Self-driving vehicles.

The Rocketry Forum

Help Support The Rocketry Forum:

This site may earn a commission from merchant affiliate links, including eBay, Amazon, and others.
I’ll also bet my left… dangly thing… that any self driving system that can deal with macropods is a looong way away. Ten years or more.

Volvo spent nearly ten years trying to develop a kangaroo avoidance system in the early 2000s. They gave up. The sheer unpredictability beat them.

Kangaroos and wallabies are completely unpredictable. They will run into the side of your car when you’re driving past at 5 mph.
The idea is that a computer is able to collect more data than a human and if it can process all of it faster than a human, it can be safer. Can a human driver avoid kangaroos and wallabies? If so, programmers will try to make the computer just as competent or more. If a human cannot avoid them, then a computer can't be worse.

Think of self-driving a truckload of different features. They can all work together, or just a few at a time. It's not just "self-driving" or "not self-driving", but a whole spectrum of possibilities and options, and they're still making choices as to what works best for what situation. Just the continuation of brakes, seatbelts, cruise control, ABS, airbags, rear view camera, warnings, lane change, etc. Each feature and situation is carefully worked out.

It's quite possible they haven't yet worked on a situation where a kangaroo and wallaby come in from either side while you take the 3rd exit of a roundabout in a snowstorm while passing a bus. That's why they collect data from users. But when it happens and they work it out, all the cars will be able to handle it, not just one driver.
 
Years ago all elevators had operators who would "drive" the elevator car to the floor you wanted, stopped level with the floor, and opened and closed the doors for folks. Eventually automated elevators were introduced and it caused quite a panic with a lot of people. They predicted that these "driverless" elevators would move when people were stepping through the doors and kill or maim, that the elevator would drop suddenly several floors, that it wouldn't stop level with the floor, or on the correct floor, etc. Folks were scared to use them.

Now you probably would think an elevator operated by a person is less safe than the automatic ones.

Personally I can't wait to have self driving car technology available to me. I had a test ride in a Tesla which was able to drive itself and I was super impressed. These cars can drop you off at the front door of a building and then go find a parking space all by themselves, and then pick you up at the door later when you're ready.

These will eventually (if not already) be MUCH safer than humans and have a much higher safety record. I'm sure issues will still occur but they'll be super rare. I'd be more worried about someone hacking into an overall grid and causing problems than the system itself making a mistake.
 
The biggest problems are not the technical ones, it's the problem of perception. People have a hard time giving up what they perceive as their ability to "control" the situation.
We have a light rail system going through tests in preparation to going into full operation soon. The trains will be driverless, operated through a central control hub. I admit, I had some questions pop up in my mind when I heard that. Old habits die hard.
 
The idea is that a computer is able to collect more data than a human and if it can process all of it faster than a human, it can be safer. Can a human driver avoid kangaroos and wallabies? If so, programmers will try to make the computer just as competent or more. If a human cannot avoid them, then a computer can't be worse.
I think this is a harder problem than you're positing, particularly for wildlife. The combo of the human eye and brain have spent several million years evolving to be extremely good at evaluating a situation and getting a gut feel for which direction an animal is going to go. Sure, that has atrophied a bit in the last 150 years or so, but I can't count the number of times I've gone by a deer (or a person) and thought "he looks like he's going to jump into the road." I didn't have evidence to back that up, just a feeling. Most of the time, I was right. I don't really know why it works, maybe I noticed what direction the animal/person is looking, maybe there's something else. Machines can learn quickly, but we have a pretty big head start. We also have an enormous amount of intuitive processing power between our ears.

The other issue that I keep seeing is about Lidar technology. I use a survey-grade Lidar scanner at work. It's absolutely fantastic, but it takes time to take the scans and link them together. You also trade off scan speed vs. density of scan points. Figuring out how things are moving based on Lidar scan data isn't going to be a trivial project, particularly in real time. It can certainly be done, but it's hard to do it in time to be useful to a moving vehicle. Most likely, there's going to need to be some mix of different sensors for different tasks.

Again, I'm not saying it's impossible or improbable, just that it's hard.
 
I think this isFiguring out how things are moving based on Lidar scan data isn't going to be a trivial project, particularly in real time. It can certainly be done, but it's hard to do it in time to be useful to a moving vehicle. Most likely, there's going to need to be some mix of different sensors for different tasks.

Again, I'm not saying it's impossible or improbable, just that it's hard.
Lidar tech for just doing that is currently in test trials on the roads with autonomous vehicles. The KPI for self driving vehicles is "interventions per 1,000,000 miles". An intervention is when a human has to take control or do an action. The target is in the single digits. Current technology (Tesla et al) is currently in 5 figures.
 
But can they optimize their parking space selection by picking the space that has the least chance of generating door dings? 😄
I'd start with things humans can do. But selecting a wide space between expensive cars and no one around would be a good start.

The biggest problems are not the technical ones, it's the problem of perception. People have a hard time giving up what they perceive as their ability to "control" the situation.
We have a light rail system going through tests in preparation to going into full operation soon. The trains will be driverless, operated through a central control hub. I admit, I had some questions pop up in my mind when I heard that. Old habits die hard.
I think people will be able to select on or off for many years, and after using it often enough, it will become a normal thing.

I think this is a harder problem than you're positing, particularly for wildlife. The combo of the human eye and brain have spent several million years evolving to be extremely good at evaluating a situation and getting a gut feel for which direction an animal is going to go. Sure, that has atrophied a bit in the last 150 years or so, but I can't count the number of times I've gone by a deer (or a person) and thought "he looks like he's going to jump into the road." I didn't have evidence to back that up, just a feeling. Most of the time, I was right. I don't really know why it works, maybe I noticed what direction the animal/person is looking, maybe there's something else. Machines can learn quickly, but we have a pretty big head start. We also have an enormous amount of intuitive processing power between our ears.
I have no doubt a computer can see an animal before a human does, and have a quicker reaction time.
The car could also be escorted by 2 drones probing the road sides and scaring away the kangaroos 🤪 (Woah ... I didn't have coffee!).
Anyway, I never met a kangaroo, but I'm pretty confident the game is not over.

The other issue that I keep seeing is about Lidar technology. I use a survey-grade Lidar scanner at work. It's absolutely fantastic, but it takes time to take the scans and link them together. You also trade off scan speed vs. density of scan points. Figuring out how things are moving based on Lidar scan data isn't going to be a trivial project, particularly in real time. It can certainly be done, but it's hard to do it in time to be useful to a moving vehicle. Most likely, there's going to need to be some mix of different sensors for different tasks.
Lidar is probably a better tech in many ways, at least because it works in darkness as well as in daylight. But it's expensive. Tesla avoids Lidar to keep costs lower, but also because some reason like "the road network is built for human eyes, so we use cameras (same color spectrum as eyes)". So they're pushing camera tech. It'll be interesting to see how far they can go.

Again, I'm not saying it's impossible or improbable, just that it's hard.
I agree. Comparable to other ambitious science and engineering projects.
 
Last edited:
Lidar tech for just doing that is currently in test trials on the roads with autonomous vehicles. The KPI for self driving vehicles is "interventions per 1,000,000 miles". An intervention is when a human has to take control or do an action. The target is in the single digits. Current technology (Tesla et al) is currently in 5 figures.
What's this number for a human driver?

Say of someone was able to intervene to avoid an accident with a human driver, I'm sure it's a similarly high number.

I'm not saying this is a apples to apples comparison, but interesting to think about.

I'm late to this thread, but I've been listening to lots of Lex Fridman lately. 😉
 
Well, I guess the software won't count as a "driver" if the lone occupant wants to use the High Occupancy Vehicle lanes
 
Lidar is probably a better tech in many ways, at least because it works in darkness as well as in daylight. But it's expensive. Tesla avoids Lidar to keep costs lower, but also because some reason like "the road network is built for human eyes, so we use cameras (same color spectrum as eyes)". So they're pushing camera tech. It'll be interesting to see how far they can go.

One of the issues with Lidar is that it's relatively slow. To take an example from my work, our Lidar scanner is reasonably modern. At a reasonably precise setting (~1/4" between points at 20m range, sorry for mixing units), it will take a scan of a 360-degree azimuth (compass direction) and dead vertical to 60 degrees below the horizon in about 3 minutes. Obviously, a car doesn't need that much scan, so you could cut that to maybe 120 degrees of azimuth (60 degrees either side of dead forward) and maybe 15 degrees of vertical range. That cuts your scan time to around 6-10 seconds. Even if you dump the data in real time to the data processing, that's still a fairly long reaction time. It's also *hard* to match up one scan to another, particularly if you have moving things in both scans. The more you cut down your scan region to save processing time, the harder it is. Oh, and the optics need to be clean. After a little driving in winter conditions, it's going to get scunged up by all of the crap that's on the road.

Granted, (as noted above by @jderimig) you can replace or assist the Lidar scanner with cameras. And it definitely works for a lot of situations. I'm just pointing out the technological limits for what we have now and why some of the important driving tasks (like figuring out if a kangaroo/child is about to jump into the road) are very difficult.

What's this number for a human driver?

Say of someone was able to intervene to avoid an accident with a human driver, I'm sure it's a similarly high number.

I'm not saying this is a apples to apples comparison, but interesting to think about.

I'm late to this thread, but I've been listening to lots of Lex Fridman lately. 😉
Current accident numbers for the US are around 40K fatalities/year, 3.4M injuries/year, and 7.9M property-damage-only/year, all over ~3.2T miles driven. That's 11.3M accidents total or a rate of around 3.5/million miles driven. I'm not sure how many of these would count as needing an intervention, but that's a good start. If the Tesla numbers are comparable, Tesla autopilot needs intervention !40% more often than the average driver.
 
One of the issues with Lidar is that it's relatively slow. To take an example from my work, our Lidar scanner is reasonably modern.
The Lidar tech that is being driven around now can generate a point cloud with each pixel having properties of refectance, vector position and vector velocity at several 10's of Hertz. These are in vehicles driving around now.
 
Last edited:
What's this number for a human driver?

Say of someone was able to intervene to avoid an accident with a human driver, I'm sure it's a similarly high number.

I'm not saying this is a apples to apples comparison, but interesting to think about.

I'm late to this thread, but I've been listening to lots of Lex Fridman lately. 😉
No, this the target for any self driving technology. Especially for trucking. That application will need a KPI probably of 0.01 or less intervention per million miles because of the consequences of a non-intervention.

Accident rates not directly comparable to interventions. A human fender bender will not bankrupt a company or an entire industry. We are a country of lawyers, not people.
 
Here are examples of current realtime lidar imaging, position, velocity tech..
 
I wish I could get a simple pick-up truck, manual transmission, no power anything, cruise, tilt & air is about the only thing I would want.
Rubber floor mat, no fancy interior, something you can hose out when needed.
A basic work or farm truck. I know a few farmers in my area that have restored their old pickups for about 1/3 of the cost of a new one.
The don't want all the fancy interiors because it will just get messed up with cow poop, hay, stray, pig poop, etc.
If an auto maker would come out with an inexpensive work truck like this, I bet they would sell more than they would expect.
I know about 20 people right now that would buy one in a 5 mile radius from me, including me.
$30K and more for a work truck to just get all shitty inside and out is just plain ridiculous, without any alternatives from any manufacturer.

Look at the Ford Maverick -- it's based on a car chassis, but it's a 4-door pick up truck. Starts new at $20k. Dunno if they make a rubber interior, but it looks like Ford is going in the direction you're looking for -- basic, simple, cheap.
 
Can a human driver avoid kangaroos and wallabies? If so, programmers will try to make the computer just as competent or more. If a human cannot avoid them, then a computer can't be worse.

Unfortunately, the roadside carnage shows most humans can't avoid wallabies. Even the best of drivers will hit a few in their lifetime. A self-drive system that 'isn't worse' is a complete waste of time and resources, IMHO.

I would really like to see a self-drive system that has an effective macropod avoidance system. I just don't see it happening any time soon.
 
Unfortunately, the roadside carnage shows most humans can't avoid wallabies. Even the best of drivers will hit a few in their lifetime. A self-drive system that 'isn't worse' is a complete waste of time and resources, IMHO.

I would really like to see a self-drive system that has an effective macropod avoidance system. I just don't see it happening any time soon.
It's not just about wallabies and kangaroos!
If FSD isn't better than humans at that, there's still everything else!
At least wait until THEY (software and hardware engineers) run out of ideas before giving up hope!

https://www.huffpost.com/entry/driverless-cars-kangaroos-volvo_n_5957d3e8e4b0da2c7323eaf0
https://spectrum.ieee.org/autonomous-cars-vs-kangaroos-the-long-furry-tail-of-unlikely-events
https://www.pcmag.com/news/volvos-self-driving-cars-confused-by-kangaroos-in-australia
com.cgs.kangroo.robot.car.transform.robot.shooting.games_411509ba-8086-4f7e-8271-72d090b3f651.png

kangaroo-box.gif
 
Last edited:
It's not just about wallabies and kangaroos!

Interesting articles. Thanks for sharing.

No, it isn't just about kangaroos and wallabies. I'm talking from a personal viewpoint about the hazards I encounter daily that I'm pretty sure self-driving cars can't handle. Moose collision avoidance is of no practical use to me on a daily basis, but I still understand its importance.

Having seen a number of cars with destroyed rims, axles and suspension components due to the potholes caused by our floods this year, I'm wondering how systems estimate the depth of a pothole and how to avoid vehicle damage. On a 350 km (217 mile) round trip last weekend I counted no less than 87 sets of roadworks.

The tech may work perfectly for most people. Aussies living outside of cities aren't most people, we're just the best. 🤠
 
Look at the Ford Maverick -- it's based on a car chassis, but it's a 4-door pick up truck. Starts new at $20k. Dunno if they make a rubber interior, but it looks like Ford is going in the direction you're looking for -- basic, simple, cheap.

+2 the Maverick has an old school dash and heat/AC control, no infuriating touch screens
 
A friend of mine has a “self driving” Tesla. It’s pretty cool in a lot of different ways but out in the country where there are zero lines on the road and/or worn out lines from the snow plows scrapping the paint off it does all sorts of rather sudden weird things. Weird enough that if one is not paying very close attention could lead to serious trouble. Not exactly confidence inspiring in those situations.

It is a pretty cool car. Not something I care to own but I do get the attraction.
 
Having seen a number of cars with destroyed rims, axles and suspension components due to the potholes
This would be self-limiting for the self-driving cars. Once they have disabled themselves, they won't be endangering anyone else on the highway.
 
Checked Google.
On the average, there are 6 MILLION accidents in the US per years, resulting in 35000 to 40000 deaths. I did not find info on the number of injuries.

Granted, the number of self driving cars is very small, percentage wise,

But I keep getting the feeling that people are expecting the self driving feature to be perfect with 0 accidents.
I would consider them a success if they were able to significantly reduce these numbers.
Yes, Ideally it would be great to get the numbers to 0, but I don't think that is realistic..........
 
Checked Google.
On the average, there are 6 MILLION accidents in the US per years, resulting in 35000 to 40000 deaths. I did not find info on the number of injuries.

Granted, the number of self driving cars is very small, percentage wise,

But I keep getting the feeling that people are expecting the self driving feature to be perfect with 0 accidents.
I would consider them a success if they were able to significantly reduce these numbers.
Yes, Ideally it would be great to get the numbers to 0, but I don't think that is realistic..........

“In the 4th quarter [2021], we recorded one crash for every 4.31 million miles driven in which drivers were using Autopilot technology (Autosteer and active safety features). For drivers who were not using Autopilot technology (no Autosteer and active safety features), we recorded one crash for every 1.59 million miles driven. By comparison, NHTSA’s most recent data shows that in the United States there is an automobile crash every 484,000 miles.”

https://www.tesla.com/VehicleSafetyReport
 
Last edited:
It's easy to imagine that, eventually, the rate of accidents between two autonomous vehicles should approach zero. But as long as human drivers are coexisting on the same roads, there will always be accidents.
 
Checked Google.
On the average, there are 6 MILLION accidents in the US per years, resulting in 35000 to 40000 deaths. I did not find info on the number of injuries.

Granted, the number of self driving cars is very small, percentage wise,

But I keep getting the feeling that people are expecting the self driving feature to be perfect with 0 accidents.
I would consider them a success if they were able to significantly reduce these numbers.
Yes, Ideally it would be great to get the numbers to 0, but I don't think that is realistic..........
You already see this with battery fires. If a Tesla catches fire in an accident, it’s Big News. On the other hand, you can pass a burning Chevelle every five miles and nobody bats an eyelash.
 
That's the first thing I turned off in my 2022 Subaru. Soooo annoying, beeping all the time, like having the mother in law nagging from the front seat.

The front collision system, on the other hand, has proved its worth many times in SF Bay Area traffic.
Lots of truck drivers turned the lane warning systems off in their trucks (those that had them) years ago, as they were annoying as crap (and still are), we also had a blind spot radar system and one that tracked off the shoulder objects...12 years ago...and they were hated then. Yes the devices have come a long way but...are still annoying, I do like the collision avoidance auto brake feature though, that can save lives more than any single feature (MY OPINION).
 
You already see this with battery fires. If a Tesla catches fire in an accident, it’s Big News. On the other hand, you can pass a burning Chevelle every five miles and nobody bats an eyelash.
Chevelle? Someone is really dating themselves here.
 
Chevelle? Someone is really dating themselves here.
OK, that assumes that there are enough Chevelles remaining in the world to have one burning by the side of the road every five miles, but you get my point. I was barely out of diapers when the last Chevelle rolled off the line, so they've always been old to me. :D
 
Back
Top