4 ms·

Disclaimer up front: I work for GM, I don't work on SDC. As I see it, self-driving is a cursed[0][1] problem. If you could choose a different problem space, or

by csours 7y ago

Disclaimer up front: I work for GM, I don't work on SDC. As I see it, self-driving is a cursed[0][1] problem. If you could choose a different problem space, or to ignore some complications, self-driving would merely be very hard. But the requirement to handle ANY external behavior with ANY external conditions and navigate to ANY destination, while also maintaining safety is impossible* satisfy. One cursed corner of the problem space is ML itself: ML is amazing in that it enables emergent behavior [3], but ML is terrible in that it gives rise to emergent behavior. The traditional engineering mindset wants a map of inputs to outputs, but you don't get to choose your inputs in the SDC world, and you can't specify all of your outputs. Another cursed corner is the Always/Never[2] problem. You want safety features like Automated Emergency Braking to Always kick in when there is a problem, and you Never want them to kick in when there isn't a problem. I really don't know how any of this gets fixed. I do think that sensor fusion and advances in AI can reduce the size of some of the cursed area of the problem, but the problem is also meta-cursed in the definition of Self-Driving: the solution to normal cursed problems is to reduce or change the solution scope, but if you reduce or change the Self-Driving solution scope, then "it's not real self driving". 0. https://twvideo01.ubm-us.net/o1/vault/gdc2019/presentations/Jaffe_Alex_Cursed_Problems_In.pdf https://twvideo01.ubm-us.net/o1/vault/gdc2019/presentations/... 1. https://en.wikipedia.org/wiki/Curse_of_dimensionality#Machine_learning https://en.wikipedia.org/wiki/Curse_of_dimensionality#Machin... 2. https://www.newyorker.com/news/news-desk/always-never https://www.newyorker.com/news/news-desk/always-never --- 3. edit: [pdf] https://core.ac.uk/download/pdf/81580604.pdf https://core.ac.uk/download/pdf/81580604.pdf Especially see section 4.3

jacquesm 7y ago
This is spot on. The Always/Never is well articulated, it is exactly why I got rid of a vehicle with auto braking. It did it twice when it wasn't supposed to and that was more dangerous than if it had never triggered at all. And yet, in an actual imminent crash it would have probably responded a lot quicker than I ever could. Still, I'm not going to be scared twice in a month by my car doing stuff 'on its own' to my possible detriment.
- GeorgeTirebiter 7y ago
  Could you share which company's auto-braking had the false positives? What is the acceptable false-positive threshold for you to use auto-braking again? Also, did your system tell you why it decided to brake (albeit incorrectly)?
erikpukinskis 7y ago
Waymo and Tesla have both solved the problem, in different ways, by gating the feature. Why would they need to force Level 5? They can operate with the promise of L5 in the future, until the moment the statistics show it’s safe to turn it on. Between now and then they can provide a steadily increasing supply of value to their customers. And steadily ramp the data they have to improve the system. I have no idea if it will be 2 years or 10 years or 100 years before they can turn on Level 5 without a geofence.... but what difference does that number make? It’s academic. It’s not an existential question for either company’s self driving value prop. And consider that different jurisdictions will set different dates. All it takes is one government to OK L5 in some area, and Tesla’s ass is covered. And every jurisdiction that signs on will put more pressure on other jurisdictions to follow suit. To me, that’s the brilliance of Tesla’s strategy. They have decoupled their forcing functions from the L4/L5 legislative question. Governments can drag their feet as long as they want but it won’t hurt Tesla’s ability to sell the tech or collect the data they need to improve the tech. They are already selling autonomy features, and they can make that value prop incrementally better year over year for as long as it takes.
GeorgeTirebiter 7y ago
Sensor fusion? Absolutely required. Also, I think "AI Fusion" is needed, where different AI approaches evaluate a situation. Maybe a supervisor DNN AI which is looking at the output from the AIs below, inferring patterns. DNNs (alone) are incapable of Level 5. OTOH, it's conceivable that in this limited domain, a "Society of AIs" might yield 'acceptable' performance ( 1K deaths per year? ). I would also like to see a given driving system rated at maximum speed for autonomous use. If you're moving at 5 MPH and your stopping distance (reaction time + coming to a stop) is 5 feet, then your 'Zone of Despair' is the 5 feet in front of you (and some on either side). If you absolutely know this Zone is clear, you can proceed with confidence. It's a hellishly difficult problem; but I don't think it requires AGI to solve 'adequately'.

by csours 7y ago

jacquesm 7y ago

This is spot on. The Always/Never is well articulated, it is exactly why I got rid of a vehicle with auto braking. It did it twice when it wasn't supposed to and that was more dangerous than if it had never triggered at all. And yet, in an actual imminent crash it would have probably responded a lot quicker than I ever could. Still, I'm not going to be scared twice in a month by my car doing stuff 'on its own' to my possible detriment.

GeorgeTirebiter 7y ago
Could you share which company's auto-braking had the false positives? What is the acceptable false-positive threshold for you to use auto-braking again? Also, did your system tell you why it decided to brake (albeit incorrectly)?

erikpukinskis 7y ago

Waymo and Tesla have both solved the problem, in different ways, by gating the feature. Why would they need to force Level 5? They can operate with the promise of L5 in the future, until the moment the statistics show it’s safe to turn it on. Between now and then they can provide a steadily increasing supply of value to their customers. And steadily ramp the data they have to improve the system. I have no idea if it will be 2 years or 10 years or 100 years before they can turn on Level 5 without a geofence.... but what difference does that number make? It’s academic. It’s not an existential question for either company’s self driving value prop. And consider that different jurisdictions will set different dates. All it takes is one government to OK L5 in some area, and Tesla’s ass is covered. And every jurisdiction that signs on will put more pressure on other jurisdictions to follow suit. To me, that’s the brilliance of Tesla’s strategy. They have decoupled their forcing functions from the L4/L5 legislative question. Governments can drag their feet as long as they want but it won’t hurt Tesla’s ability to sell the tech or collect the data they need to improve the tech. They are already selling autonomy features, and they can make that value prop incrementally better year over year for as long as it takes.

GeorgeTirebiter 7y ago

Sensor fusion? Absolutely required. Also, I think "AI Fusion" is needed, where different AI approaches evaluate a situation. Maybe a supervisor DNN AI which is looking at the output from the AIs below, inferring patterns. DNNs (alone) are incapable of Level 5. OTOH, it's conceivable that in this limited domain, a "Society of AIs" might yield 'acceptable' performance ( 1K deaths per year? ). I would also like to see a given driving system rated at maximum speed for autonomous use. If you're moving at 5 MPH and your stopping distance (reaction time + coming to a stop) is 5 feet, then your 'Zone of Despair' is the 5 feet in front of you (and some on either side). If you absolutely know this Zone is clear, you can proceed with confidence. It's a hellishly difficult problem; but I don't think it requires AGI to solve 'adequately'.