4 ms·

I know this is a heated debate, but it's interesting to see the divergent approaches that are being taken to achieve self driving. In one corner, it's lidar-bas

I know this is a heated debate, but it's interesting to see the divergent approaches that are being taken to achieve self driving. In one corner, it's lidar-based, equipping vehicles produced by other companies, organically and incrementally growing the navigation technology. In the other, it's video-based with now 100s of thousands of vehicles feeding data to a massive neural net. Each has very well thought-out arguments for the approach they are taking, and I expect that both would ultimately 'work' by reducing the inherent risks of taking a car ride by a substantial amount (90%?). I think the winner will be the one who gets to that milestone soonest. It's entertaining to watch.

krona 8y ago
You forgot radar. Most startups are using a combination of all the perception approaches, especially vehicle manufacturers.
Shivetya 8y ago
While I own a TM3 I have little faith in their full self driving solution and did not opt for that. I did take the enhanced auto pilot and recent additions show how much more work it will take. An example would be that it now shows different vehicle types and even pedestrians near you yet when all surround objects are obviously not moving it still has them moving on the display. It still has issues with certain bridges and apparently it is hard coded disabled around certain bridges to the point there are videos of it not recognizing a merging semi. with regards to Waymo's solution, I am concerned that it should have something in the driver's seat. I don't care if its a Total Recall "Jiffy Cab" type solution. not so much for the riders but for those who are also sharing the same road be that other vehicles or pedestrians. Something that clearly indicates it is under control.
- rbowlby83 8y ago
  Perhaps that blowup pilot from Airplane?
WhompingWindows 8y ago
Definitely entertaining to watch, I agree heartily. I don't agree that whoever gets there first will be the winner. As we see with Waymo one, it's quite probable that there will be a limited roll-out of features/locations, such that it'll be much like these electric scooters: available in some places, but not others. The roll-out should be much slower in winter months and in places with inclement weather, however I see no reason that warm/dry places couldn't expect to see a big roll-out in the next decade. Phoenix is the beginning for Waymo, what about Tucson, Dallas, Houston, Miami, San Diego, etc? Whereas, with Tesla, we'll likely see features rolling out slowly over time, an incremental push towards FSD, with owners being responsible for safety monitoring their own vehicle.
maxerickson 8y ago
Does Tesla do intersections yet?
aantix 8y ago
Curious; If you train a neural net to drive a car with X number of samples, and the car does something erratic, how do you determine which data item introduced the erratic behavior and how do you go about untraining that behavior?
- Hydraulix989 8y ago
  Gradient descent is done in batches and over the aggregate collection of the training data so it's pretty lossy. Usually the error is more of the fault of the learned model itself, and then the culprit is attributed to neuron "mis"-activations post-hoc. Often, problems take the form of generalization error instead. And yes, all of these problems are still quite unsolved, which is why we don't have end-to-end neural network-based self-driving cars, we just have rule based systems with a perception framework that just uses convnets for vision.
  - ghaff 8y ago
    I think there's this popular perception that you just stream in lots of data from road vehicles and the driving just gets better over time--rather than continuing to discover individual edge cases more or less manually and updating rules based on those.
- repsilat 8y ago
  That is one reason they don't do that. (Another is that the net would be useless if you upgraded the cameras...) Whatever the sensing technology, everyone uses it to build a "state of the world" model which is then fed into algorithms for planning movement. Nobody (to my knowledge) just wires up input pixel neurons to output pedal neurons.
- piccolbo 8y ago
  It's simple, you retrain the model several times removing one sample at a time (leave-one-out cross-validation), until the erratic behavior disappears. You repeat until there are no more instances of erratic behavior. Sorry couldn't resist the attempt at humor, but also consider it as an argument by contradiction that what you are asking is not possible. It's like asking which lesson caused me to hit a forehand over the tramlines. The answer is the one I didn't have time for.
forgot-my-pw 8y ago
In 2004-2007, DARPA had a few self driving contests in the desert. The documentary was an entertaining watch. IIRC, most of the cars didn't even make it to the end. https://en.wikipedia.org/wiki/DARPA_Grand_Challenge_(2004) https://en.wikipedia.org/wiki/DARPA_Grand_Challenge_(2004) https://www.pbs.org/wgbh/nova/darpa/ https://www.pbs.org/wgbh/nova/darpa/ Would be fun to get another test with current technology to see how far we've come. Though self driving cars nowadays are more focused on safety, not speed.
- dsfyu404ed 8y ago
  Off road racing has come a long way since 2004. Go watch some old King of the Hammers races from the early 2000s and watch some from recent years. Even if the AI goes off coarse it's going to be a lot harder to get into a in situations where "just point in the direction you want to go and pin the throttle" is not an acceptable solution in 2018 compared to 2004. The 2018 vehicle will be able to just ignore things that the 2004 vehicle had to actually navigate around. We also have cars that are almost fully drive by wire in stock form which frees up a ton of resources for software and testing. Basically in 2018 the vehicle side of things is a trivially solved problem and even with a "dumb" 2004 level AI at the wheel half the field won't DNF from mechanical failure
asdkhadsj 8y ago
In your two camps, I'd be curious to see why someone would specifically choose just video based? Ie, I'd expect (as a layman) neural nets to be able to consume many different data types. Why limit to just one? I wonder what is better about just video, rather than video+lidar+audio+x+y. Hell, one of the most neat things to me was how Teslas can scan multiple cars ahead to know of non-visible traffic blockages/etc. Using just video seems like a step backwards. Though, if it can produce better AI, I suppose it's still forwards.
- Traster 8y ago
  Well, the only person who appears to be stating that you can use video only is Tesla, and Elon Musk seems to be the only person stating that that's the right strategy. From what I've seen he basically says: If you focus on vision and solve it, you've got autonomous driving whereas if you use a kludge of different sensors none of them will be great and the final solution won't be great. As an engineer whose worked on similar projects this just seems down-right incorrect frankly- especially in safety critical systems. However, it's also worth noting that when he said that it was suggested an autonomous Tesla drive across the country would complete by EOY 2017[1]. It's ALSO worth noting that Tesla have really upsold this idea of 'Buy a Tesla now, and it'll magically become autonomous'[2] [1] https://electrek.co/2017/04/29/elon-musk-tesla-plan-level-5-full-autonomous-driving/ https://electrek.co/2017/04/29/elon-musk-tesla-plan-level-5-... [2] https://www.theverge.com/2018/6/11/17449076/tesla-autopilot-full-self-driving-elon-musk https://www.theverge.com/2018/6/11/17449076/tesla-autopilot-...
- dominotw 8y ago
  > i'd be curious to see why someone would specifically choose just video based? other than cost? Teslas are consumer vehicles that are price sensitive.
opencl 8y ago
It really doesn't seem like much of a heated debate. The entire industry aside from one company has decided to combine video+radar+LIDAR and the one company that isn't using LIDAR has shown no meaningful progress (in actual self-driving, not referring to driver assist tech) in years.

krona 8y ago

You forgot radar. Most startups are using a combination of all the perception approaches, especially vehicle manufacturers.

Shivetya 8y ago

While I own a TM3 I have little faith in their full self driving solution and did not opt for that. I did take the enhanced auto pilot and recent additions show how much more work it will take. An example would be that it now shows different vehicle types and even pedestrians near you yet when all surround objects are obviously not moving it still has them moving on the display. It still has issues with certain bridges and apparently it is hard coded disabled around certain bridges to the point there are videos of it not recognizing a merging semi. with regards to Waymo's solution, I am concerned that it should have something in the driver's seat. I don't care if its a Total Recall "Jiffy Cab" type solution. not so much for the riders but for those who are also sharing the same road be that other vehicles or pedestrians. Something that clearly indicates it is under control.

rbowlby83 8y ago
Perhaps that blowup pilot from Airplane?

WhompingWindows 8y ago

Definitely entertaining to watch, I agree heartily. I don't agree that whoever gets there first will be the winner. As we see with Waymo one, it's quite probable that there will be a limited roll-out of features/locations, such that it'll be much like these electric scooters: available in some places, but not others. The roll-out should be much slower in winter months and in places with inclement weather, however I see no reason that warm/dry places couldn't expect to see a big roll-out in the next decade. Phoenix is the beginning for Waymo, what about Tucson, Dallas, Houston, Miami, San Diego, etc? Whereas, with Tesla, we'll likely see features rolling out slowly over time, an incremental push towards FSD, with owners being responsible for safety monitoring their own vehicle.

maxerickson 8y ago

Does Tesla do intersections yet?

aantix 8y ago

Curious; If you train a neural net to drive a car with X number of samples, and the car does something erratic, how do you determine which data item introduced the erratic behavior and how do you go about untraining that behavior?

Hydraulix989 8y ago
Gradient descent is done in batches and over the aggregate collection of the training data so it's pretty lossy. Usually the error is more of the fault of the learned model itself, and then the culprit is attributed to neuron "mis"-activations post-hoc. Often, problems take the form of generalization error instead. And yes, all of these problems are still quite unsolved, which is why we don't have end-to-end neural network-based self-driving cars, we just have rule based systems with a perception framework that just uses convnets for vision.
- ghaff 8y ago
  I think there's this popular perception that you just stream in lots of data from road vehicles and the driving just gets better over time--rather than continuing to discover individual edge cases more or less manually and updating rules based on those.
repsilat 8y ago
That is one reason they don't do that. (Another is that the net would be useless if you upgraded the cameras...) Whatever the sensing technology, everyone uses it to build a "state of the world" model which is then fed into algorithms for planning movement. Nobody (to my knowledge) just wires up input pixel neurons to output pedal neurons.
piccolbo 8y ago
It's simple, you retrain the model several times removing one sample at a time (leave-one-out cross-validation), until the erratic behavior disappears. You repeat until there are no more instances of erratic behavior. Sorry couldn't resist the attempt at humor, but also consider it as an argument by contradiction that what you are asking is not possible. It's like asking which lesson caused me to hit a forehand over the tramlines. The answer is the one I didn't have time for.

forgot-my-pw 8y ago

In 2004-2007, DARPA had a few self driving contests in the desert. The documentary was an entertaining watch. IIRC, most of the cars didn't even make it to the end. https://en.wikipedia.org/wiki/DARPA_Grand_Challenge_(2004) https://en.wikipedia.org/wiki/DARPA_Grand_Challenge_(2004) https://www.pbs.org/wgbh/nova/darpa/ https://www.pbs.org/wgbh/nova/darpa/ Would be fun to get another test with current technology to see how far we've come. Though self driving cars nowadays are more focused on safety, not speed.

dsfyu404ed 8y ago
Off road racing has come a long way since 2004. Go watch some old King of the Hammers races from the early 2000s and watch some from recent years. Even if the AI goes off coarse it's going to be a lot harder to get into a in situations where "just point in the direction you want to go and pin the throttle" is not an acceptable solution in 2018 compared to 2004. The 2018 vehicle will be able to just ignore things that the 2004 vehicle had to actually navigate around. We also have cars that are almost fully drive by wire in stock form which frees up a ton of resources for software and testing. Basically in 2018 the vehicle side of things is a trivially solved problem and even with a "dumb" 2004 level AI at the wheel half the field won't DNF from mechanical failure

asdkhadsj 8y ago

In your two camps, I'd be curious to see why someone would specifically choose just video based? Ie, I'd expect (as a layman) neural nets to be able to consume many different data types. Why limit to just one? I wonder what is better about just video, rather than video+lidar+audio+x+y. Hell, one of the most neat things to me was how Teslas can scan multiple cars ahead to know of non-visible traffic blockages/etc. Using just video seems like a step backwards. Though, if it can produce better AI, I suppose it's still forwards.

Traster 8y ago
Well, the only person who appears to be stating that you can use video only is Tesla, and Elon Musk seems to be the only person stating that that's the right strategy. From what I've seen he basically says: If you focus on vision and solve it, you've got autonomous driving whereas if you use a kludge of different sensors none of them will be great and the final solution won't be great. As an engineer whose worked on similar projects this just seems down-right incorrect frankly- especially in safety critical systems. However, it's also worth noting that when he said that it was suggested an autonomous Tesla drive across the country would complete by EOY 2017[1]. It's ALSO worth noting that Tesla have really upsold this idea of 'Buy a Tesla now, and it'll magically become autonomous'[2] [1] https://electrek.co/2017/04/29/elon-musk-tesla-plan-level-5-full-autonomous-driving/ https://electrek.co/2017/04/29/elon-musk-tesla-plan-level-5-... [2] https://www.theverge.com/2018/6/11/17449076/tesla-autopilot-full-self-driving-elon-musk https://www.theverge.com/2018/6/11/17449076/tesla-autopilot-...
dominotw 8y ago
> i'd be curious to see why someone would specifically choose just video based? other than cost? Teslas are consumer vehicles that are price sensitive.

opencl 8y ago

It really doesn't seem like much of a heated debate. The entire industry aside from one company has decided to combine video+radar+LIDAR and the one company that isn't using LIDAR has shown no meaningful progress (in actual self-driving, not referring to driver assist tech) in years.