Latest / What's Your Problem? / Why Amazon Built a Spatula-Wielding Robot

Transcript

Why Amazon Built a Spatula-Wielding Robot · What's Your Problem?

• 0:15
  Speaker 1: Pushkin.
• 0:20
  Speaker 2: One of the things I look at in robotics as
• 0:22
  Speaker 2: a big field. There are so many amazing demonstrations of mobility,
• 0:27
  Speaker 2: robots doing backflips, robots running down hills, and that's really
• 0:32
  Speaker 2: impressive to me because I can't do a backflip, or
• 0:35
  Speaker 2: I might trip if I run down the hill. But
• 0:37
  Speaker 2: where the really valuable parts of robotics are going to
• 0:41
  Speaker 2: be are in manipulation. So my kid can take a
• 0:46
  Speaker 2: blueberry out of her cereal bowl because she doesn't want
• 0:48
  Speaker 2: to eat it, and that is an incredibly hard task
• 0:51
  Speaker 2: for a robot. And you don't see any of those demos.
• 0:54
  Speaker 2: And I think we're like kind of inherently programmed as
• 0:58
  Speaker 2: people to like be biased towards the backflip being more
• 1:01
  Speaker 2: impressive and in reality like the business value and the
• 1:04
  Speaker 2: harder thing for the robot is to like take the
• 1:07
  Speaker 2: blueberry out of the cereal bowl.
• 1:15
  Speaker 1: I'm Jacob Goldstein and this is What's Your Problem, the
• 1:17
  Speaker 1: show where I talk to people who are trying to
• 1:19
  Speaker 1: make technological progress. My guest today is Aaron Parness. Aarin
• 1:24
  Speaker 1: spent the earlier part of his career building space robots
• 1:27
  Speaker 1: at NASA's Jet Propulsion Laboratory JPL. Six years ago, he
• 1:32
  Speaker 1: went to work at Amazon. Now, Aaron is a director
• 1:35
  Speaker 1: of Applied Science at Amazon Robotics. I wanted to talk
• 1:39
  Speaker 1: to Erin about a robot arm called Vulcan. He and
• 1:43
  Speaker 1: his team developed Vulcan to do a job that is
• 1:46
  Speaker 1: surprisingly hard for robots to do, taking stuff that gets
• 1:50
  Speaker 1: delivered to Amazon warehouses and putting it onto shelves. In
• 1:55
  Speaker 1: order to solve this problem, Aaron and his team had
• 1:58
  Speaker 1: to build a robot that had a sense of touch,
• 2:01
  Speaker 1: that could deal with complicated, unpredictable situations, and that could
• 2:05
  Speaker 1: look at a shelf and plan out a course of action.
• 2:09
  Speaker 1: As you'll hear in the interview, all of those traits
• 2:11
  Speaker 1: may someday be helpful, not just in stocking shelves in
• 2:14
  Speaker 1: a warehouse, but in doing lots of boring sounding but
• 2:18
  Speaker 1: complicated real world tasks like, for example, taking a blueberry
• 2:23
  Speaker 1: out of a bowl of cereal. To start, I asked
• 2:26
  Speaker 1: Aaron to tell me the problem that Vulcan was designed
• 2:28
  Speaker 1: to solve at Amazon's warehouses.
• 2:31
  Speaker 2: So new inventory comes into the building. You know, trucks
• 2:35
  Speaker 2: pull up and they unload new stuff. We need to
• 2:38
  Speaker 2: store that stuff while it's waiting for someone to click
• 2:40
  Speaker 2: the buy button. We store it in these large fabric bookcases.
• 2:46
  Speaker 2: It's about eight feet tall. It has about forty different
• 2:49
  Speaker 2: shelves on it. It's four sided, so you can store
• 2:53
  Speaker 2: stuff from any of the different faces of the case.
• 2:57
  Speaker 2: What's really interesting is this stuff is randomly stowed, so
• 3:01
  Speaker 2: it's not like all the iPhones are in one shelf.
• 3:04
  Speaker 2: It'll be all different stuff, all mixed together.
• 3:08
  Speaker 1: When you say random, do you mean random or do
• 3:11
  Speaker 1: you mean it would look random to the untrained eye.
• 3:14
  Speaker 2: I mean literally random. Really wherever there is space you
• 3:18
  Speaker 2: can put the item, because.
• 3:20
  Speaker 1: That's what's optimal. It turns out the optimal way to
• 3:22
  Speaker 1: store stuff is random.
• 3:24
  Speaker 2: That's right. Why the stems actually from Jeff Bezos is
• 3:28
  Speaker 2: like original vision, I think, and it's it's incredible. So
• 3:33
  Speaker 2: you want to have the most selection, and you want
• 3:36
  Speaker 2: to have speed of delivery, and you want to have
• 3:39
  Speaker 2: low cost, and that's what the customer wants, right. The
• 3:42
  Speaker 2: customer is using Amazon dot Com because we have selection,
• 3:46
  Speaker 2: we have speed, and we have low cost. In order
• 3:49
  Speaker 2: to achieve that, you have to have these massive warehouses
• 3:53
  Speaker 2: located really close to your customers, and you have a
• 3:56
  Speaker 2: lot of customers in Tokyo, in New York City and
• 3:59
  Speaker 2: San Francisco where real estate's really expensive, so you have
• 4:03
  Speaker 2: to figure out a way to put all of this
• 4:05
  Speaker 2: different stuff in like the densest packing area, you can
• 4:10
  Speaker 2: and have access to it immediately so that you can
• 4:13
  Speaker 2: you can deliver in hours instead of days. And what
• 4:17
  Speaker 2: that means is that random is better than structured. So
• 4:22
  Speaker 2: anywhere there's a space, you can add that item into
• 4:25
  Speaker 2: the inventory, and that means it comes up for sale
• 4:28
  Speaker 2: immediately on the website, and then when someone places in order,
• 4:32
  Speaker 2: you don't have to wait for that iPhone bookcase to
• 4:34
  Speaker 2: make its way all the way across the warehouse. You
• 4:37
  Speaker 2: probably have a thousand iPhones in the warehouse, and whichever
• 4:40
  Speaker 2: one is closest can go to whichever pickstation is eligible,
• 4:44
  Speaker 2: and it ends up being actually substantially faster. So that last.
• 4:49
  Speaker 1: Sentence seems to be the key. The idea is like, yes,
• 4:52
  Speaker 1: given you have whatever a thousand iPhones in the warehouse
• 4:56
  Speaker 1: in the universe where a human had to know where
• 4:57
  Speaker 1: they all were, you'd put them all on one shelf.
• 5:00
  Speaker 1: But you're saying at any given time that means that
• 5:02
  Speaker 1: shelf is probably going to be pretty far away, whereas
• 5:05
  Speaker 1: if you randomly distribute them throughout the shelves and warehouse
• 5:10
  Speaker 1: at any given time, one of those thousand iPhones is
• 5:12
  Speaker 1: probably going to be pretty close to where it needs
• 5:14
  Speaker 1: to be. And because you have a whatever, a computerized
• 5:17
  Speaker 1: system that can keep track of everything all the time,
• 5:19
  Speaker 1: it makes sense to randomly distribute all the things.
• 5:22
  Speaker 2: Yeah, that's exactly right. And it works on the flip
• 5:25
  Speaker 2: side as well. So when you have a new item
• 5:27
  Speaker 2: that's come in, rather than waiting for the shelf that
• 5:30
  Speaker 2: has the right size thing to put the new dog
• 5:33
  Speaker 2: toy in, you just put the dog toy anywhere you
• 5:36
  Speaker 2: can find space for it.
• 5:37
  Speaker 1: H It's like my house. We have a lot of
• 5:41
  Speaker 1: dog toys in my house. Also, Yeah, that's really interesting.
• 5:45
  Speaker 1: It's great for the customer, and that's optimal.
• 5:48
  Speaker 2: It's optimal, and it creates an incredibly difficult environment for robotics. Huh,
• 5:54
  Speaker 2: because now you have to deal with all this clutter.
• 5:57
  Speaker 2: We can have more than a million unique items in
• 6:00
  Speaker 2: one warehouse. Yeah, so it's not like you have a
• 6:03
  Speaker 2: model of each of those items. And we sell more
• 6:06
  Speaker 2: third party items than you know, Amazon owns themselves. Right.
• 6:10
  Speaker 2: We are a platform for third party fulfillment, and so
• 6:14
  Speaker 2: you don't have all the data about all those items,
• 6:17
  Speaker 2: and so you have to handle all this uncertainty, all
• 6:19
  Speaker 2: this clutter, and everything's tightly packed.
• 6:22
  Speaker 1: And so still in most places as a result, when
• 6:26
  Speaker 1: stuff comes into the warehouse every day off a truck
• 6:28
  Speaker 1: to people take the things out of the truck and
• 6:32
  Speaker 1: stick them randomly on shelves wherever they can find space.
• 6:35
  Speaker 1: Is that the system?
• 6:36
  Speaker 2: That is exactly the system, and it's in you know,
• 6:39
  Speaker 2: hundreds of buildings around the world.
• 6:42
  Speaker 1: And just to be clear, I mean it's pretty clear,
• 6:44
  Speaker 1: but just to really put a point on it, why
• 6:48
  Speaker 1: is this a hard environment for robots?
• 6:51
  Speaker 2: Traditional industrial robots do not handle contact well, so like
• 6:57
  Speaker 2: touching their environments, and they don't handle clutter or you know, uncertainty,
• 7:03
  Speaker 2: and so it's hard because to put that last book
• 7:07
  Speaker 2: onto the bookshelf, squeeze that teddy Bear into the just
• 7:12
  Speaker 2: small enough space that it'll fit, you have to push
• 7:15
  Speaker 2: the other stuff around that's already on that bookshelf. And
• 7:18
  Speaker 2: a traditional robot doesn't have sensors, it doesn't even know
• 7:22
  Speaker 2: how to do that. So if you think of like
• 7:24
  Speaker 2: a car manufacturing line, you're like nineteen nineties two thousands,
• 7:28
  Speaker 2: you know, welding robot or loading sheet metal into a press.
• 7:32
  Speaker 2: It's doing all of that only knowing its position in space.
• 7:35
  Speaker 2: So it has no force sensing. If it runs into something,
• 7:38
  Speaker 2: it either is like an emergency stop because it's like broken,
• 7:42
  Speaker 2: or it just smashes that thing and keeps going and
• 7:44
  Speaker 2: it doesn't even know it's smashed anything. It literally has
• 7:47
  Speaker 2: no sensing.
• 7:48
  Speaker 1: That is an incredibly homogeneous environment. Right, It's doing like
• 7:53
  Speaker 1: the exact same thing at a very high level of
• 7:55
  Speaker 1: precision forever one thing.
• 7:57
  Speaker 2: That's exactly right. And so this extension, the fundamental breakthrough
• 8:02
  Speaker 2: for science for robotics manipulation that my team is trying
• 8:05
  Speaker 2: to make is one giving the robot a sense of
• 8:08
  Speaker 2: touch and using that along with site and along with
• 8:11
  Speaker 2: like knowing where your robot is to do meaningfull tasks
• 8:16
  Speaker 2: in like very high contact, high clutter environments. And then
• 8:20
  Speaker 2: there's a brain part. It's also much more difficult to
• 8:23
  Speaker 2: kind of predict how this random assortment of items is
• 8:27
  Speaker 2: going to move or change as you push on it.
• 8:30
  Speaker 2: And so there's an AI piece, there's a brain piece
• 8:33
  Speaker 2: that's saying this item will fit in that bin. This
• 8:36
  Speaker 2: is actually one of the most frustrating things when you
• 8:37
  Speaker 2: try and do the job yourself. I'm like an optimist.
• 8:40
  Speaker 2: I'm always oh, yeah, this will fit. And I go
• 8:42
  Speaker 2: up there and I try and play Tetris and I
• 8:44
  Speaker 2: try and rearrange the shelf and like, it clearly isn't
• 8:47
  Speaker 2: going to fit. And then I've wasted thirty seconds or
• 8:49
  Speaker 2: forty seconds and I have to try something else.
• 8:53
  Speaker 1: That's a good statement of the problem. Well, like when
• 8:56
  Speaker 1: did you come onto the scene?
• 8:58
  Speaker 2: So I was working on some other stuff and there
• 9:01
  Speaker 2: was a recent PhD that had joined our team. He was,
• 9:06
  Speaker 2: you know, one year out of school something like this,
• 9:08
  Speaker 2: and he says, I'm going to go try and solve
• 9:10
  Speaker 2: stowing items into these bookshelves. And my thought was, Oh,
• 9:15
  Speaker 2: how naive, Like the real world is going to teach
• 9:18
  Speaker 2: this new grad. That's just way too hard a problem
• 9:21
  Speaker 2: for robotics to solve. But I was helping him because
• 9:24
  Speaker 2: it's fun, right, Like you like to work on hard
• 9:26
  Speaker 2: problems when you're a researcher. And he was a very
• 9:29
  Speaker 2: nice guy, and so I was, you know, helping him,
• 9:32
  Speaker 2: but never thought it was going to work. And there
• 9:34
  Speaker 2: were a couple of kind of moments where we made
• 9:36
  Speaker 2: these simplifications that turn the problem from I have to
• 9:41
  Speaker 2: try and do every possible game of Tetris that a
• 9:45
  Speaker 2: person can do into a problem where you're like, oh,
• 9:50
  Speaker 2: it's not that this is never going to work, it's
• 9:52
  Speaker 2: that this is the future, Like this is robotics two
• 9:54
  Speaker 2: point zero, Like this is I have to work on this.
• 9:57
  Speaker 2: I can't do anything else anymore. I'm like, I'm all
• 10:00
  Speaker 2: in on this problem.
• 10:01
  Speaker 1: Tell me about one of those simplifications one of those moments.
• 10:05
  Speaker 2: It was the gripper is one. The design, the mechanic
• 10:09
  Speaker 2: design of the robotic hand was actually a big breakthrough.
• 10:13
  Speaker 2: And when we started, we were trying to push items
• 10:17
  Speaker 2: with the item we were gripping. So imagine you're pinching
• 10:19
  Speaker 2: a book and you're trying to use that book to
• 10:22
  Speaker 2: like push this dog toy over to the side.
• 10:25
  Speaker 1: I see, So you want to put the book in
• 10:27
  Speaker 1: a bin. Yeah, dog toys in the way. So you're like, okay,
• 10:30
  Speaker 1: pick up the book and use the book kind of
• 10:32
  Speaker 1: like a brush to sweep the dog toy out of
• 10:34
  Speaker 1: the way.
• 10:35
  Speaker 2: Okay, And I say, okay, like I understand, but it's
• 10:37
  Speaker 2: never going to work. What if you don't have a book.
• 10:39
  Speaker 2: What if you have a T shirt? Yeah, what if
• 10:41
  Speaker 2: you have an iPhone and it's very expensive? Are you
• 10:44
  Speaker 2: going to actually want to start pushing on stuff with
• 10:46
  Speaker 2: the phone? And so we came up with this strategy
• 10:50
  Speaker 2: to have like a spatula that would extend into the
• 10:53
  Speaker 2: bin and you'd push everything with this spatula that was
• 10:56
  Speaker 2: part of your hand. So imagine like you're like Wolverine
• 11:00
  Speaker 2: and you can shoot out, you know, but instead of
• 11:02
  Speaker 2: like the Adamantium claws, you're shooting out a spatula.
• 11:07
  Speaker 1: So it's like a pincher grip. A little spatula shoots
• 11:10
  Speaker 1: forward out of the pincher grip.
• 11:12
  Speaker 2: Is the thing that's right.
• 11:13
  Speaker 1: It's so simple when you put it that way. I mean,
• 11:16
  Speaker 1: I'm sure making it was not low tech, but it
• 11:18
  Speaker 1: sounds very like. It's not like some crazy AI thing.
• 11:22
  Speaker 1: It's like just what if there was another little thing
• 11:24
  Speaker 1: that came out and push stuff out of the way.
• 11:26
  Speaker 2: But those ideas are like the really powerful ones when
• 11:29
  Speaker 2: you have a simple, elegant solution and you're like, okay,
• 11:34
  Speaker 2: that could work. That's different than like a five fingered
• 11:38
  Speaker 2: hand that has twenty five motors embedded in it. Yeah,
• 11:42
  Speaker 2: like Oh, it's just dispatulate.
• 11:44
  Speaker 1: Fingers are famously difficult. Why didn't anybody think of it before?
• 11:49
  Speaker 2: So we had been working on it as a company
• 11:51
  Speaker 2: back to the Amazon Picking challenge, which was, you know,
• 11:54
  Speaker 2: twenty fifteen. But I think a lot of robotics researchers
• 11:59
  Speaker 2: like myself, were scared that this problem was just too hard.
• 12:02
  Speaker 2: There was easier things to go try and work on,
• 12:05
  Speaker 2: and there were a couple of simplifications, so using this
• 12:08
  Speaker 2: spatulo as one, and then you watch people do the
• 12:11
  Speaker 2: task and you realize they're kind of doing the same
• 12:14
  Speaker 2: strategies over and over again. It's like insert dispatchela on
• 12:19
  Speaker 2: the D and sweep to one side. For this kind
• 12:23
  Speaker 2: of page turn mechanism. Something's fallen over and you need
• 12:26
  Speaker 2: to sort of flip it back up to make space.
• 12:30
  Speaker 1: So you put this spatula underneath it and flip the
• 12:33
  Speaker 1: thing up ninety degreees basically, Yeah.
• 12:35
  Speaker 2: And you realize that accounts for like ninety percent of
• 12:38
  Speaker 2: the actions you do when you try and stow into
• 12:41
  Speaker 2: these pins.
• 12:42
  Speaker 1: And did you figure that out by watching people stow.
• 12:45
  Speaker 2: We did and doing it yourself.
• 12:48
  Speaker 1: How much stowing did you do couple of days? Okay,
• 12:52
  Speaker 1: it's a hard job, thousands of items probably, I imagine.
• 12:55
  Speaker 2: Yeah, exactly. And we tried to wear go pro cameras
• 12:59
  Speaker 2: on our heads so we could look at the videos later,
• 13:00
  Speaker 2: which turns out as a recipe for motion sickness. It's
• 13:03
  Speaker 2: very difficult to watch those videos, but you go and
• 13:06
  Speaker 2: you do it, and you build up this intuition. And
• 13:08
  Speaker 2: I I think the other piece of the problem that
• 13:11
  Speaker 2: made it tractable and made me this like huge believer
• 13:14
  Speaker 2: was recognizing we didn't have to get to one hundred percent.
• 13:17
  Speaker 2: So in some automation scenarios, you have to solve the
• 13:21
  Speaker 2: whole problem, and if you don't, you have nothing, so
• 13:24
  Speaker 2: like landing on the moon. And what we realized was
• 13:27
  Speaker 2: there was a way to like make the business logic
• 13:30
  Speaker 2: work that the robot could handle seventy five percent of
• 13:34
  Speaker 2: the stoves and it just had to not make a
• 13:36
  Speaker 2: mess and let like work alongside people to do the
• 13:41
  Speaker 2: other twenty five percent, and the sum of the parts
• 13:44
  Speaker 2: is actually much better than either all robots or all
• 13:47
  Speaker 2: employees would be on their own. And making that realization
• 13:52
  Speaker 2: all of a sudden meant that it could be a
• 13:54
  Speaker 2: two or three year project instead of a twenty year
• 13:57
  Speaker 2: project because chasing this long tail. You know, we have
• 14:01
  Speaker 2: a million unique items in the building, but we also
• 14:04
  Speaker 2: process a million items per day. So I have a
• 14:06
  Speaker 2: phrase like, if something goes wrong one in a million,
• 14:09
  Speaker 2: it happens every day in every Amazon building. And to
• 14:13
  Speaker 2: try and solve all of those is it is a
• 14:15
  Speaker 2: twenty year problem.
• 14:17
  Speaker 1: I feel like that part of the solution generalizes in
• 14:20
  Speaker 1: a really nice way, right, Like I mean, I guess
• 14:23
  Speaker 1: the eighty twenty problem is a sort of cliche. But
• 14:27
  Speaker 1: the idea that like, oh, if you think of the
• 14:30
  Speaker 1: problem the right way, it's like, no, we don't have
• 14:32
  Speaker 1: to build a robot that does it every time. Before
• 14:34
  Speaker 1: we build a robot that does it seventy five percent
• 14:36
  Speaker 1: of the time. That is a huge efficiency gain and
• 14:38
  Speaker 1: maybe the optimal point on the curve. Right yep, if
• 14:42
  Speaker 1: the robot is doing everything, you're working too hard to
• 14:45
  Speaker 1: make the robot work.
• 14:46
  Speaker 2: Probably exactly that.
• 14:49
  Speaker 1: So, Okay, so you have these two big ideas. Do
• 14:53
  Speaker 1: you want to tell me the sort of story of
• 14:54
  Speaker 1: making it work? You want to tell me how it works.
• 14:57
  Speaker 2: We've been running six of these robots at a warehouse
• 15:00
  Speaker 2: in Spokane, Washington, Okay, since November of last year and
• 15:06
  Speaker 2: so we've done over half a million stows it this point.
• 15:10
  Speaker 2: We also have another product that's picking those items out
• 15:12
  Speaker 2: of the bins, and so that's my team in Germany,
• 15:16
  Speaker 2: and so we have a warehouse in Homburg where we've
• 15:19
  Speaker 2: been picking items. And picking is a slightly harder problem
• 15:22
  Speaker 2: in some ways because you have to identify the item.
• 15:25
  Speaker 2: So for stow, you have to identify free space. It's
• 15:28
  Speaker 2: either occupied or you can make space to put the
• 15:32
  Speaker 2: next item in. For pick, you want to make sure
• 15:34
  Speaker 2: I get you the red T shirt, not the red sweatpants,
• 15:36
  Speaker 2: or I get you the Harry Potter volume two and
• 15:39
  Speaker 2: not Sapiens or some other book.
• 15:41
  Speaker 1: Tell me how it works. Let's do this stewing first,
• 15:44
  Speaker 1: since that's what we've been talking about. So there's this
• 15:46
  Speaker 1: warehouse in Spokane where this robot that you built is
• 15:49
  Speaker 1: in use. Like what happens there? A truck pulls in
• 15:53
  Speaker 1: and then what happens.
• 15:54
  Speaker 2: The way the system works is one of these pods,
• 15:57
  Speaker 2: one of these bookcases pulls up to the station, so
• 16:00
  Speaker 2: it pulls in front of the robot. We have stereo
• 16:04
  Speaker 2: camera towers and so we're looking with the eyes first,
• 16:07
  Speaker 2: and we are creating a three D representation of the scene.
• 16:12
  Speaker 2: So we're modeling, you know, all the items that are
• 16:16
  Speaker 2: in the in the pod already. But the really interesting
• 16:20
  Speaker 2: part is we're actually predicting on top of that, how
• 16:23
  Speaker 2: we can move those items around to make more empty space,
• 16:28
  Speaker 2: how can we squeeze more stuff in. So it's not
• 16:31
  Speaker 2: just identifying vacant space. You have to predict where you
• 16:35
  Speaker 2: can make that vacant space by pushing stuff with this spatula. Okay,
• 16:40
  Speaker 2: then we do this matching algorithm. So we have about
• 16:43
  Speaker 2: forty or fifty items waiting for us to stow, and
• 16:46
  Speaker 2: so we have a variety of stuff, and we're matching
• 16:49
  Speaker 2: those forty or fifty items to the thirty ish shelves
• 16:53
  Speaker 2: that are in front of the robot. Which items should
• 16:56
  Speaker 2: go where, and then how do we make that space?
• 16:59
  Speaker 2: And so that's where a lot of the AI in
• 17:01
  Speaker 2: the system is active and operating. It's predicting success, it's
• 17:05
  Speaker 2: minimizing risk, it's trying to optimize for a bunch of
• 17:08
  Speaker 2: different parameters. Once we've made that selection, we grasp the item,
• 17:14
  Speaker 2: so that item we've selected for putting into the given
• 17:18
  Speaker 2: shelf passes into our hand and our hand is two
• 17:21
  Speaker 2: conveyor belt paddles, so you can think of it kind
• 17:24
  Speaker 2: of like a Panini press, like a George Foreman grill.
• 17:27
  Speaker 2: It is a George Foreman grill where each side has
• 17:30
  Speaker 2: a conveyor built into it.
• 17:32
  Speaker 1: Like a little belt, Like just a little belt going around.
• 17:35
  Speaker 2: That's right. So each face of the grill, the top
• 17:37
  Speaker 2: face and the bottom face have a conveyor belt. And
• 17:40
  Speaker 2: that's important because you can control the pose of the
• 17:43
  Speaker 2: item and you can feed it into the bin rather
• 17:47
  Speaker 2: than like throwing it into the bin. One of the
• 17:50
  Speaker 2: early versions we had kind of dropped it and tried
• 17:53
  Speaker 2: to punch it to put the item into the bin,
• 17:55
  Speaker 2: and that predictably failed in a lot.
• 17:57
  Speaker 1: Of Well you say predictably now, but if you try
• 18:00
  Speaker 1: to it wasn't predictable.
• 18:02
  Speaker 2: Yeah. Yeah. I'm a huge believer in iterative design, and
• 18:05
  Speaker 2: so we try and build early, build often build and
• 18:09
  Speaker 2: learn from those builds. So it's actually really important to
• 18:13
  Speaker 2: keep sixed off pose control of the item. So you
• 18:17
  Speaker 2: want to make sure the item isn't rotating as you
• 18:20
  Speaker 2: shoot it out. You want to make sure that you
• 18:22
  Speaker 2: keep the orientation of the item because it's fitting tightly,
• 18:26
  Speaker 2: so you don't want it to run into the bookshelf
• 18:28
  Speaker 2: above it or below it, or the items thirty in there.
• 18:31
  Speaker 2: Yeah yeah, yeah. We started by trying to shoot it out,
• 18:34
  Speaker 2: and then we had all kinds of problems when it
• 18:36
  Speaker 2: would like collide with stuff and fall on the floor.
• 18:39
  Speaker 2: The worst case is, you know, you would shoot it out,
• 18:42
  Speaker 2: it would bounce off the back of the back of
• 18:44
  Speaker 2: the bookcase and then come back and hit you in
• 18:46
  Speaker 2: the face or hit you in your.
• 18:47
  Speaker 1: Did that happen?
• 18:48
  Speaker 2: Yeah? Oh yeah, good.
• 18:49
  Speaker 1: That's robot comedy, yesh, ro but physical comedy.
• 18:52
  Speaker 2: Yeah.
• 18:53
  Speaker 1: So, and that's it.
• 18:54
  Speaker 2: That's the stow process. And we want to do that
• 18:57
  Speaker 2: a few hundred times an hour, and we want to
• 18:59
  Speaker 2: do it on the top shelves of those bookcases where. Yeah.
• 19:03
  Speaker 2: That's one of the ways we are really complimentary to
• 19:06
  Speaker 2: the employees is if the robots can do the top shelves,
• 19:09
  Speaker 2: it saves a lot of ergonomic tasks. It allows the
• 19:13
  Speaker 2: employees to work in their power zone, like you know,
• 19:17
  Speaker 2: shoulder level. That makes them faster too. So if you
• 19:20
  Speaker 2: put robots in, people get faster at the job.
• 19:22
  Speaker 1: I mean, presumably as the robot gets better, it'll also
• 19:26
  Speaker 1: be better at putting things on the middle shelf.
• 19:28
  Speaker 2: Right, Well, there's this like sweet spot.
• 19:29
  Speaker 1: The robot's going to get better faster than people will
• 19:31
  Speaker 1: get better.
• 19:33
  Speaker 2: Yeah, we want the robots to be as good as
• 19:34
  Speaker 2: it can and not chase one hundred percent. We don't
• 19:37
  Speaker 2: really believe in one hundred percent automation. We want to
• 19:40
  Speaker 2: find that sweet spot where we're maximizing product.
• 19:43
  Speaker 1: Mean, the sweet spot's going to keep moving, right, The
• 19:45
  Speaker 1: robot's going to get better and better and be able
• 19:47
  Speaker 1: to do more and more faster and faster.
• 19:49
  Speaker 2: Presumably, and my science team's actually really excited about that.
• 19:52
  Speaker 2: As you get more and more data. So we have
• 19:55
  Speaker 2: five hundred thousand stows that we've done so far, but
• 19:58
  Speaker 2: when we get to five hundred million STOs, you can
• 20:00
  Speaker 2: leverage some of these techniques to start learning the motions
• 20:03
  Speaker 2: and learning some of these strategies and refining them to
• 20:06
  Speaker 2: be specific to the item that you are molding in
• 20:09
  Speaker 2: your hand. There's a lot of opportunity as you get
• 20:12
  Speaker 2: more and more data.
• 20:14
  Speaker 1: Well, right, so we haven't really I mean, you mentioned
• 20:16
  Speaker 1: the software side, the AI side, but we haven't really
• 20:19
  Speaker 1: talked about it, and it is I mean, in talking
• 20:22
  Speaker 1: to other people working on robotics. It's plainly a data
• 20:27
  Speaker 1: game because there's no Internet of the physical world, right,
• 20:30
  Speaker 1: because large language models work so well, because there's this
• 20:33
  Speaker 1: huge data set and everybody is trying to get data
• 20:36
  Speaker 1: from the physical world, and you seem very well positioned
• 20:40
  Speaker 1: to get a lot of data from the physical world.
• 20:43
  Speaker 2: I think that's true. So one of the joys of
• 20:46
  Speaker 2: being a roboticist at Amazon is all the data that
• 20:48
  Speaker 2: we have access to. But I will push back a
• 20:51
  Speaker 2: little bit that it's just a data problem. It's a
• 20:54
  Speaker 2: highly debated topic. Some people in the world believe that
• 20:57
  Speaker 2: you can apply the same sort of transformer architectures that
• 21:02
  Speaker 2: work so well for search and so well for natural
• 21:05
  Speaker 2: language processing and apply those to robotics. If we only
• 21:08
  Speaker 2: had the data. I would not put myself in that camp.
• 21:12
  Speaker 2: I am not a believer that all we need is
• 21:15
  Speaker 2: more torque data from robotic grippers and will solve it.
• 21:19
  Speaker 2: Natural language is already tokenized in a way that's very
• 21:23
  Speaker 2: amenable to those methods, and language and search are also
• 21:28
  Speaker 2: very tolerant of sloppiness. So you and I can have
• 21:33
  Speaker 2: a conversation. I don't have to get every single word correct,
• 21:36
  Speaker 2: but if you mess up a torque on a gripper,
• 21:38
  Speaker 2: you can crush your iPhone, or you can sort of
• 21:41
  Speaker 2: smash something else that's there, or drop something, or just
• 21:45
  Speaker 2: fail the task. And that's because you have physics and
• 21:48
  Speaker 2: this nonlinear, very sort of difficult to model real world
• 21:54
  Speaker 2: that these robots have to interact with. And so I
• 21:57
  Speaker 2: think those techniques certainly accelerate us in a lot of places,
• 22:01
  Speaker 2: but they don't just solve the problem. I think we
• 22:04
  Speaker 2: need all of the rest of robotics, like hardware design
• 22:08
  Speaker 2: and classical control theory to solve those problems.
• 22:12
  Speaker 1: Compelling. Although you did start this part of the conversation,
• 22:15
  Speaker 1: you brought it up by saying, the science team is
• 22:18
  Speaker 1: really excited for what the model's going to learn once
• 22:21
  Speaker 1: you have hundreds of millions of stoves.
• 22:24
  Speaker 2: That's right, and that both things are true.
• 22:26
  Speaker 1: I know, yes, plainly, we're just talking about sort of
• 22:28
  Speaker 1: the margins, right, what is true? At what margins? I mean?
• 22:31
  Speaker 1: I did wonder as I was reading about this, you know,
• 22:34
  Speaker 1: I thought of AWS of Amazon Web Services, which of course,
• 22:37
  Speaker 1: like was an internal Amazon thing that at some point
• 22:40
  Speaker 1: Amazon was like, oh, maybe other people would find this
• 22:42
  Speaker 1: service useful, And now it's a giant part of Amazon's business,
• 22:45
  Speaker 1: and so I wondered, like, are you building Amazon robotics services.
• 22:51
  Speaker 2: Yet not today? There's so much value that we can
• 22:55
  Speaker 2: provide to our fulfillment business that we are one hundred
• 23:00
  Speaker 2: percent focused on that. Certainly as a roboticist, though, I
• 23:04
  Speaker 2: take great joy that the work we're doing is advancing
• 23:08
  Speaker 2: the field of robots, and so it's definitely like in
• 23:11
  Speaker 2: the makes my job better that we're advancing the state
• 23:14
  Speaker 2: of the art. But from a business perspective, it's all
• 23:17
  Speaker 2: hands on making the fulfillment process better for Amazon dot Com.
• 23:25
  Speaker 1: We'll be back in just a minute. I think I
• 23:40
  Speaker 1: read you say that you're building a foundation model of items.
• 23:45
  Speaker 1: Is that right? And I sort of know what that means,
• 23:48
  Speaker 1: But tell me what that means when you say that.
• 23:50
  Speaker 2: So, when a robot handles an item, it would do
• 23:54
  Speaker 2: better if it takes into account the properties of that item.
• 23:58
  Speaker 2: So if you're trying to hand a bowling ball to someone,
• 24:02
  Speaker 2: you should do that in a different way than if
• 24:04
  Speaker 2: you're handing them a bouncy ball or a light bulb.
• 24:07
  Speaker 2: At its core, a foundation model for items is simply
• 24:11
  Speaker 2: a model that encodes all of those attributes of an item.
• 24:15
  Speaker 2: And makes them available to the robotic systems that are
• 24:19
  Speaker 2: going to use it. And one of the things that
• 24:21
  Speaker 2: makes it a foundation model instead of just you know,
• 24:23
  Speaker 2: some custom bespoke thing is that you can transfer it
• 24:26
  Speaker 2: across lots of different applications. So if it's you know, stowing,
• 24:30
  Speaker 2: you can use it. If you're packing it into a
• 24:32
  Speaker 2: delivery box, you can use it. If you're putting it
• 24:35
  Speaker 2: onto a shelf in a physical store like for grocery
• 24:38
  Speaker 2: or whole foods or something, you can use it. And
• 24:40
  Speaker 2: so that like commonality across applications is one of the
• 24:43
  Speaker 2: things that's important.
• 24:45
  Speaker 1: Is part of the notion there that like the model
• 24:49
  Speaker 1: would allow a robot to sort of look at some
• 24:52
  Speaker 1: novel item and make a reasonable inference about the properties
• 24:56
  Speaker 1: of that item.
• 24:57
  Speaker 2: Yeah, absolutely that. And the other thing that's a little
• 25:00
  Speaker 2: non intuitive is that by understanding how to handle that
• 25:05
  Speaker 2: item in all those different applications a grocery a you know,
• 25:10
  Speaker 2: stowing picking, you get better at the individual application. So
• 25:14
  Speaker 2: by training on all of this data across these different domains,
• 25:18
  Speaker 2: you actually get better at the individual task that your
• 25:22
  Speaker 2: specific robot is trying to do. Doesn't like it takes
• 25:26
  Speaker 2: a while to like understand that it is not intuitive.
• 25:29
  Speaker 1: SayMore, what do you mean, like, I don't know that
• 25:31
  Speaker 1: I fully get it.
• 25:32
  Speaker 2: Understanding how an item behaves when you gift wrap it,
• 25:36
  Speaker 2: Uh huh shouldn't really inform how it's going to behave
• 25:40
  Speaker 2: when you're picking it off of a bookshelf.
• 25:42
  Speaker 1: Oh, I mean yes, it should, right, Like if you
• 25:45
  Speaker 1: think of like a whatever, a stuffed animal versus a book. Yeah,
• 25:49
  Speaker 1: maybe that's too easy of a case, but like if
• 25:52
  Speaker 1: a thing is squishy or rigid, that seems like as
• 25:56
  Speaker 1: a human being, I feel like we sort of port
• 25:59
  Speaker 1: that knowledge from one use case to another, right.
• 26:02
  Speaker 2: Yeah, it's a good point. And maybe that's because we
• 26:04
  Speaker 2: are inherently sort of we think and manipulate items in
• 26:08
  Speaker 2: the world more similarly to how these foundation models do.
• 26:11
  Speaker 2: But ten years ago it was totally not the case.
• 26:15
  Speaker 2: You would train your model in a very narrow domain,
• 26:17
  Speaker 2: and if you gave it data from some other domain,
• 26:21
  Speaker 2: it would kind of corrupt the results that you had,
• 26:23
  Speaker 2: and so you were very careful to curate all the
• 26:27
  Speaker 2: data that you were using to be very specific to
• 26:29
  Speaker 2: the task that you wanted it to do, and that
• 26:32
  Speaker 2: made the performance better. But it also meant that the
• 26:35
  Speaker 2: model you had was only good at that one very
• 26:37
  Speaker 2: narrow thing.
• 26:38
  Speaker 1: It was why we were always so far from the
• 26:40
  Speaker 1: general purpose robot. Yeah, because, as you're describing it, trying
• 26:43
  Speaker 1: to make a robot do more than one thing just
• 26:46
  Speaker 1: meant it couldn't even do.
• 26:47
  Speaker 2: One We couldn't even do one thing, and so you're
• 26:49
  Speaker 2: putting all your effort into making it do that one
• 26:51
  Speaker 2: thing just a little better. I think there's another really
• 26:56
  Speaker 2: interesting piece here, which is our team, the Vulcan team
• 27:01
  Speaker 2: at Amazon, is trying to use touch and vision together,
• 27:06
  Speaker 2: and that is how people interact with the world. That's
• 27:09
  Speaker 2: how people manipulate the world. And so the example I
• 27:14
  Speaker 2: like to give is picking a coin up off a
• 27:16
  Speaker 2: table ten years ago, when a robot would try and
• 27:19
  Speaker 2: do that, I mean it's impossible, Like, robot can't pick
• 27:22
  Speaker 2: a coin up off table, it's too hard a task.
• 27:25
  Speaker 2: My five year old can pick a coin up off
• 27:27
  Speaker 2: the table in half a second without you noticing. Well,
• 27:31
  Speaker 2: the reason is your strategy. So when you pick a
• 27:34
  Speaker 2: coin up off the table, you actually don't grasp the coin.
• 27:38
  Speaker 2: You go and you touch the table and then you
• 27:40
  Speaker 2: slide your fingers along the surface of the table until
• 27:44
  Speaker 2: you feel the coin, and when you feel the coin,
• 27:47
  Speaker 2: that's your trigger to like rotate it up into a grasp.
• 27:51
  Speaker 2: You're not going to some millimeter precision the way your
• 27:54
  Speaker 2: grandfather's robot and the welding line would do. And you're
• 27:59
  Speaker 2: not just watching with your eyes. You're using your eyes
• 28:01
  Speaker 2: and your fingertips both your.
• 28:04
  Speaker 1: Sense of touch. Yes, sense of touch is central to.
• 28:07
  Speaker 2: Pick and we are trying to do those same kind
• 28:11
  Speaker 2: of behaviors that are not only reacting to touch, but
• 28:16
  Speaker 2: planning for touch. So the same way you plan to
• 28:19
  Speaker 2: touch the table first, we plan to put our spatula
• 28:23
  Speaker 2: against the side of the bookcase before we try to
• 28:26
  Speaker 2: extend it in between this you know, small gap between
• 28:30
  Speaker 2: the T shirt and the bag and the side of
• 28:32
  Speaker 2: the bookcase. So we are building our plans and our
• 28:36
  Speaker 2: controllers around having sight and touch.
• 28:40
  Speaker 1: I mean when you say touch in the context of
• 28:42
  Speaker 1: the robot, does that mean that it is getting feedback
• 28:45
  Speaker 1: from the stuff it is coming into contact with? And
• 28:49
  Speaker 1: is that novel? And how does that work?
• 28:51
  Speaker 2: So the sensor is a force torque sensor. It looks
• 28:54
  Speaker 2: like a hockey puck and a thousand times a second,
• 28:58
  Speaker 2: it's telling you what it feels in the six degrees
• 29:03
  Speaker 2: of freedom, So up, up and down is one, left
• 29:05
  Speaker 2: and right is two, in and out as three, and
• 29:08
  Speaker 2: then you've got roll pitch y'ah as the three torques.
• 29:12
  Speaker 2: So a thousand times per second, you're sensing, you're feeling
• 29:16
  Speaker 2: what the world is pushing on you with, and we
• 29:20
  Speaker 2: use that to control the motion but also to plan
• 29:22
  Speaker 2: the motion.
• 29:24
  Speaker 1: When you say plan the motion, it's like, given the
• 29:27
  Speaker 1: sense of touch that is happening right now, what should
• 29:29
  Speaker 1: I do next? Yep.
• 29:31
  Speaker 2: So in a like high level view, it's like touch
• 29:33
  Speaker 2: the table first, slide along the table while keeping you know,
• 29:38
  Speaker 2: sort of one pound of force pushing into the table
• 29:42
  Speaker 2: until you touch the coin, and then you know, rotate.
• 29:45
  Speaker 2: That's at a high level, but then even at a
• 29:48
  Speaker 2: low level, the thousand times per second is so that
• 29:51
  Speaker 2: as you slide your fingers along the table, you're sort
• 29:54
  Speaker 2: of maintaining that accurate force.
• 29:57
  Speaker 1: Yeah. Or like if you're putting a thing on the shelf,
• 30:00
  Speaker 1: you can sort of tell if you've pushed it too
• 30:02
  Speaker 1: far because the shelf is pushing back.
• 30:04
  Speaker 2: At you exactly, or you can tell it slipping and
• 30:06
  Speaker 2: you're about to like push over the top of it,
• 30:09
  Speaker 2: so you can like, oh, it's about to fall over,
• 30:12
  Speaker 2: so I can react. And those dynamics are happening at
• 30:16
  Speaker 2: tens or hundreds of hurt since you need to sense
• 30:18
  Speaker 2: them at a thousand hurts.
• 30:21
  Speaker 1: What's the frontier right now for stewing? What are you
• 30:24
  Speaker 1: trying to figure out?
• 30:26
  Speaker 2: One of the things is getting the fullness of those
• 30:31
  Speaker 2: bins all the way up to where they are today,
• 30:34
  Speaker 2: so as a person you can pack those bins really,
• 30:37
  Speaker 2: really densely, and so the robot's close but not quite
• 30:41
  Speaker 2: as good as a person is today at getting as
• 30:45
  Speaker 2: much stuff into the bookcase as it can. That's one frontier,
• 30:49
  Speaker 2: and that is because one we're conservative, like our brain
• 30:53
  Speaker 2: is telling us there's no space when really there is space.
• 30:57
  Speaker 2: And two it's because those motions are not sophisticated enough yet.
• 31:01
  Speaker 2: So we're trying to improve our video streaming. We're trying
• 31:04
  Speaker 2: to get the eyes better to help as well as
• 31:07
  Speaker 2: those low level touch centers to those behaviors to be better.
• 31:12
  Speaker 2: So that's one of the major frontiers. The other one
• 31:16
  Speaker 2: is the negative. The robot makes too many mistakes, so
• 31:20
  Speaker 2: defects and exception handling are so important in robotic systems,
• 31:26
  Speaker 2: and this is another thing I think the world on
• 31:27
  Speaker 2: the Internet doesn't appreciate enough. Like you can do a
• 31:30
  Speaker 2: demo and a happy path. Hey, it worked once. I
• 31:33
  Speaker 2: can submit a paper to a conference, or I can
• 31:35
  Speaker 2: put a cool video on YouTube. That's great. You have
• 31:38
  Speaker 2: a demo. To have a product, you have to make
• 31:41
  Speaker 2: sure it's working, you know, ninety nine percent of the time,
• 31:44
  Speaker 2: or ninety nine and a half percent, or you know
• 31:47
  Speaker 2: in some cases four nines or five nines. And a
• 31:50
  Speaker 2: lot of the work you have to do is to
• 31:52
  Speaker 2: recover and handle those rare exceptions or prevent or recover
• 31:57
  Speaker 2: from those defects. And so the robot still drops too
• 32:00
  Speaker 2: much stuff on the floor. One of our frontiers is
• 32:03
  Speaker 2: not dropping crap on the floor, like, we need to
• 32:06
  Speaker 2: get about three times better at that.
• 32:08
  Speaker 1: Umly, the robot is already skipping some universe of items
• 32:13
  Speaker 1: that the robot can't handle.
• 32:14
  Speaker 2: Yeah, and so we need to get smarter about which
• 32:17
  Speaker 2: items we skip and which items we take. We also
• 32:19
  Speaker 2: need to get better at inserting those items in such
• 32:23
  Speaker 2: a way that they're not going to fall back out?
• 32:25
  Speaker 1: What items are particularly hard for the robot?
• 32:28
  Speaker 2: So tight fitting items are the hardest.
• 32:31
  Speaker 1: Uh huh. And so that's not the nature of the item,
• 32:34
  Speaker 1: but the nature of the particular relationship between the item and.
• 32:37
  Speaker 2: The shelf exactly.
• 32:38
  Speaker 1: Yeah, Like, is there a kind of thing that the
• 32:42
  Speaker 1: robot just can't do because of its shape or something.
• 32:46
  Speaker 2: There is a particular rubber fish that we really hate.
• 32:52
  Speaker 2: It's a dog toy.
• 32:53
  Speaker 1: It's floppy. Is that what?
• 32:55
  Speaker 2: Sticky?
• 32:56
  Speaker 1: Oh? Sticky? Interesting? Yeah, And they don't put it in
• 33:00
  Speaker 1: a bot, Nope, they just send you the sticky fish.
• 33:04
  Speaker 2: Yeah, and it sort of gets hung up on whenever
• 33:07
  Speaker 2: it makes contact. It doesn't slide, it like it wants
• 33:09
  Speaker 2: to rotate about whatever it's made contact with. And so
• 33:13
  Speaker 2: there's this particular dog toy and so we use it.
• 33:15
  Speaker 2: We've bought like fifty of them and now we have
• 33:18
  Speaker 2: them in the lab and this is like our diabolical
• 33:21
  Speaker 2: item set.
• 33:22
  Speaker 1: Is that a term of art that diabolical?
• 33:24
  Speaker 2: I don't know, Yeah, it's our term of art. Yeah.
• 33:27
  Speaker 2: Also bagged items where the bag is really loose. So
• 33:31
  Speaker 2: imagine having like a T shirt in a bag, but
• 33:34
  Speaker 2: the bag is like twice as big as the T shirt.
• 33:37
  Speaker 1: Floppy? Is that the floppy problem?
• 33:40
  Speaker 2: Floppy but also transparent, so sometimes you can see through
• 33:44
  Speaker 2: the bag.
• 33:44
  Speaker 1: Or so the robot gets confused about is the bag
• 33:48
  Speaker 1: the item.
• 33:49
  Speaker 2: Yeah or not? Sometimes you want one and sometimes you
• 33:51
  Speaker 2: want the other. So like if it's just floppy plastic bag,
• 33:55
  Speaker 2: it probably will fit. Like if I just push it
• 33:57
  Speaker 2: into the bin, the bag is going to conform and
• 34:00
  Speaker 2: slide in, but you can't be sure about that. You know,
• 34:03
  Speaker 2: you get into a bunch of those edge cases that
• 34:04
  Speaker 2: are in that long tail of being robust.
• 34:07
  Speaker 1: I mean, it's interesting, right because the robot is dealing
• 34:10
  Speaker 1: with this sort of human optimized world. Like it reminds
• 34:13
  Speaker 1: me of the way I think is it. Ikea designs
• 34:16
  Speaker 1: its furniture to fit optimally on a pallette, so you
• 34:20
  Speaker 1: can fit the most of them, like not just the
• 34:21
  Speaker 1: flat pack, but like in more subtle ways. And can
• 34:25
  Speaker 1: you imagine that there is some shift in the world
• 34:28
  Speaker 1: where I mean, obviously you're trying to make the robot better,
• 34:31
  Speaker 1: but also people are trying to make things work better
• 34:33
  Speaker 1: for the robot.
• 34:34
  Speaker 2: Yes. Absolutely, And there is a different team within Amazon
• 34:38
  Speaker 2: that's imagining a future world and future bookcases that are
• 34:43
  Speaker 2: friendly for robots.
• 34:45
  Speaker 1: Uh huh.
• 34:46
  Speaker 2: However, there are currently five million of those bookshelves in
• 34:53
  Speaker 2: warehouses holding inventory that's for sale on Amazon dot Com.
• 34:58
  Speaker 2: And so it's a really really big lift to go
• 35:02
  Speaker 2: replace all of those books.
• 35:04
  Speaker 1: Shehlds interesting. So it's a whole other team that's just like,
• 35:07
  Speaker 1: let's imagine the you know, a much more robot centric warehouse. Yeah,
• 35:13
  Speaker 1: those guys like you don't even talk to them. They're
• 35:15
  Speaker 1: just off in their own.
• 35:16
  Speaker 2: I mean, they're friends, but yeah, we are spacing very
• 35:19
  Speaker 2: different problems. And so we took a tenant very early on.
• 35:24
  Speaker 2: It's like, the world exists, the robot needs to perform
• 35:29
  Speaker 2: in the world as it exists. And this team they
• 35:32
  Speaker 2: get their green field, so they get to think of
• 35:34
  Speaker 2: like a new field. We are a brownfield, meaning we
• 35:38
  Speaker 2: have to retrofit into these existing buildings. You know, we
• 35:41
  Speaker 2: have like ten year leases on some of these building
• 35:43
  Speaker 2: They're going to be there for a long long time.
• 35:45
  Speaker 1: And then somebody else is out there. So they're building
• 35:47
  Speaker 1: a whole other kind of robot. Your robot is optimized
• 35:50
  Speaker 1: for the world today, and somebody else is building a
• 35:52
  Speaker 1: robot for the robot world.
• 35:53
  Speaker 2: That's right. I love that they have a building that
• 35:56
  Speaker 2: they've built in Louisiana. It's in Shreveport, Louisiana. It has
• 36:00
  Speaker 2: ten times the number of robots that traditional building has.
• 36:05
  Speaker 2: It's a completely reimagined way of fulfilling your order. It
• 36:09
  Speaker 2: also has a lot of people still working in those buildings,
• 36:12
  Speaker 2: but they're working in maintenance and robotics quarterbacks jobs, and
• 36:17
  Speaker 2: so they're higher skilled. And so we have a bunch
• 36:19
  Speaker 2: of programs that are trying to transition our very talented
• 36:23
  Speaker 2: workforce into the jobs of the future. One of the
• 36:26
  Speaker 2: things I really like to say is, you don't need
• 36:28
  Speaker 2: a college degree to work in robotics. At Amazon. It's
• 36:31
  Speaker 2: about twenty twenty five percent of my team doesn't have
• 36:33
  Speaker 2: a college degree but are enormously valuable. Like some of
• 36:37
  Speaker 2: our top ten people on our team are those people.
• 36:41
  Speaker 1: That facility in Treeport. Is it live? Like, is real
• 36:45
  Speaker 1: stuff going in and real orders going out?
• 36:47
  Speaker 2: Yeah, it's live. We could follow up with exactly the date,
• 36:49
  Speaker 2: but it's been up for about a year. I think, so.
• 36:52
  Speaker 1: Interesting thing like that, Well, I would be interested in
• 36:55
  Speaker 1: talking to your counterpart there as well. That show would
• 36:57
  Speaker 1: pair interestingly with this show. So okay, let's talk about
• 37:01
  Speaker 1: the rest of the process. You know, the rest of
• 37:04
  Speaker 1: what's going on in the warehouse and where else you're
• 37:06
  Speaker 1: working on robots. So the piece we've been talking about
• 37:10
  Speaker 1: this whole time is getting stuff as it comes in
• 37:13
  Speaker 1: from the truck onto the shelf, which naively I wouldn't
• 37:17
  Speaker 1: even think of that part, but it turns out to
• 37:19
  Speaker 1: me this scrape big problem. What are the other pieces?
• 37:23
  Speaker 2: What's interesting is the science we're building giving robots a
• 37:26
  Speaker 2: sense of touch has applicability and lots and lots of
• 37:30
  Speaker 2: places across that whole chain. Anytime the robots need to
• 37:34
  Speaker 2: be physically interacting, like contacting, touching items is a good
• 37:40
  Speaker 2: place for our core technology. So if we're packing four
• 37:44
  Speaker 2: items into a box because we want to send you
• 37:46
  Speaker 2: the four things you bought in one shipment, not in
• 37:48
  Speaker 2: four separate packages, you need to touch the box. You
• 37:52
  Speaker 2: need to touch the other items that are already in
• 37:53
  Speaker 2: the box. You need to play that game of tetris.
• 37:56
  Speaker 1: Yes, I mean it's a stowing problem again, right, I
• 37:59
  Speaker 1: know it's called packing, but it's a version of that same.
• 38:01
  Speaker 2: Problem, that's right, And those problems recur over and over again.
• 38:04
  Speaker 2: So getting all of the packages, all of the cardboard
• 38:07
  Speaker 2: boxes and paper mailers into a cart that can go
• 38:11
  Speaker 2: onto the back of the truck, that is a stowing
• 38:13
  Speaker 2: problem in the cart.
• 38:15
  Speaker 1: Putting things in a thing, yeah, is a great, big
• 38:18
  Speaker 1: problem in many ways.
• 38:19
  Speaker 2: But you can also expand to think about grocery. So
• 38:22
  Speaker 2: if you order produce, you don't want your grandfather's welding
• 38:28
  Speaker 2: robot handling your peaches. It's gonna smash them, like, you
• 38:31
  Speaker 2: need a robot with a sense of touch. If you
• 38:33
  Speaker 2: think about household tasks, if you want a robot, you know,
• 38:37
  Speaker 2: picking up your kid's toys or dealing with laundry, like,
• 38:41
  Speaker 2: those robots need to have a sense of touch. They're
• 38:43
  Speaker 2: physically interacting in a dexterous way with the world. And
• 38:47
  Speaker 2: so one of the things that we're so excited about
• 38:49
  Speaker 2: not only these big applications for stowing and picking off
• 38:52
  Speaker 2: of you know, these bookcases, but everything that gets unlocked
• 38:57
  Speaker 2: once the robot has that sense of touch.
• 39:00
  Speaker 1: When you talk that way, it feels like a beyond
• 39:05
  Speaker 1: what is typically considered Amazon kind of thing. It seems
• 39:08
  Speaker 1: like a thing either Amazon's going to get into lots
• 39:11
  Speaker 1: of other sort of non retail businesses or license the
• 39:16
  Speaker 1: technology or sell you know, robotic touch as a service
• 39:19
  Speaker 1: or whatever.
• 39:20
  Speaker 2: Yeah, I think there are probably five or ten applications
• 39:25
  Speaker 2: in how we process orders today that are all within
• 39:29
  Speaker 2: the warehouses and delivery stations, and those are my first
• 39:33
  Speaker 2: hill to climb. Then we do have a consumer robotics team.
• 39:37
  Speaker 2: So there was a cool robot we released called Astro.
• 39:41
  Speaker 2: It didn't have any manipulation capabilities, right, It would drive
• 39:44
  Speaker 2: around your house. It had a camera on a mast
• 39:46
  Speaker 2: that would extend up and down. You could talk to
• 39:48
  Speaker 2: it the way you can talk to an Alexa device.
• 39:51
  Speaker 2: The future versions of those robots are going to want
• 39:53
  Speaker 2: to do more useful things, and so they're going to
• 39:56
  Speaker 2: need this kind of underlying technology, and so that's a
• 39:58
  Speaker 2: business opportunity in the long term. You know, that's not
• 40:02
  Speaker 2: a thing my team is focused on now, but I
• 40:05
  Speaker 2: get excited about it when I think about what we unlock.
• 40:12
  Speaker 1: We'll be back in a minute with the lightning round.
• 40:25
  Speaker 1: Let's do a lightning round. If you listen to the show,
• 40:28
  Speaker 1: you have a sense of what this is. Tell me
• 40:30
  Speaker 1: about the last time you were in zero gravity.
• 40:33
  Speaker 2: I flew an experiment to try and drill into rocks,
• 40:38
  Speaker 2: which was going to be applied to asteroids. And of course,
• 40:42
  Speaker 2: if you're drilling into an asteroid, any amount you're pushing
• 40:45
  Speaker 2: into the rock is pushing you back off into space
• 40:47
  Speaker 2: because asteroids have almost zero gravity.
• 40:51
  Speaker 1: Right, so you gotta have somebody push it on the
• 40:54
  Speaker 1: other side. How do you solve that? What do you do?
• 40:56
  Speaker 1: You grab it? How do you even do that?
• 40:58
  Speaker 2: Says my passion for robot hands. We built a robot
• 41:02
  Speaker 2: hand that would grab the rock with a bunch of claws.
• 41:05
  Speaker 2: I think it had a thousand claws, and the claws
• 41:08
  Speaker 2: were actually fish hooks. So imagine a bunch of fish
• 41:10
  Speaker 2: hooks grabbing onto a rock to react the force of
• 41:13
  Speaker 2: pushing a drill bit down the center.
• 41:15
  Speaker 1: Did it work?
• 41:17
  Speaker 2: It did work, but it only worked on rocks that
• 41:19
  Speaker 2: were pretty rough, that had a lot of spots for
• 41:22
  Speaker 2: the fish hooks to grab. But it turns out asteroids
• 41:25
  Speaker 2: are really rough. Most of the smooth rocks you find
• 41:27
  Speaker 2: on Earth have been processed by liquid water or ice,
• 41:31
  Speaker 2: and that's not happening on asteroids. No liquid water.
• 41:35
  Speaker 1: And so this was on the on the plane, on
• 41:37
  Speaker 1: that NASA plane that flies what is it? Yeah, vomit
• 41:41
  Speaker 1: combat fine curve basically, yeah, what was it like?
• 41:45
  Speaker 2: The vomit comets actually very zen. So when you're in
• 41:48
  Speaker 2: zero gravity, when you're floating, it's like very peaceful. It's
• 41:52
  Speaker 2: when you're in double gravity, where you're the bottom of
• 41:55
  Speaker 2: the parabola and you're like being glued and pushed against
• 41:58
  Speaker 2: the floor. If you like turn your head very quickly,
• 42:00
  Speaker 2: that's where you get like into serious trouble. And so
• 42:03
  Speaker 2: the trick is just to like go into your zone
• 42:06
  Speaker 2: for the bottom of the parabola and then you've become
• 42:08
  Speaker 2: like very free and zen like in the zero G portion.
• 42:13
  Speaker 1: You think you'll ever go to space?
• 42:15
  Speaker 2: No, I think now that I have three kids, I
• 42:18
  Speaker 2: think I'm landlocked.
• 42:20
  Speaker 1: You seem a little bit sad about that. Does everybody
• 42:24
  Speaker 1: who works at JPL kind of want to go to space?
• 42:26
  Speaker 2: Yes, everybody that works at JPL, I think does think
• 42:30
  Speaker 2: about going to space. I think what makes me sad
• 42:32
  Speaker 2: is we could be doing so much more at building
• 42:36
  Speaker 2: civilization out into space, at the scientific exploration of all
• 42:41
  Speaker 2: of the interesting places in space, and I think we're
• 42:45
  Speaker 2: kind of tripping ourselves up in a couple of places
• 42:48
  Speaker 2: as a species. I wish we would get unblocked and
• 42:51
  Speaker 2: get some of that eagerness you see some of the
• 42:52
  Speaker 2: private investment, Like we're doing well in rockets, but we're
• 42:56
  Speaker 2: not yet doing well in the spacecraft and the scientific
• 42:59
  Speaker 2: instruments and the pieces that have to fly on top
• 43:03
  Speaker 2: of the rockets.
• 43:04
  Speaker 1: When you say we're tripping ourselves up in a couple places,
• 43:07
  Speaker 1: in what places? Like, what do you mean?
• 43:10
  Speaker 2: I think we became very conservative, like our risk posture
• 43:13
  Speaker 2: about going to space. We stopped treating it as this
• 43:17
  Speaker 2: like very dangerous activity and tried to make it extremely safe,
• 43:20
  Speaker 2: and that slowed us down.
• 43:22
  Speaker 1: To bring back to cowboys a little bit.
• 43:24
  Speaker 2: Yeah. Interesting, Yeah, and then there's a lot of bureaucracy,
• 43:28
  Speaker 2: of course, that built up over fifty years. I still
• 43:32
  Speaker 2: have very optimistic there's a lot of smart people working
• 43:35
  Speaker 2: in that area and a lot of exciting things happening,
• 43:39
  Speaker 2: So we're going to get through it.
• 43:46
  Speaker 1: Aaron Parness is a director of Applied Science at Amazon Robotics.
• 43:52
  Speaker 1: Please email us at problem at Pushkin dot fm. We
• 43:55
  Speaker 1: are always looking for new guests for the show. Today's
• 43:59
  Speaker 1: show was produced by Trinomnino and Gabriel Hunter Chang. It
• 44:03
  Speaker 1: was edited by Alexander Garretton and engineered by Sarah Bruguerrett.
• 44:07
  Speaker 1: I'm Jacob Goldstein, and we'll be back next week with
• 44:09
  Speaker 1: another episode of What's Your Problem.