Daimler tests self-driving truck on German highway | Hacker News Reader

Daimler tests self-driving truck on German highway

news.yahoo.com

49 pointsspking23 comments

I'd be curious to know what progress they've made since I worked on VITA-2 as a Daimler intern as part of PROMETHEUS back in 1992. We also had self-driving on highway with a driver sitting attentive next to a red E-stop button to revert to manual control.

We couldn't get the bus to the highway autonomously, but once there, it could self-drive. (I'm sure it's a lot better now, and the hardware much smaller (we had 3 40U racks of equipment) but I'm curious.)

Well, this sounds really interesting, I was just reading the wikipedia entry on Erns Dickmans https://en.wikipedia.org/wiki/Ernst_Dickmanns this and the project was truly fascinating. What did you do there? What did you end up doing afterwards?

It was amazingly interesting. My degree is in Mechanical Engineering, but I taught myself programming at age 7 on an old TRS-80 and have basically only worked in computing my whole life.

At Daimler, I worked on the computer vision system, doubling the throughput (by adding more processing nodes primarily). IIRC, we went from ~12.5 Hz to about 25Hz. That was fairly straightforward, but I also added a lot of inter-frame aware filtering, created some graphical output changes to help us better understand what the system was doing internally at any given time. That was all in Occam on transputers. I think each card held two nodes and cost about $25K. Each node was about a 486-25.

We had a couple mile stretch of abandoned Autobahn a few miles away from our lab building that we would test on about once a week. (Other teams also needed the vehicle to test their systems.) We also took it to an abandoned airfield outside Munich for longer, higher-speed tests, and demos, and took it to Turin a few times for tests and pan-PROMETHEUS demos and world record runs. On a few of those test sessions, if our system was working well, I was also able to pitch in with other teams on control system algorithms, a radar ranging system, and lane/sign detection vision system work.

(This is all 23+ years old, so forgive some minor discrepancies.) It was a great experience all around! (You also got me to send a LinkedIn invite to my boss there, Dr Andreas Kuehnle, now at Bendix, so thanks for the followup Qs.)

When I came back and graduated, I interviewed at the US Big 3 (back when there were 3, they were big, and US... :) ) and a Ford engineer urged me to "go anyplace other than US auto". He said if I wanted to come back into motorsports or high-perf/speciality street vehicles that I was better off doing something else for 5 years and applying back directly into those units. I wish I had that guy's name and address as it was great advice.

Went to a few startups that failed, went to a motorsports simulation game company that was successful and bought by Sierra-Online, then went to a satellite office of DEShaw (a hedge fund), bounced around in finance, and finally ended up at Vistaprint a few years before we went public in 2005 and I'm still there and crazy happy.

I welcome any followup questions (for HN sake, followups on your former question are more interesting than the latter, but I'll answer either...)

One funny story. Our vision system relied (heavily, not exclusively) on sensing prominent horizontal features symmetric about a common centerline. (Cars and trucks, especially at the time, have a lot of horizontal lines: bumpers, window top/bottom, valence, etc.)

During a test on the Munich airfield, we had had rain overnight, but that had stopped a few hours ago and the demos were going well. As my boss^3 (unsure if it was Herr Dickmanns or not) was on-board watching the demo, standing between my position mid-ship and the two front seats, and as we completed a turn from a taxiway onto a disused runway, I watched in horror as the vision system "locked on" to some prominent horizontal features provided by the uneven drying on the pavement. Almost immediately, this lock turned into a "valid" signal post-filtering and induced a panic stop of the bus. The boss^3 fell forward and it had to be very uncomfortable though didn't seem injurious.

After we came to a stop, the vehicle refused to move as our (my?) vision system was convinced there was a car immediately ahead. It was like a $2MM recalcitrant horse...

That day was not my best demo ever. ;)

Heh, I graduated as a mechatronics engineer(so my education leans mostly towards the mechanical Eng. side) and wound up as a firmware engineer for the auto industry. These days I'm quite staggered, frankly, how inefficient the auto industry is at writing good firmware and would love the chance to chip at something more useful than what I'm doing right now.

I was asking what you wound up doing afterwards because I was personally curious what career trajectory one can expect going from such cool projects.(esp at that time)

It's really curious how this bit of history isn't that well known. I've only just heard about it 3 days ago. Since they have been working at it since the 80's, I would have expected Daimler to have the lead in the sdc game. How brilliant were the people working on it at the time? How come they didn't try to hit the market with what they had? I guess it wasn't feasible in '92, but I see the project running till' at least 2005, so I guess they could have tried putting some commercial applications of it out.

Hardware cost (and size/power) was the big issue. VITA-2 was literally a 9m+ bus and was jammed full of equipment. It surely wasn't a limit of brilliance (but like anything, hard and well-directed work is a crucial ingredient).

I believe that you could directly trace some of the lab work our group did to Distronic (adaptive cruise control) and Parktronic (parking assist) in Mercedes road cars over the years. I doubt any of my personal work can be directly traced to a shipping feature in a production car, but I still learned a ton (and I believe contributed a lot as well).

I'd love to learn more about auto firmware, especially as I've gotten into hardware and electronics a bit more as a hobbyist, but my sense is that the auto industry cares deeply about COGS and that probably drives a lot of their seemingly inefficient decision-making. (Firmware writing is NRE, module and ECU production is COGS and if you sell enough units, COGS dominates.)

I don't know you can extrapolate from anyone's career path, due to the insane amount of randomness in outcomes and choices, but if you're "good with computers", there's probably never been a better time to be in the field, especially since I'd basically do it for free (given what I do for hobbies).

"The grass is always greener" is a very real thing, so don't be in a huge hurry to jump fields/sub-fields/companies assuming that it's just your particular company that has flaws. They all do. ;)

The thing idn't that clear-cut in the COGS vs NRE area. From what I've seen, the firmware for each car model is VERY customized and for each generation they're solving the same problems all over again.

At one time I was writing some firmware and it was very fun, but it felt sort of pointless, since I'm sure someone else should have already solved that problem at least once in the past in the same company. So they keep reinventing the wheel, with each product, with each car.

It was very educational for me, but I'd love that once something is good, let's build on it... make it better, expand it. It could use from more transparency or some (internal, in the least) transition to a kind of opensource model within the company, where the good code survives and evolves and the bad one gets left behind...

This is a comment made at Yahoo news page:

  Driving trucks has been a good paying job for lots of people
 in the past. Between robot trucks and robot burger flippers,
 everything might be cheaper, but still people won't be able
 to afford anything.

We are certainly moving to uncharted waters now.

After all the automation—both mechanical and software affecting service (brain?) and labour (brawn?) industries—would there be need for humans to get things done?

Like all automation (the wheel, the combine, ...) it pushes humans up the value chain. There is no market for coolies any more either, just like there won't be for truck drivers in 15 years.

The thing that is different is this time is that we have run the course on almost all unskilled labor now.

The implication of that is that people need to move up the value chain or perish.

> The implication of that is that people need to move up the value chain or perish.

Nice. So will there be enough jobs in this value chain for all the unemployed truck drivers?

The very clear issue is that not everybody can or wants to be pushed up the value chain and 'perishing' is not an option for large numbers of people in a society without that society itself perishing.

Less and less. When joblessness goes up more, we need to go to shorter work weeks so the benefits are distributed evenly.

The goal should be a future where everybody works a day every week or so (or not at all) and can afford a decent life style.

That is excellent news for road safety! I remember working with companies on lane departure detection systems back in 2007. We have come a long way... Next steps will include to have vehicle convoys where vehicles sync together to drive at same speed.

> That is excellent news for road safety!

For that to be a certainty we need a lot more data than what is available today. This is very very early in the deployment of this kind of technology to make such blanket statements. It may be excellent news for road safety.

Same with many charities as well; they're created for PR & tax purposes, not to serve any actual public good.

Not in this case. The "last mile" of freight distribution currently needs trucks (just imagine a city full of freight train tracks), and if these trucks could operate on their own, the likelihood of truck-bicyclist accidents alone could drop massively - not to mention the massive reduction in freight costs because the most expensive part (the human) is eliminated.

Not to nitpick, but fuel is the most expensive part of trucking. [1] And Level 4 automation ain't happening in long-haul trucking (in the U.S.) for a while [2].

1: http://news.trucker.com/trucker-jobs/equipment-driver-pay-pu...

2: http://news.trucker.com/technology/long-road-autonomous-truc...

> because the most expensive part (the human) is eliminated.

I think the fuel (6 mpg!) and capital costs of the truck are considerably larger than the wages of the man driving it.

Perhaps once labour costs are eliminated you can also improve efficiency by driving more slowly (in a truck-only lane on the freeway). You could also draft just behind other trucks. Finally I suppose you could be whatever shape is most efficient for slipping your cargo through the air without needing to worry about accommodating a pesky human.

With all of the advances in self-driving cars, I wonder why we don't have research into self-driving trains.

Self-driving trains do exist. There are quite a few cities that use driverless trains. https://en.m.wikipedia.org/wiki/Automatic_train_operation

It's a matter of scale. A single train operator hauls thousands of people or thousands of tons of material. Getting rid of that one person hardly gains you everything. On the other hand a truck driver hardly hauls anything, maybe ten or twenty tons of cargo per driver.

https://en.wikipedia.org/wiki/Docklands_Light_Railway

Reply on news.ycombinator.com