6 ms·

There is no such thing as Level-3 Autonomy, the car can either drive itself safely or not. A slow moving lane of cars is a dangerous environment where things c

by yholio 5y ago

There is no such thing as Level-3 Autonomy, the car can either drive itself safely or not. A slow moving lane of cars is a dangerous environment where things can turn catastrophic in an instant, with zero time to react for a human driver reading their news. Bikes and motorcycles come out of nowhere and pass you between the lanes. Pedestrians and children will attempt to cross the road, encouraged by the apparent slow speed. You will not see them since your view is blocked by close large vehicles. The same for essential signs and signaling. Slow moving can turn to fast moving intermittently and frustrated drivers will unexpectedly go full throttle to cut in front of you. You might find yourself suddenly stranded in the middle of an intersection with lateral trafic. Trafic is slow because your lane is blocked by construction, which requires carefully observing the human readable cues of the drivers in the other lanes, which might yield to you or not. All manner of bad things can and will happen with a scary probability. If an automated car can handle all that safely without reverting to human assistance, what exactly can't it handle, and why is it not full self driving? The 3 and 4 SAE levels were always a dangerous fiction, they imply you can bring to market uncomplete products that take some part of the human effort. This is simply not how the minds of human drivers works: if drivers lose focus to the road they can't regain it instantly in a fraction of a second when prompted by the failing AI algorithm. A half- self driving car is a defective product that is bound to kill someone, regulators should reject any products short of real self driving.

babyshake 5y ago
SAE vehicle autonomy levels are mostly a fiction in terms of not being highly contextual. The practical approach for self-driving is to take an actuarial, data-driven approach. For example, on certain streets or neighborhoods at certain times of day, you can be sure with a 99% confidence level that you will not encounter certain types of situations. A Level 3 vehicle in Phoenix is not necessarily a Level 3 vehicle in Manhattan or Nairobi.
ghaff 5y ago
I generally disagree with respect to L4. A car that is fully autonomous under a precisely defined set of circumstances, especially certain types of roads and possibly under certain environmental conditions could still be extremely useful--assuming you can fully hand off control. A lot of driving is fairly predictable highway driving even if it doesn't help people who want a door-to-door robo-taxi.
- yholio 5y ago
  If you can geofence those specific circumstance and guarantee they won't happen on useful stretches of road, then congratulations, you have a car that reaches level 5 on certain roads. But that's not what level 4 generally implies, they call it "High Driving Automation" and the expectation is that the driver is on hand to correct the AI in certain limited scenarios.
  - ghaff 5y ago
    This is how JD Power describes it: Referred to as high-driving automation, Level 4 autonomy does not require any human interaction in the vehicle’s operation because it is programmed to stop itself in the event of system failure. Since a human driver is never needed, a Level 4 vehicle may not have a steering wheel and pedals. And yes, at Level 4, you can take a nap while riding in the vehicle. Level 4 driving automation technology is for use in driverless taxis and public transportation services. Such vehicles will be programmed to travel between Point A and Point B and restricted to specific geographic boundaries by geofencing technology. Certain conditions may limit or cancel Level 4 autonomous vehicle operation, such as severe weather. I admit that I thought of L4 as being fully autonomous under specific pre-defined conditions (given that's what would actually be useful and that's probably why I've heard it described that way.) The key would seem to be that the vehicle won't try to hand off control unexpectedly given, in part, that a vehicle can't always stop safely.
- nradov 5y ago
  Highway driving stops being predictable when a deer steps onto the road.
  - ghaff 5y ago
    True. But that's actually the sort of thing I'd expect an autonomous driving system to likely be able to handle as well as a person.
azinman2 5y ago
Exactly. Level 5 or bust. It’s fiction to believe anyone can quickly take over — it is not how our brains work. If anything less than Level 5 does a good job much of the time, that only reinforces to the driver that they can completely disengage. It takes many seconds to not only respond to a call that they need to put down their phone and look up, but also to then grok the scene and keep moving. It’s just not cognitively realistic.
skohan 5y ago
Couldn't you imagine a world where infrastructure is adapted for safe self-driving? For instance, what if you had roads (think highways) where self-driving was mandatory? By applying certain constraints, you might be able to create situations where autonomous vehicles are practical. Maybe it's something like a carpool lane. Autonomous vehicles get the far left lane, and are allowed to drive much faster and closer together than manually controlled vehicles.
- yholio 5y ago
  I can imagine it for the sake of argument, but without strong govt action it has a bootstrap problem: no one will create such a lane if there are no such cars, and no one will buy and develop them if there are no roads you can use them on. Moreover, if the goal is to achieve autonomy with relatively dumb algorithms that fall short of full self driving, you have to do grade separation, fencing, tunnels or overground etc., otherwise you still need full self driving to handle those case where bikes or children roam around in your "dedicated" lane, cars and garbage bins are parked on it, drunk drivers suddenly invade it etc. And if you go the full trouble of dedicated and grade separated infrastructure, why not build a tramway or PRT (like the Booring Co Loop system) that any citizen can benefit from? All in all, I can't see the scenario you are describing happening in any real city with real politics, maybe some oil prince can decree it.
  - skohan 5y ago
    I mean if hyperloop can exist, I wouldn't rule it out completely. One path I could imagine: a city like Phoenix parners with autonomous vehicle manufacturers to create a bubble where self driving cars can exist. It goes really well, other people want it, and other cities fall suit. I'm not saying it's likely, but I would not put it out of the question.
    - nradov 5y ago
      Hyperloop doesn't exist in any meaningful sense.
- 5y ago
ripe 5y ago
> If an automated car can handle all that safely without reverting to human assistance, what exactly can't it handle, and why is it not full self driving? The answer is in the article, and in the definition of Level 3: the automated feature is restricted to a certain operational design domain. For Drive Pilot, this domain comprises specific sections of highway, maximum speed, and certain weather conditions. It refuses to operate outside that domain, even if technically the software might be capable of doing so. When it detects that it is leaving or has left that domain, it tries to hand back control to the human within a short period, failing which it brings the vehicle to a safe halt and calls for help.
- yholio 5y ago
  You misunderstood the question. Everything points to the conclusion that solving the self driving problem for that specific operational domain, congested highway, is almost just as hard as solving self-driving in the general case. The logical conclusion is then that the failure of the AI and calling for help from a human driver will either never happen at all, since you have a safe level 5 algorithm, or will happen abruptly in dangerous trafic situations that the driver cannot be reasonably expected to recover from.
  - ripe 5y ago
    Ah, I think I understand the question better now. Thanks for clarifying it. > Everything points to the conclusion that solving the self driving problem for that specific operational domain, congested highway, is almost just as hard as solving self-driving in the general case. The general self-driving problem is a much harder AI problem than the problem that Mercedes is solving with Drive Pilot. The operational domain of Drive Pilot is not just any congested highway, it's particular stretches of limited-access freeways in Germany that have been identified and studied by Daimler engineers. These roads have machine-detectable lane markings and no tunnels, toll booths or traffic control devices (stop signs or lights). The system uses satellite positioning with live corrections applied for transmission errors along these particular roads. The vendor has prepared detailed digital maps of particular signs, fence posts, and guardrails within this geo-fenced environment, and the sensors compare detected objects with the map, combining this positioning data with the satellite method. This level of preparation reduces the long tail of uncovered cases for this domain greatly, making this solution much less risky for the car and for its neighbors, to the point where the vendor can assume liability and regulators can adopt a framework to support it and evaluate its use. The general Level 5 self-driving problem throws too many uncovered corner cases at an autonomous car driving on random roadways designed for people. In my opinion, the current approach of throwing more and more data at ML networks cannot by itself bring down the risk to a low enough level to make it practical in my lifetime. Neural networks are not sufficient, and if we choose to enhance the roadways, then neural networks might not be necessary. Conceptual breakthroughs are needed. I should mention that I build robots for a living, though not self-driving cars, so maybe these other vendors and their investors know something I don't.

by yholio 5y ago

babyshake 5y ago

SAE vehicle autonomy levels are mostly a fiction in terms of not being highly contextual. The practical approach for self-driving is to take an actuarial, data-driven approach. For example, on certain streets or neighborhoods at certain times of day, you can be sure with a 99% confidence level that you will not encounter certain types of situations. A Level 3 vehicle in Phoenix is not necessarily a Level 3 vehicle in Manhattan or Nairobi.

ghaff 5y ago

I generally disagree with respect to L4. A car that is fully autonomous under a precisely defined set of circumstances, especially certain types of roads and possibly under certain environmental conditions could still be extremely useful--assuming you can fully hand off control. A lot of driving is fairly predictable highway driving even if it doesn't help people who want a door-to-door robo-taxi.

yholio 5y ago
If you can geofence those specific circumstance and guarantee they won't happen on useful stretches of road, then congratulations, you have a car that reaches level 5 on certain roads. But that's not what level 4 generally implies, they call it "High Driving Automation" and the expectation is that the driver is on hand to correct the AI in certain limited scenarios.
- ghaff 5y ago
  This is how JD Power describes it: Referred to as high-driving automation, Level 4 autonomy does not require any human interaction in the vehicle’s operation because it is programmed to stop itself in the event of system failure. Since a human driver is never needed, a Level 4 vehicle may not have a steering wheel and pedals. And yes, at Level 4, you can take a nap while riding in the vehicle. Level 4 driving automation technology is for use in driverless taxis and public transportation services. Such vehicles will be programmed to travel between Point A and Point B and restricted to specific geographic boundaries by geofencing technology. Certain conditions may limit or cancel Level 4 autonomous vehicle operation, such as severe weather. I admit that I thought of L4 as being fully autonomous under specific pre-defined conditions (given that's what would actually be useful and that's probably why I've heard it described that way.) The key would seem to be that the vehicle won't try to hand off control unexpectedly given, in part, that a vehicle can't always stop safely.
nradov 5y ago
Highway driving stops being predictable when a deer steps onto the road.
- ghaff 5y ago
  True. But that's actually the sort of thing I'd expect an autonomous driving system to likely be able to handle as well as a person.

azinman2 5y ago

Exactly. Level 5 or bust. It’s fiction to believe anyone can quickly take over — it is not how our brains work. If anything less than Level 5 does a good job much of the time, that only reinforces to the driver that they can completely disengage. It takes many seconds to not only respond to a call that they need to put down their phone and look up, but also to then grok the scene and keep moving. It’s just not cognitively realistic.

skohan 5y ago

Couldn't you imagine a world where infrastructure is adapted for safe self-driving? For instance, what if you had roads (think highways) where self-driving was mandatory? By applying certain constraints, you might be able to create situations where autonomous vehicles are practical. Maybe it's something like a carpool lane. Autonomous vehicles get the far left lane, and are allowed to drive much faster and closer together than manually controlled vehicles.

yholio 5y ago
I can imagine it for the sake of argument, but without strong govt action it has a bootstrap problem: no one will create such a lane if there are no such cars, and no one will buy and develop them if there are no roads you can use them on. Moreover, if the goal is to achieve autonomy with relatively dumb algorithms that fall short of full self driving, you have to do grade separation, fencing, tunnels or overground etc., otherwise you still need full self driving to handle those case where bikes or children roam around in your "dedicated" lane, cars and garbage bins are parked on it, drunk drivers suddenly invade it etc. And if you go the full trouble of dedicated and grade separated infrastructure, why not build a tramway or PRT (like the Booring Co Loop system) that any citizen can benefit from? All in all, I can't see the scenario you are describing happening in any real city with real politics, maybe some oil prince can decree it.
- skohan 5y ago
  I mean if hyperloop can exist, I wouldn't rule it out completely. One path I could imagine: a city like Phoenix parners with autonomous vehicle manufacturers to create a bubble where self driving cars can exist. It goes really well, other people want it, and other cities fall suit. I'm not saying it's likely, but I would not put it out of the question.
  - nradov 5y ago
    Hyperloop doesn't exist in any meaningful sense.
5y ago

ripe 5y ago

> If an automated car can handle all that safely without reverting to human assistance, what exactly can't it handle, and why is it not full self driving? The answer is in the article, and in the definition of Level 3: the automated feature is restricted to a certain operational design domain. For Drive Pilot, this domain comprises specific sections of highway, maximum speed, and certain weather conditions. It refuses to operate outside that domain, even if technically the software might be capable of doing so. When it detects that it is leaving or has left that domain, it tries to hand back control to the human within a short period, failing which it brings the vehicle to a safe halt and calls for help.

yholio 5y ago
You misunderstood the question. Everything points to the conclusion that solving the self driving problem for that specific operational domain, congested highway, is almost just as hard as solving self-driving in the general case. The logical conclusion is then that the failure of the AI and calling for help from a human driver will either never happen at all, since you have a safe level 5 algorithm, or will happen abruptly in dangerous trafic situations that the driver cannot be reasonably expected to recover from.
- ripe 5y ago
  Ah, I think I understand the question better now. Thanks for clarifying it. > Everything points to the conclusion that solving the self driving problem for that specific operational domain, congested highway, is almost just as hard as solving self-driving in the general case. The general self-driving problem is a much harder AI problem than the problem that Mercedes is solving with Drive Pilot. The operational domain of Drive Pilot is not just any congested highway, it's particular stretches of limited-access freeways in Germany that have been identified and studied by Daimler engineers. These roads have machine-detectable lane markings and no tunnels, toll booths or traffic control devices (stop signs or lights). The system uses satellite positioning with live corrections applied for transmission errors along these particular roads. The vendor has prepared detailed digital maps of particular signs, fence posts, and guardrails within this geo-fenced environment, and the sensors compare detected objects with the map, combining this positioning data with the satellite method. This level of preparation reduces the long tail of uncovered cases for this domain greatly, making this solution much less risky for the car and for its neighbors, to the point where the vendor can assume liability and regulators can adopt a framework to support it and evaluate its use. The general Level 5 self-driving problem throws too many uncovered corner cases at an autonomous car driving on random roadways designed for people. In my opinion, the current approach of throwing more and more data at ML networks cannot by itself bring down the risk to a low enough level to make it practical in my lifetime. Neural networks are not sufficient, and if we choose to enhance the roadways, then neural networks might not be necessary. Conceptual breakthroughs are needed. I should mention that I build robots for a living, though not self-driving cars, so maybe these other vendors and their investors know something I don't.