Jessica Chobot: Hey there I’m Jessica Chobot and this is AI: Hype vs Reality. An original podcast from Dell Technologies. I’m at the University of Toronto’s self driving car lab. There’s a group of students here competing in the auto-drive challenge, trying to build a fully autonomous car.
In fact these guys were the winners of round one with their car Zeus. Which, to tell you the truth looks a little like the google car but a little more homemade. You know it’s got a huge spinning scanner thing on top, cameras pointing in all directions and wires running all over it. So, with that being said let’s meet the guys who built this.
Jessica Chobot: Jessica.
Jessica Chobot: Keenan. Jessica. Nice to meet you.
Jessica Chobot: Sepehr. So this is Zeus.
Jessica Chobot: All right. So how far along in building Zeus are you guys currently.
Keenan: Right so the autobot challenge is this three year competition. We’re currently in the second year of the competition and so this year we’re only six weeks away from the second installment of the competition which is down at the University of Michigan. Where we’re going to see traffic lights as well as dynamic pedestrians and a lot of different signs that we have to deal with.
Jessica Chobot: All right. So, I’m legit very nervous. Believe it or not about entering this car because: a: I’m concerned for my safety, b: you have a lot of very expensive equipment and I don’t have that kind of money to cover those checks. So how often do people actually get behind the wheel and drive this thing?
Keenan: On a fairly semi regular basis we’ll let people sit in the back of the car and they’ll get a ride. But letting people actually sit behind the driver’s wheel and actually get that experience that we basically never do that.
Jessica Chobot: Never. Okay. Yeah. Awesome. Well great. Can we get started?
Jessica Chobot: Fantastic.
Sepehr: Let’s go.
Jessica Chobot: All right, so we’re doing it. Is this it? Autonomous? Self driving?
Sepehr: One. Two. Three. Go.
Jessica Chobot: I hear things. Oh (beep)
Okay, now before we hear the rest of that experimental drive, we need to dissect the hype around self driving cars.
The greatest disruption in transportation since the Model T Ford assembly line. Fleets of fully autonomous cars filling city streets. Driverless taxis replacing all ride sharing services. Cars produced without steering wheels, gas or brake pedals. All of this any day now.
And to give us a reality check. I’m talking with Dave Graham. He lives and breathes emerging tech for Dell Technologies. So Dave, would you say that AI currently or potentially in the near future is safer than human drivers?
Dave Graham: I think the potential is there. Absolutely. I think right now we have a low statistical probability that’s going to help. We just don’t have enough data. And I say that we have petabytes of data on interactions with this stuff.
But we don’t have enough out there. So as those sample sizes increase, as people start to use this more as we kind of demonstrate this I think you’ll see, yeah, the potential for this is incredible. That’s me being an optimist in the face of a lot of pessimism around some of these things, around this machine making decisions for you.
Jessica Chobot: So I’m actually going to be testing out an experimental self driving car later on in this show. But, before that, you’ve actually got some examples of other technologies and advances around autonomous cars.
Dave Graham: Yes. One of the biggest issues for self driving cars when it comes to safety is obviously the pedestrians, right? We don’t want to hit them, we don’t want to hurt them, we don’t want to do anything that would endanger their lives. Same thing with passengers. It’s always that delicate balance.
So, thing of it is there’s a lot of data out there from labs but not a lot from real world situations where people carrying groceries or walking around with their heads buried in their cell phones for example. So, I got to hear from Ram Vasudevan, professor of mechanical engineering at the University of Michigan, who took his research to the streets by filming real pedestrians in a real car.
Ram Vasudevan: A lot of the work in pedestrian prediction has in the past, has involved sort of looking at an individual and sort of just taking information about context. So that is, if they’re on the sidewalk or if they’re in the middle of the road and making predictions about what they’re capable of doing given that contextual information. The work we’ve done recently is basically looking not only at the body pose of the individual but looking at the biomechanics of the individual in order to make reasonable predictions of what they’re capable of doing next.
So what that means is, for instance, if I’m looking at a person and they’re pointed in a certain direction and if they’re walking forwards. For example if their torso is pointed into the sidewalk and their hands are swinging, right, that’s telling us something about where they’re capable of going next. Now, you predict where they’ll be several seconds from now and then what you want to be able to do is you get new information as the car moves forward or they move forward and now you want to update that prediction.
Right, and so it’s a balance of basically making long predictions but at the same time being nimble enough to update that prediction quickly, given new information. The difference between where our predictions say the person is going to be and where the person actually ended up on the order of centimeters.
Jessica Chobot: And so essentially, tell me if I’m wrong, but it sounds like what he’s saying is that it’s just consistently making assumptions based off of what it quote unquote, “sees,” happening in front of it. Which is exactly what we do when we’re driving. We’re looking behind us, we’re checking to the side, we see somebody start looking to cross the street and we kind of just start to slow down assuming that what if that person decides to make a run for it. Even though the crosswalk is only five steps away this person might do that and so we adjust for it. Welcome to LA, that happens all the time.
Dave Graham: Like a lot of that, you have to make a decision in the moment, “Is this object big enough to cause damage?” Like understanding kind of the impact of that image would have. A moose is a lot bigger than a squirrel right? A lot of this has to do with understanding the consequences of actions right? So when I write these maths I write these software bits and pieces that determine this, I have to look at the consequences. I have to understand what happens if.
Jessica Chobot: So Dave, in order to kind of prep for this conversation about self driving cars, who else did you talk to to get information on where it’s going and where we’re at right now.
Dave Graham: Well as we talked about, there’s a lot of data that has to be going back and forth between the vehicles and kind of base stations if you will or processing stations right? We want to learn more about the data we’re collecting as we go along. So obviously there’s things like 5G networks right, or LTE networks that kind of come into play there. So with all that data going back and forth, there’s a concept of having these smart cities right?
So as a car goes through a city it’s able to beacon off of these antennas and do all this type of thing, obviously run themselves. So there’s one company called Integrated Roadways that wants to actually turn the road itself into a network versus having to go off these towers and they have a product called Smart Pavement. This is company founder Tim Sylvester.
Tim Sylvester: Smart pavement is a prefabricated paving system with embedded sensing and communication technologies. It acts as the road’s driving surface, it collects data about traffic and it provides wireless network services for connected and autonomous vehicles as well as smart cities. Most people think about autonomous vehicles as being an onboard technology that’s independent of any sort of network changes.
But I would like to remind everyone that every type of communication and data driven technology that we have ever built required a network. From radio to television to telephones to cable, so I think it’s kind of, I’m struggling between whether I want to say silly or optimistic to believe that we’re going to be able to do all of this in vehicle. But let me ask you, if you lose your call when you’re talking on the phone regularly, that’s inconvenient. But if you are asleep in an autonomous vehicle and it loses its network connection, do you really want to take that risk?
What we do is, by turning the road into a touch pad where we can see at a thousand frames per second exactly where every vehicle is, we take out all of the guess work for the vehicle and we can tell the car, “Here’s where you are. Here’s where everybody else is and in the next millisecond, this is your safety zone that you can travel into with no concern about collision.”
Jessica Chobot: All right, so he makes a good point that maybe autonomous cars are being asked to do too much on their own. What do you think about that?
Dave Graham: I think, again, it’s kind of a sliding scale right? Where we’re at right now, they’re being asked to do a lot. I think long term again, kind of falling in that curve of miniaturization to compute increasing and some of the ancillary products that get installed in some of these vehicles like sensors and whatever. I think there a lot of opportunity for the cars to make better use of the resources that it has.
So I think there’s certainly room to grow within there. So again, I take it both and approach to what he’s saying. There’s augmentation so I think there’s reason to believe that what these smart pavements would augment the capability of the car not replace, not overtake. But in the same sense I think the car is going to be doing enough. It won’t be saturated I think that’s again one of those peripheries that we test to make sure they don’t exceed.
Jessica Chobot: Got it. Well Dave, thank you for the rundown. I am actually headed off to the University of Toronto to test an experimental self driving car. But before that I’m going to hear from one of the researchers who actually helped to make it happen. Wish me luck. Because I don’t know how this is going to turn out.
Dave Graham: Enjoy the ride.
Jessica Chobot: Thanks.
Ragavan T.: My name is Ragavan Thurairatnam and I’m a cofounder and the chief of machine learning at Dessa.
Jessica Chobot: Dessa works with a bunch of different companies helping them develop AI solutions. Last year they were approached by a group of students at the University of Toronto’s self driving car team looking for guidance. They had just entered the auto-drive challenge, a three year, international competition for students to build a self driving car.
Ragavan T.: We ended up meeting with the leadership team every two weeks or so. Checking in on the project progress, suggesting ways to improve the system to make sure that they hit their goal for the competition.
Jessica Chobot: On thing that Dessa helped the University of Toronto team with was the difference between testing and autonomous car and in the lab versus the real world.
Ragavan T.: To give an example, if I have a self driving car, and I see something in front of me and I think it might be a human, or I think it might be a piece of road. If I mistake a road for a human and I stop it’s not a big deal, it’s just kind of annoying, but if I mistake a human for a piece of road, and I drive through that, or try to drive through that human, I will injure or kill that human. So the cost of an error is very very important when you’re trying to build something in the real world.
Jessica Chobot: One advantage the University of Toronto team has over other competitors is location. Even though the car is tested in Arizona, they’re building it under the relatively harsh conditions of Canada.
Ragavan T.: The thing about Arizona is that it’s probably the easiest place to build and test self driven cars. They have very simple weather compared to Canada for example, which has lots more complicated weather. They have very big roads, they have few pedestrians, Canada’s the hardest place to make this work. So if they can make it work here they will make it work anywhere.
Jessica Chobot: To win the auto drive competition, they goal is to create a fully autonomous car. But is that even possible?
Ragavan T.: I think it’s likely possible. Because there’s so many crazy scenarios that we can handle as humans because we have the ability to reason from all this information we have about the world. So for example, let’s imagine I see a pick up truck and there’s some stop signs in the back of the pick up truck. As a human I know, okay, stop signs are moveable and also construction crews might be installing stop signs somewhere. If a stop sign isn’t fixed in a road, it’s probably not a stop sign I need to stop for.
But, an automated car can’t reason that same way. So that unusual situation, it might actually just stop. I think we can make cars smart enough that they can handle the vast vast majority of situations like 99.9999 etc situations. That would be far better than what we have now with humans driving and ethically it would be a better choice to allow that to happen than wait for self driving cars to be absolutely perfect which might not happen.
Jessica Chobot: So self driving cars may never be perfect but what about the car the University of Toronto students are experimenting with. The one I’m about to get into? Well the team did win the round of the AutoDrive challenge but that was mostly conceptual with some simple driving tasks. The next round and what I’ll be testing means actually driving down simulated city streets even trying to avoid pedestrians. Does Ragavan think they, and I, are going to be okay?
Ragavan T.: I think what they’ve done, is really remarkable for a bunch of students and it kind of speaks to the culture of AI. They come from the birthplace of deep learning, University of Toronto. I’m very confident that they will probably win again.
Jessica Chobot: Probably? Okay Ragavan, I will take your word for it but I am still pretty legitimately nervous the words, “experimental,” and “self-driving,” are kind of freaking me out right now.
Did you guys ever get pulled over by the cops?
Sepehr: No, thankfully, not yet.
Jessica Chobot: Maybe this’ll be the time. Maybe this is when we’ll get lucky. You got a lot of stuff going on in this car.
Keenan: Yeah we do.
Jessica Chobot: There are lots of things. I guess we’ll just start with the thing closest to me. What am I looking at on this monitor?
Keenan: Right so this visualization that we have here, is on the bottom left we’ve got some boxes put around pedestrians that we can detect so we can see that guy that’s far away. On the top left hand side we’ve got top down visualization of that guy that’s also walking towards us and on the right is actually this 3D cloud of points that’s being produced by a rotating laser sensor that we’ve got on the top of the car.
Jessica Chobot: Got it and so then the car takes all this information and processes it and then that’s how it drives?
Keenan: Yeah. Exactly.
Jessica Chobot: Oh, Okay. Well, seems simple enough and then you’ve got a big gigantic warning red button in the front and another big gigantic red button here on the middle console. So what are those for?
Sepehr: So these ones basically are if we encounter a problem that we’ve never seen before and we’re unable to conduct a safe takeover of an autonomous drive. This one will kill power to all of our computing systems and cause the car to just roll to a stop. That one will disable all of the autonomy components of the car.
Jessica Chobot: Have you ever gotten in an accident with it where you’ve had to use these?
Keenan: No. No and hopefully we never do.
Jessica Chobot: That’s what I meant.
So it’s very loud in here. Where is that sound coming from?
Keenan: Right so we’ve got some liquid cooling for our computer in the back of the trunk and that’s important because when we’re running all these algorithms on our processors they’ll actually overheat unless we have liquid cooling running through on top of them to keep them. So that sound in the back is the fans that’s actually circulating air over the coolant to allow us to keep our computers at a nice comfortable temperature.
Sepehr: All right. Ready to go?
Sepehr: All right. Let’s rock and roll.
Jessica Chobot: Is that a dead raccoon?
Jessica Chobot: Is that real? Did you guys hit that thing?
Jessica Chobot: Are you sure?
Sepehr: So what’s going to happen here is the car is going to go detect a right turn only sign and turn right because of it.
Jessica Chobot: Okay. So this is kind of a make it or break it test, that they’ll have.
Sepehr: Looks all clear we’re about to launch. Maybe I can ask Jessica to flip the car into autonomous mode when we’re ready to go.
Jessica Chobot: All I’m seeing is warning, warning, warning, warning. Geeze.
Keenan: As long as it’s not in air.
Jessica Chobot: Actually yeah, that’s a good point. All right so we’re doing it. Is this it? Autonomous? Self driving?
Sepehr: All right. So, let’s go. Yeah, this is it.
Jessica Chobot: Okay. I don’t want to end up like that raccoon that we saw.
Sepehr: All right are we ready to go?
Keenan: Ready to go.
Jessica Chobot: Yeah. I’m ready.
Sepehr: One. Two. Three. Go.
Jessica Chobot: I hear things. Oh (beep). Oh (beep). It takes off fast and then it slows down real fast. All right.
Keenan: So the reason why it accelerates so quickly in the beginning is that we’re basically telling it, “You need to immediately get to what we’re going which is about 8 miles per hour.” So it will do what you tell it to do. But when we’re driving in a map or doing some more complex autonomous driving tasks we’ll tell it to like nicely ramp up to that speed.
Jessica Chobot: So yeah. Your hands are off the wheel. Foot on the gas.
Keenan: Hands are off.
Jessica Chobot: It’s driving by itself. I love that it turns on your signal. Also I see, so this is where we would turn right wait.
Keenan: Yeah, so that one didn’t work.
Jessica Chobot: That one didn’t work. Okay.
Keenan: We’ll try it again.
Jessica Chobot: It is interesting how quickly I went from, “Yeah I could see myself just chilling looking at my phone while the car’s driving,” to, “Oh my gosh!” Trouble.
Keenan: This is a robot. But you made it into a robot and it’s a big heavy scary robot.
Jessica Chobot: So is it my turn?
Keenan: It is yeah.
Jessica Chobot: All right. Which program are we going to do?
Keenan: So we’re going to do the pedestrian detection this time.
Jessica Chobot: Good. That’s the one I wanted. Awesome. All right, I’ll switch out here.
Sepehr: Do the big switcheroo.
Jessica Chobot: Okay. Okay. Okay. I’m just cutting in here for a second to explain that there isn’t actually a real live human pedestrian. It’s this life size mannequin that they call Fred Buster, he’s on a wheeled cart with a really long rope and someone pulls him in front of the car as if he’s jaywalking. So, to be clear, we’re not at risk of killing somebody but if we were to hit him it would do serious damage.
All right. Buckle up. Because we don’t know how this is going to work.
Keenan: Tighten your restraints. You’ve watched me take manual control, that was only once though.
Jessica Chobot: So I grab the wheel. Jerk it opposite direction. Gently, but strongly press on the brake and then switch to manual.
Keenan: That’s right. Yeah. You’re going to want to hover your foot over the brake pedal instead.
Jessica Chobot: Oh that’s it. That’s the gas? Sorry. It’s just habit.
Supair: What we can start with is flipping the car into autonomous mode. So Jessica are you ready?
Jessica Chobot: I am ready. Oh gosh I’m a little nervous.
Sepehr: We’ll look here.
Jessica Chobot: I think I’m more scared of disappointing you guys and ruining the car than I am of actually of us like getting us hurt.
Keenan: Final check. The vehicle enter autonomous mode.
Sepehr: On three. One. Two. Three.
Jessica Chobot: I’m so excited.
Sepehr: And we’re rolling.
Jessica Chobot: That is weird. Okay. Now I’m calm again.
Keenan: Yeah so same as before. We’re just driving kind of within the cones here and then we’re going to see that pedestrian as we come up to it.
Jessica Chobot: Okay. Making our turn. I see them in the distance.
Keenan: All right. So we should start slowing down.
Jessica Chobot: Slowing down.
Keenan: Now the pedestrian’s going to start going across.
Jessica Chobot: Taking their sweet time when I’ve got somewhere I need to be.
Sepehr: Soon as they’ve crossed the road.
Keenan: Back up to speed.
Jessica Chobot: And we’re going. Yeah all right.
Sepehr: And then they’ve crossed the street.
Jessica Chobot: That’s awesome.
Keenan: All right. Now you can take back control.
Jessica Chobot: Okay.
Keenan: And just jerk the steering wheel. There you go.
Jessica Chobot: Jerk the steering wheel.
Sepehr: And we can switch back to manual mode.
Jessica Chobot: I’m fully braked. Switch to manual.
Jessica Chobot: And we’re good.
Sepehr: Put the car in park.
Jessica Chobot: Put the car in park.
Keenan: Now we have –
Sepehr: That was it.
Jessica Chobot: All right. Yay. I did it. That was super cool.
Sepehr: All right I think I’m just going to walk back.
Jessica Chobot: Nobody wants to drive back with me?
Keenan: I’ll drive.
Jessica Chobot: Seriously.
Sepehr: We could drive back. Keenan you want to walk?
Keenan: I’ll drive.
Jessica Chobot: All right. I’m not walking are you kidding, it’s cold outside.
I survived. I made it. Now the scariest part was probably the acceleration but what surprised me was how much fun it was. So is the hype justified? Will we see autonomous cars take over our streets any day now? As it currently stands based on the people that I talked to and my experience behind the wheel, not quite yet. They’re just not quite there. However, will we see more autonomous features built into our regular cars? You bet and hopefully that will lead to the dream behind all of this which is safer driving for everybody.
That’s AI: Hype vs Reality from Dell Technologies and hey want to see what it looked like while I was in that autonomous car, just check out Delltechnologies.com/hypevreality. Next time on the podcast, are artificially intelligent robots going to take all of our jobs? See you then.