lead to some interesting side effects because so much is changing in this space. If cars are autonomous even if they're owned they can park themselves, instantly this means you can put a lot more cars into the same space if all the cars in that space can drive autonomously and of all the autonomous tasks manoeuvring in a controlled space at low speed is one of the easiest, hell, self parking is a pretty common feature in cars now. As things move to ride sharing then everything from building design to town planning will change. Already we're seeing apartment blocks going up with less parking as rather than having a standard 1 or 2 carparks per apartment the builders are assigning parks to car sharing and then only having enough parks for a small percentage of the apartments to option in at purchase. The amount of space we'll assign to parking is going to dramatically reduce.

Traffic and things is an interesting one. What I think will start to happen is that certain lanes/roads/areas will become autonomous only. Once all cars are autonomous and connected things like traffic signals become redundant because it is far more efficient to have the whole thing orchestrated by some big arse traffic control computer. Or lots of little ones, I'm not sure which. (actually, probably both)

I'm not sure about retro-fitting, samoht would probably be best placed to talk to that. I can see retro fitting of telematics/connected car equipment but retro-fitting autonomous capabilities I'm a bit dubious on.

I don't think cars will ever be banned, best analogy is the horses imo. They make no sense from a transport efficiency perspective but there's still lots of people owning, riding, sharing horses and still a lot of sport based around horses. I'll still own a car but in 50 years will I be able to drive that to work? Unlikely, but I'll certainly be able to drive it somewhere whether that be a race track or an area with country roads where crazy people in crashable hand driven cars can still play.

Theres a bunch of other impacts arising from other aspects like rise in electric cars, car sharing etc. For example I wouldn't be investing in a dealership these days, they are dead men walking and are desperately trying to find ways to make money (servicing basically but electric cars takes a lot of that away as well). There's things like whether there will be a rise of coach builders because its so much easier now to design a car, will existing OEMs split into designers and manufacturers (think about an iPhone) and will anyone still do both. If you have your own car and have some sort of blockchain based finance on it if you miss a payment will the car start in the morning or will it drive back to the financier to repossess itself. Or to the auction house because the financier is actually 10,000 individuals spread around the globe.

There are a whole bunch of technologies converging atm and while that won't mean existing cars will suddenly disappear (just have a look at how many fixed line phones still exist. I mean wtf!) that also doesn't mean that there won't be radical and profound shifts in how we transport ourselves and I'm very much in the camp that thinks this will happen faster rather than slower.

Dicknose
October 21st, 2018, 03:01 PM
Traffic and things is an interesting one. What I think will start to happen is that certain lanes/roads/areas will become autonomous only. Once all cars are autonomous and connected things like traffic signals become redundant because it is far more efficient to have the whole thing orchestrated by some big arse traffic control computer. Or lots of little ones, I'm not sure which. (actually, probably both)

Skynet!

I think while there are still humans driving they will keep lights. It could be a long while till all cars/trucks/emergency/military vehicles are auto-drive.
But I could see lots more use of "flashing lights" to say you must slow and give way and the auto-drives just work it out, maybe even communicate to each other.

Traffic jams should be lessened if the cars and the traffic signals can talk. I think there will also be an increase in the smarts of traffic lights. Some trick down of the car technology - if traffic lights have better sensors and can decide to change phase based on waves or gaps in the traffic flow. This could happen before majority auto-drive if the technology becomes cheap and trusted. It would still have basic safety sanity (not green in all directions!). But the timing and/or sequence could make much smarter decisions. So often Im at a red light, no cars coming the other way, then just as the first cars come the lights change and now they wait. If only it changed 10s earlier, it could have finished and gone back for them, much less waiting.