6 ms·

This was a fun interview, thanks Trevor! There are a few controverial bits that I'd love to hear what HN thinks. 1. Boston Dynamics Spot, where you have a tec

This was a fun interview, thanks Trevor! There are a few controverial bits that I'd love to hear what HN thinks. 1. Boston Dynamics Spot, where you have a technical achievement without a clear application 2. What it takes to win in self driving. At one point Musk said "If you have accurate vector space representation, then you're kind of like a video game." That's just not true.

rozgo 6y ago
2. "you're kind of like a video game." I think this is a fair characterization. It's obviously not a video game, but there is enormous overlap, once you've successfully modeled your perception in vector space. This is actually my entire strategy for approaching autonomy, I make as much of the problem space a game; just a serious one. And the most successful optimizations and solutions are the ones that adopt techniques from games. Many academic solutions are formalizations for old-school game "hacks". In my line of work, training, in many occasions, is about making deep computer vision match ground truth, generated by game tech.
- bumby 6y ago
  I actually think this is one area where people are conflating a perception problem with a much more complex problem. In many ways, “perception” is solved or close to it. But this is just a subset of the self-driving problem
  - jacquesm 6y ago
    Perception is where it begins. Then you can start to reason. Case in point: just the other day I drove past a line of cars. Right hand side passenger door opens, so I slow down by lifting a bit and sure enough, less than a second later the driver side door gets flung open. If not for reasoning then I would not have lifted and likely would have hit either the door or the former driver of the car when they got out. Those little hints that we humans are good at processing (passenger disembarks: driver likely to follow) would be very hard to teach a computer in such a way that it would not lead to false positives all the time.
    - inventtheday 6y ago
      couldn't that specific correlation be learned with enough data?
      - bumby 6y ago
        I think maybe you could, but I think it would take a lot of data to correlate the edge cases. Think about seeing a playground ball rolling into the street. Most human drivers would anticipate a child jumping into the street after the ball even though they’ve never experienced that situation before because they can correlate the perception with non-driving scenarios (I.e. playground + ball is associated with children). A self-driving car may not have the non-driving context and I’m not sure society is willing to accept the outcome while the cars gather enough data to learn. And that’s to say nothing of all the off-nominal maintenance related conditions to account for. Maybe there can be enough data sharing to update models effectively. It will be interesting to see this industry evolve
      - DayDollar 6y ago
        Comp vis often reclassifys objects, loosing history.. You see them get in- you know they are closing the doors. Comp vis sees a butterfly, and sees a car with doors half open. If you do not allow for frequent doubt, aka re-classification, you get hallucinating systems, imagining a bike-rishka as a slow car with people hanging out
        bumby 6y ago
        There’s been some cases that highlight the danger of this as well as bad examples of how mitigations were poorly implemented. I believe a case of a pedestrian being killed by a self driving car was related to the system reclassifying enough to initiate a deliberate delay timer. In a system traveling 60mph this delay in decision was enough to cause a mishap
        Animats 6y ago
        Cite? The Uber killing was simply that the self-driving system was not allowed to do an emergency stop. All it could do was alert the driver. Presumably because they had a high false alarm rate. The serious accidents involving self-driving cars have all been huge failures at the basic "don't run into things" task. Google/Waymo, which seems to be past that, has subtle problems, like "projected that bus couldn't fit through space to left of car in wide lane and so turned slightly into path of bus and was scraped." The standard Waymo accident is "advanced into intersection, detected cross traffic, stopped, was rear-ended by human driver at low speed".[1] [1] https://www.dmv.ca.gov/portal/vehicle-industry-services/autonomous-vehicles/autonomous-vehicle-collision-reports/ https://www.dmv.ca.gov/portal/vehicle-industry-services/auto...
Animats 6y ago
Spot does have that problem. Back in the 1990s, before Boston Dynamics had legged running really working, I was looking at that and running simulations. The only market that looked reasonable was toys. A $1000 toy seemed marginal. The Sony Aibo was in that space, and it was withdrawn. Sony had a follow-on humanoid, but it never shipped as a product. I met the designer of the Furby and found out what its manufacturing cost is. Let's just say it's less than a meal at most fast food places. Big Dog and the Legged Squad Support System were impressive, but not useful. The USMC rejected the LS3 and bought small ATVs instead. Much more practical. (Reminds me of the U.S. Army's flying saucer, the AvroCar. That was a disk-shaped aircraft that was supposed to be a flying Jeep. It was unstable, underpowered, and couldn't get out of its own ground effect. Really cool, though. The Army had another project at the same time, the Utility Helicopter program. That resulted in the UH-1 "Huey", one of the most successful helicopters of all time, with over 16,000 made and many still in use after six decades. It's not exotic but usually will get you there and back.)
- beambot 6y ago
  I heard through the grapevine that SoftBank sold something like 22,000 Pepper robots at $12k each. I never understood the utility -- I only ever saw them used as chintzy marketing gimmicks. I could easily see 1000 qty sales of SpotMini for the cool factor alone -- i.e. research labs, corporate labs, startups, etc without ever finding a real "value" application. This drives me nuts as a roboticist & CEO of a service robotics company. But so it goes -- at least it's better than their past product: Youtube Videos.
  - ivankirigin 6y ago
    Your company https://cobaltrobotics.com/ https://cobaltrobotics.com/ is really cool. I think it highlights that a vertical with specific needs can be met with the right hardware. No need for a face or arms.
    - beambot 6y ago
      Thanks for the kind words.
tlb 6y ago
You can divide the problem into - perception: sensors -> vector space (a list of objects with their locations) - planning: vector space -> path Both are hard, but for the planning problem you can do most of the training and testing in simulation. It's fairly easy to simulate the response of the car. The perception problem is more expensive to solve, because you have to collect so much data under every weather and lighting condition.

rozgo 6y ago

2. "you're kind of like a video game." I think this is a fair characterization. It's obviously not a video game, but there is enormous overlap, once you've successfully modeled your perception in vector space. This is actually my entire strategy for approaching autonomy, I make as much of the problem space a game; just a serious one. And the most successful optimizations and solutions are the ones that adopt techniques from games. Many academic solutions are formalizations for old-school game "hacks". In my line of work, training, in many occasions, is about making deep computer vision match ground truth, generated by game tech.

bumby 6y ago
I actually think this is one area where people are conflating a perception problem with a much more complex problem. In many ways, “perception” is solved or close to it. But this is just a subset of the self-driving problem
- jacquesm 6y ago
  Perception is where it begins. Then you can start to reason. Case in point: just the other day I drove past a line of cars. Right hand side passenger door opens, so I slow down by lifting a bit and sure enough, less than a second later the driver side door gets flung open. If not for reasoning then I would not have lifted and likely would have hit either the door or the former driver of the car when they got out. Those little hints that we humans are good at processing (passenger disembarks: driver likely to follow) would be very hard to teach a computer in such a way that it would not lead to false positives all the time.
  - inventtheday 6y ago
    couldn't that specific correlation be learned with enough data?
    - bumby 6y ago
      I think maybe you could, but I think it would take a lot of data to correlate the edge cases. Think about seeing a playground ball rolling into the street. Most human drivers would anticipate a child jumping into the street after the ball even though they’ve never experienced that situation before because they can correlate the perception with non-driving scenarios (I.e. playground + ball is associated with children). A self-driving car may not have the non-driving context and I’m not sure society is willing to accept the outcome while the cars gather enough data to learn. And that’s to say nothing of all the off-nominal maintenance related conditions to account for. Maybe there can be enough data sharing to update models effectively. It will be interesting to see this industry evolve
    - DayDollar 6y ago
      Comp vis often reclassifys objects, loosing history.. You see them get in- you know they are closing the doors. Comp vis sees a butterfly, and sees a car with doors half open. If you do not allow for frequent doubt, aka re-classification, you get hallucinating systems, imagining a bike-rishka as a slow car with people hanging out
      - bumby 6y ago
        There’s been some cases that highlight the danger of this as well as bad examples of how mitigations were poorly implemented. I believe a case of a pedestrian being killed by a self driving car was related to the system reclassifying enough to initiate a deliberate delay timer. In a system traveling 60mph this delay in decision was enough to cause a mishap
        Animats 6y ago
        Cite? The Uber killing was simply that the self-driving system was not allowed to do an emergency stop. All it could do was alert the driver. Presumably because they had a high false alarm rate. The serious accidents involving self-driving cars have all been huge failures at the basic "don't run into things" task. Google/Waymo, which seems to be past that, has subtle problems, like "projected that bus couldn't fit through space to left of car in wide lane and so turned slightly into path of bus and was scraped." The standard Waymo accident is "advanced into intersection, detected cross traffic, stopped, was rear-ended by human driver at low speed".[1] [1] https://www.dmv.ca.gov/portal/vehicle-industry-services/autonomous-vehicles/autonomous-vehicle-collision-reports/ https://www.dmv.ca.gov/portal/vehicle-industry-services/auto...

Animats 6y ago

Spot does have that problem. Back in the 1990s, before Boston Dynamics had legged running really working, I was looking at that and running simulations. The only market that looked reasonable was toys. A $1000 toy seemed marginal. The Sony Aibo was in that space, and it was withdrawn. Sony had a follow-on humanoid, but it never shipped as a product. I met the designer of the Furby and found out what its manufacturing cost is. Let's just say it's less than a meal at most fast food places. Big Dog and the Legged Squad Support System were impressive, but not useful. The USMC rejected the LS3 and bought small ATVs instead. Much more practical. (Reminds me of the U.S. Army's flying saucer, the AvroCar. That was a disk-shaped aircraft that was supposed to be a flying Jeep. It was unstable, underpowered, and couldn't get out of its own ground effect. Really cool, though. The Army had another project at the same time, the Utility Helicopter program. That resulted in the UH-1 "Huey", one of the most successful helicopters of all time, with over 16,000 made and many still in use after six decades. It's not exotic but usually will get you there and back.)

beambot 6y ago
I heard through the grapevine that SoftBank sold something like 22,000 Pepper robots at $12k each. I never understood the utility -- I only ever saw them used as chintzy marketing gimmicks. I could easily see 1000 qty sales of SpotMini for the cool factor alone -- i.e. research labs, corporate labs, startups, etc without ever finding a real "value" application. This drives me nuts as a roboticist & CEO of a service robotics company. But so it goes -- at least it's better than their past product: Youtube Videos.
- ivankirigin 6y ago
  Your company https://cobaltrobotics.com/ https://cobaltrobotics.com/ is really cool. I think it highlights that a vertical with specific needs can be met with the right hardware. No need for a face or arms.
  - beambot 6y ago
    Thanks for the kind words.

tlb 6y ago

You can divide the problem into - perception: sensors -> vector space (a list of objects with their locations) - planning: vector space -> path Both are hard, but for the planning problem you can do most of the training and testing in simulation. It's fairly easy to simulate the response of the car. The perception problem is more expensive to solve, because you have to collect so much data under every weather and lighting condition.