Marc Raibert: Boston Dynamics | MIT Artificial Intelligence (AI)

00:00:00.000 | Welcome back to 6S099, Artificial General Intelligence.

00:00:05.000 | Today, we have Mark Raybert.

00:00:09.600 | (audience applauding)

00:00:13.840 | He is the, he really doesn't need an introduction,

00:00:19.840 | but we'll give him one anyway.

00:00:21.760 | He's the founder and CEO of Boston Dynamics.

00:00:24.820 | He founded the CMU Leg Lab in 1980,

00:00:27.360 | the MIT Leg Lab in 1986.

00:00:29.720 | Boston Dynamics in 1992.

00:00:32.000 | He and his team have developed

00:00:33.920 | some of the most amazing robots ever built,

00:00:36.840 | including Big Dog, Atlas, Handle, Spot, Spot Mini.

00:00:41.240 | These robots move with the agility, dexterity,

00:00:43.840 | and even grace that rivals

00:00:45.520 | and often supersedes that of human movement.

00:00:48.440 | He continues to inspire us

00:00:50.120 | with what robots are capable of achieving in the real world

00:00:53.120 | and what physical form future intelligence systems may take

00:00:57.080 | as they become integrated in our daily lives.

00:00:59.760 | So please give Mark a warm welcome.

00:01:02.480 | (audience applauding)

00:01:03.520 | - Thank you, thank you.

00:01:05.440 | This is our grand mission, our aspiration,

00:01:12.640 | which is to make robots that are equal to

00:01:15.320 | or greater than people and animals.

00:01:18.680 | And it's a daunting mission

00:01:20.720 | because we're so good at things.

00:01:24.020 | It seems effortless.

00:01:26.400 | I'm standing here knowing questions

00:01:28.560 | that I could stand here like this, but a lot's going on.

00:01:31.620 | I can manipulate things.

00:01:33.920 | I can pick up this water or I could reach in my pocket

00:01:36.600 | and use my hands with all the sensors in my hands

00:01:39.640 | and coordinate that.

00:01:41.120 | And maybe most of all, our perception systems.

00:01:43.740 | You know, this audience has, what is it,

00:01:45.760 | 250 people in it or something?

00:01:47.720 | And I can look out there and see every one of you

00:01:51.520 | stabilized in space even while I'm moving.

00:01:54.240 | It's just astounding.

00:01:56.240 | And robots aren't there yet.

00:01:59.280 | But I think they can be.

00:02:00.840 | And our goal is to keep chipping away to try and get there.

00:02:05.840 | Before I get started, I wanted to say that

00:02:07.720 | I got my start in robotics here at MIT.

00:02:11.560 | I was a graduate student.

00:02:12.800 | I was in what was then called the psychology department,

00:02:15.800 | the brain and cognitive science department.

00:02:17.820 | But I was taking an IAP course, just like you are,

00:02:20.400 | and it might have been exactly this time of year,

00:02:24.900 | when I followed my professor, Bertolt Horn, back.

00:02:29.100 | I was jabbering away at him, asking him some questions

00:02:32.200 | about this or that, and we walked back to Tech Square,

00:02:35.160 | which is where the AI lab was in those days.

00:02:37.860 | And we went up to the ninth floor,

00:02:39.960 | and Russell Novsker, who was a guy working in the lab,

00:02:43.960 | had an arm all taken apart on the table.

00:02:47.280 | It was like 1,000 pieces.

00:02:49.200 | And I was a roboticist from that day on.

00:02:53.320 | I didn't switch my major, but I got Bertolt to be an advisor.

00:02:58.320 | I found a topic that had to do with robotics.

00:03:02.980 | In that time, it was a manipulation thing,

00:03:04.920 | but eventually became a legged thing.

00:03:07.980 | And it was amazing, and I've never looked back.

00:03:12.980 | So here are some animals doing things

00:03:16.280 | that are very exciting, climbing around

00:03:19.140 | on very rough terrain, very short-footed.

00:03:23.280 | Using a mixture of their proprioception and their vision.

00:03:26.960 | And look, there's even a baby that probably

00:03:29.160 | is only a couple of months old,

00:03:30.720 | has no trouble at all doing these things.

00:03:34.440 | And look at the grace and suppleness

00:03:38.320 | and the fearlessness of these animals.

00:03:42.340 | It's amazing.

00:03:43.180 | Here are animals running for their lives.

00:03:47.800 | The gazelle is trying to stay alive

00:03:50.700 | for the next 10 minutes, and the cheetah's

00:03:52.880 | trying to get a meal so that it can stay alive in general.

00:03:56.920 | Sorry.

00:03:57.760 | And even people can do things that are breathtaking.

00:04:04.800 | I assume all of you were out this morning

00:04:06.760 | getting a little exercise, climbing up

00:04:08.360 | the green building that we're in now,

00:04:09.840 | and maybe the other places around here.

00:04:12.240 | It's funny, I bumped into some people

00:04:14.520 | when I came into the room who were climbing the stairways.

00:04:17.480 | I think they were going on a trek up and down them.

00:04:19.380 | But I'd like to see the going outside.

00:04:22.840 | So probably most of you have seen this video.

00:04:25.880 | This is sort of where we were after about 10 years of work

00:04:30.880 | attempting to make machines that could work out

00:04:35.080 | in the real world that were dynamically stabilized.

00:04:38.040 | Dynamics is a big deal for our company and for what we do.

00:04:42.480 | So some active sensing and control

00:04:46.040 | and understanding of the physics.

00:04:51.600 | This robot has all its control on board,

00:04:54.000 | and it has reflexes and sensors.

00:04:56.780 | And this is an extension to a 1,000-pound robot

00:05:00.440 | that could carry about 400 pounds of payload.

00:05:03.360 | And we took it all around the United States

00:05:05.120 | testing it in various situations.

00:05:08.300 | Here we have it in Virginia doing some bushwhacking.

00:05:13.140 | It's actually following a person,

00:05:15.120 | but the person is only in and out of view intermittently,

00:05:17.400 | so it has to be able to keep track of where the person is

00:05:20.340 | and deal with that.

00:05:22.040 | And then back in good old Boston, 10 inches of snow,

00:05:26.120 | just marches right up the hill.

00:05:27.720 | Here's the cheetah.

00:05:30.960 | Now, you know, MIT has its own cheetah.

00:05:33.160 | This is our cheetah.

00:05:34.440 | People know Songbe, who's doing the MIT cheetah.

00:05:40.760 | A very dynamic machine.

00:05:46.800 | And basically an experiment in seeing

00:05:48.640 | how fast we could make something like this run.

00:05:51.480 | Although you notice it's on a parking lot,

00:05:53.540 | so it wasn't doing this on rough terrain.

00:05:56.180 | And getting both the efficiency and the speed

00:05:59.340 | in the context of a machine that also can do rough terrain

00:06:03.100 | is a really big challenge that remains with us.

00:06:06.300 | So this is just a snapshot of most of the robots

00:06:11.840 | we've built at Boston Dynamics over the years.

00:06:14.320 | And I'm not gonna talk about most of them.

00:06:16.820 | I'm just gonna talk about the last four.

00:06:19.640 | These are all robots that we developed

00:06:21.940 | since we've been part of Google,

00:06:23.580 | which has been the last four years.

00:06:25.980 | Spot Mini, there's a Spot Mini on the floor here,

00:06:28.260 | which we'll demo a little later.

00:06:29.980 | Spot, Atlas, the Humanoid.

00:06:32.860 | Some of you may remember the Humanoid

00:06:34.960 | that we used in the DARPA Robotics Challenge.

00:06:37.220 | You have one here.

00:06:38.740 | And then Handle, which is our latest version.

00:06:41.900 | So I'll have a few words to say about each of them.

00:06:45.240 | So we had been developing Big Dog

00:06:48.500 | and those other quadrupeds that I showed you

00:06:50.300 | for quite some number of years.

00:06:52.100 | And it was amazing for us to find out

00:06:54.180 | when we did this project on Spot,

00:06:57.460 | this is the predecessor to that,

00:06:59.660 | that there was still a lot to learn.

00:07:01.480 | And we kind of revolutionized the hardware design

00:07:05.200 | and how the control worked

00:07:07.020 | and got a much higher level of rough terrain performance.

00:07:11.220 | And part of the solution to that

00:07:12.700 | was to be able to decompose the control problem

00:07:16.180 | into many separate controllers

00:07:19.600 | that operated in different regions of state space.

00:07:22.460 | And that allowed us both to have programmers

00:07:25.260 | work on multiple solutions to the problem

00:07:27.780 | and also have the complexity of each controller simplified

00:07:34.300 | by only having to operate

00:07:35.900 | in a small part of the dynamic space.

00:07:39.100 | Here we've added a robot arm

00:07:41.540 | to the previous version of Spot.

00:07:44.420 | And we believe that mobile manipulation,

00:07:48.540 | that is manipulation when you can move the base,

00:07:51.060 | is really a powerful way of doing things.

00:07:55.520 | Now this is probably the most important thing

00:07:58.300 | I wanna show tonight.

00:07:59.500 | And I'll show it three different times.

00:08:01.580 | The idea that we don't build controllers

00:08:03.740 | that just do one particular thing,

00:08:06.500 | but that they can determine where they are

00:08:08.340 | in the execution, here's another version of it,

00:08:10.620 | and then adjust what they're doing

00:08:12.980 | in order to compensate for disturbances in the real world.

00:08:17.220 | I know this class is about AI and probably autonomy.

00:08:20.940 | I think that one of the most important ways

00:08:26.580 | of getting to autonomy

00:08:27.980 | is to have the low-level implementations

00:08:30.100 | very robust to disturbances

00:08:32.440 | so that the planning steps don't have to take care

00:08:34.780 | of all the minutia of the details of the real world.

00:08:38.940 | And that's what we've been trying to do there.

00:08:42.020 | We've been experimenting with doing delivery

00:08:44.180 | of packages to people's houses.

00:08:46.260 | These are all employees of Boston Dynamics,

00:08:48.660 | so we didn't go crashing ordinary people's houses.

00:08:52.180 | And it turns out that there's just lots

00:08:55.860 | of different kinds of stairways and entranceways.

00:08:58.680 | And the robot's doing very well.

00:09:02.740 | We're up to something between 70 and 80%

00:09:06.020 | of the kinds of stairs and access places we encounter

00:09:11.020 | after collecting data

00:09:14.220 | and making improvements and adjustments.

00:09:17.020 | So I'm gonna say a few philosophical things

00:09:23.160 | or approach things.

00:09:24.380 | A lot of people think that this is the model

00:09:29.140 | of how a computer and a robot interact.

00:09:33.700 | That is, there's the robot,

00:09:36.080 | which is hardware and electronics and sensors.

00:09:39.900 | And then there's a computer.

00:09:41.980 | And that the computer listens to the sensors on the robot

00:09:45.940 | and then gives it instructions and tells it what to do.

00:09:49.100 | And while I think that's actually going on,

00:09:52.380 | there's another part to the story,

00:09:54.200 | which is that the physical world

00:09:55.860 | is also giving instructions to the robot.

00:09:59.420 | And that means that the energies stored in the robot,

00:10:02.900 | either in its springs or in its motion,

00:10:07.900 | those are all important determinants

00:10:09.980 | of how the robot's gonna behave in the time coming forward.

00:10:14.260 | And so we like to think in terms of designing

00:10:17.140 | the hardware of the robot, the physical world,

00:10:20.420 | and the computer all as one holistic thing

00:10:24.300 | where we take into account those interactions.

00:10:27.160 | Sometimes we call this a harmony.

00:10:30.620 | A harmonic system is one usually where you have energy

00:10:33.740 | oscillating back and forth.

00:10:35.740 | Almost all legged locomotion has some amount

00:10:38.500 | of harmony going on between potential energy of elevation,

00:10:43.340 | potential energy of elastic deformation,

00:10:46.340 | kinetic energy of motion,

00:10:48.420 | and inverted pendulum things and the like.

00:10:53.280 | Another part of our approach we call

00:10:56.920 | build it, break it, fix it.

00:10:58.600 | Now I have friends who build their robots

00:11:02.800 | and are so into the beauty of what they've created

00:11:07.800 | that they kind of put it on an altar

00:11:10.800 | and afraid of actually hurting it.

00:11:14.720 | So, and in fact, I even have friends here at MIT

00:11:18.060 | that have done that, where they have a gold-plated robot

00:11:20.700 | and they're afraid of taking it out into the world.

00:11:23.840 | I mean, we're just the opposite.

00:11:26.080 | Every one of our robots is designed to get bashed to bits.

00:11:30.160 | We have staff who are there to fix the robot

00:11:33.040 | on a daily basis as we break it.

00:11:35.960 | And I think doing that, build it, break it, fix it,

00:11:38.560 | means that we're able to learn a lot

00:11:40.800 | from the actual physical robot working in the world.

00:11:44.240 | And we can use that knowledge in order to improve the robot,

00:11:48.000 | improve its behavior, and we really like to go around

00:11:50.940 | that loop as quickly as we can, early in the process,

00:11:55.620 | and do it as many times as we can.

00:12:00.340 | So here's what build it, break it, fix it looks like.

00:12:02.940 | This is in Somerville.

00:12:08.180 | Our engineers, this is a Boston driver.

00:12:18.700 | (audience laughing)

00:12:21.700 | Now this robot's supposed to be using its visual system

00:12:30.140 | to avoid the trees.

00:12:32.180 | I think it might have fallen in love with this tree.

00:12:40.220 | We don't purposely give them any emotion,

00:12:42.060 | but boy, it's hard not to see that.

00:12:45.340 | And here's the first time we tested the push response

00:12:48.900 | to this robot.

00:12:49.860 | And you-- - That's your new guy's car.

00:12:55.300 | - Did you hear that?

00:12:56.140 | That's the new guy's car.

00:12:57.340 | So some guy who just started that week, Trent,

00:13:00.760 | had $5,000, which we paid for,

00:13:04.780 | in repairs to his vintage BMW.

00:13:08.000 | So the last thing sort of about philosophy

00:13:13.660 | is long-term versus short-term.

00:13:15.680 | Our company is 25 years old,

00:13:19.220 | and we've mostly been a long-term robotics company.

00:13:23.660 | That is, we're interested in moving the boundary forward

00:13:27.340 | in what robots can do, and we're interested in making it

00:13:32.300 | so robots meet the dream of being the equal

00:13:35.500 | or better than people and animals.

00:13:37.180 | But now we've started--

00:13:39.260 | (microphone thudding)

00:13:42.420 | Malfunction.

00:13:45.260 | (microphone thudding)

00:13:50.060 | Okay.

00:13:55.420 | We still on?

00:13:56.740 | Can you hear me?

00:13:57.580 | But lately, we've started to realize

00:14:03.340 | that some of our robots have enough capability

00:14:06.260 | that maybe it's time to try and productize them,

00:14:08.660 | and we will learn a lot by doing that, too.

00:14:10.580 | One of the things, for instance, that I've always claimed

00:14:13.500 | is that we always spent a lot of money

00:14:17.880 | on building our robots and used that

00:14:20.840 | as a competitive advantage.

00:14:22.340 | That is, DARPA was a frequent funder of us.

00:14:24.860 | DARPA always said, "Let's take money out of the equation

00:14:27.460 | "and just figure out how to get the solution

00:14:31.920 | "and then worry about getting the cost down later."

00:14:34.560 | So I've always assumed and argued

00:14:36.820 | that once we get a robot doing things

00:14:38.400 | that are interesting, then you can go and redesign it

00:14:41.060 | to make it lower cost.

00:14:42.960 | Well, we're gonna test that, because it might not be true.

00:14:46.060 | It might be that we've designed ourselves

00:14:47.620 | into an expensive corner and that it might be too late.

00:14:51.900 | But the robot that we'll show in a little bit

00:14:54.060 | is much significantly cost-reduced

00:14:56.420 | from the prototype of it, and it'll be interesting

00:14:58.700 | to see whether we can get it down

00:15:00.260 | to the kind of prices that are useful.

00:15:03.980 | So this is just a picture, again,

00:15:05.540 | of the idea of aiming long but also aiming short.

00:15:10.480 | And I think it's gonna be a challenge

00:15:13.400 | to see whether we can keep the culture of the company

00:15:16.460 | to support both of these directions,

00:15:18.220 | because people manufacturing stuff

00:15:20.740 | have a different mindset than people trying

00:15:22.700 | to get out to the future horizons,

00:15:26.260 | and it's gonna be a challenge

00:15:28.780 | to keep both those kinds of people happy.

00:15:34.340 | Here's some of the things that,

00:15:36.620 | some of the kinds of applications you can look at

00:15:39.820 | based on modest technical capabilities.

00:15:44.820 | I've shown mobility and manipulation here,

00:15:47.700 | but you could put cost, reliability,

00:15:49.620 | there's many things that could be on these axes.

00:15:52.260 | You know, entertainment, like robots in theme parks

00:15:55.260 | is something that I think we should be able to do.

00:15:58.420 | I already talked about home delivery.

00:16:00.380 | I think home delivery is waiting for self-driving cars

00:16:03.500 | to get all the way there, self-driving trucks,

00:16:05.620 | and once they do, then we will be working

00:16:08.600 | on getting it from the truck to the home.

00:16:11.200 | Logistics, there's about a trillion boxes

00:16:17.620 | moved every year around the world,

00:16:20.480 | and most of it's done by hand,

00:16:22.720 | and so there's really a big opportunity

00:16:24.420 | to having robots help with moving those trillion boxes.

00:16:28.020 | Security, which could mean either commercial security,

00:16:32.240 | like patrolling your shopping center,

00:16:34.900 | or the military type security.

00:16:37.540 | Construction, a lot of people have been coming to us

00:16:41.660 | with their construction applications asking if we can help,

00:16:45.820 | and you know, I'm not gonna talk about it now,

00:16:47.260 | but if afterwards you wanna ask about that,

00:16:48.820 | I can fill you in a little more,

00:16:51.140 | and I think this is really the ultimate

00:16:53.940 | home run application,

00:16:56.100 | care for the elderly and the disabled.

00:17:00.460 | I used to say that I wanted to have robots

00:17:03.840 | that would help me take care of my parents and older people,

00:17:08.840 | but I realize now that it's probably gonna be my children

00:17:13.180 | using them to help take care of me,

00:17:15.680 | but you guys, you're all a little bit younger,

00:17:18.540 | and I think there'll be a time when you could use robots

00:17:23.540 | to help make your parents' lives better.

00:17:26.560 | Now, some of you may think that your parents

00:17:28.240 | don't want that, but I think it's a complex question.

00:17:31.720 | We've seen some surveys that say that, you know,

00:17:34.680 | people aren't totally happy with the idea of their kids

00:17:37.460 | taking care of them on a moment by moment basis,

00:17:40.100 | and I think there's gonna be an opportunity

00:17:41.980 | for doing something, but technically,

00:17:43.680 | this is still a ways off, it's a tough thing.

00:17:46.400 | Okay, let's get back to the robots.

00:17:49.080 | Spot Mini is a robot that weighs about 60 pounds.

00:17:53.200 | That previous Spot weighed about 180 pounds.

00:17:55.720 | This one weighs about 60 pounds, and here's some anatomy.

00:18:00.720 | It's got an arm with five degrees of freedom.

00:18:04.880 | Each leg has three degrees of freedom.

00:18:06.920 | It's got about a 500 watt hour battery.

00:18:13.020 | Batteries for these things are a challenge,

00:18:15.840 | because, you know, you can have consumer products

00:18:19.240 | like electric drills that have relatively small batteries,

00:18:22.640 | and then there's electric cars that have big batteries,

00:18:24.960 | and there's not really much available in between,

00:18:26.880 | so we've done a lot of work on the battery technology

00:18:29.560 | for these things to make them safe and reliable

00:18:32.080 | and hot swappable and things like that.

00:18:34.920 | Then there's radios and computers.

00:18:36.860 | The previous version had three quad core i7s.

00:18:41.760 | This one has two.

00:18:43.080 | We're trying to cut back on the cost.

00:18:45.660 | And then there can be some sensors,

00:18:50.040 | lidars, stereo, and the like.

00:18:54.400 | (audience member speaking faintly)

00:18:57.040 | So you can see, Spot Mini's a little bit smaller than Spot.

00:19:00.200 | This isn't a real house, and those aren't real people.

00:19:09.360 | Those are engineers.

00:19:10.760 | (audience laughing)

00:19:12.280 | This is inside of a warehouse we have out on 128

00:19:16.120 | where we've built a house.

00:19:17.640 | You can see that we don't mind scuffing up the walls here,

00:19:20.480 | and there is a lot of scuffing that happens.

00:19:22.960 | Some of you may recognize Zach Tuchowski,

00:19:25.000 | who's an MIT alum, and he's, again, disturbing the robot.

00:19:29.440 | Here, the robot's using its vision

00:19:31.280 | to do some stepping stone type operations,

00:19:35.360 | and I think Gene is gonna talk a little bit more

00:19:37.260 | about this in a couple of minutes.

00:19:38.960 | And here's a case where it's doing stepping stones

00:19:41.960 | on real stones, and it's keeping its balance,

00:19:45.080 | figuring out where to put the feet.

00:19:50.720 | And again, this robot only has stereo looking out the front,

00:19:54.160 | whereas this one has stereo on all four sides.

00:19:57.920 | Now, one of the cool things about animals

00:20:05.800 | is that they have these stabilization mechanisms

00:20:08.320 | for their sensors.

00:20:09.200 | That was a real chicken.

00:20:10.760 | No robotics involved.

00:20:12.520 | And here's our attempt to show that this robot

00:20:15.560 | can do the same sort of thing.

00:20:19.120 | And if you think about it, when you're manipulating,

00:20:21.400 | you really want the hand to be stabilized in space,

00:20:24.080 | and so you'd like the body to be able

00:20:26.480 | to kind of coordinate with the hand

00:20:28.240 | so that you can concentrate on what the real world task is.

00:20:33.240 | (audience laughing)

00:20:34.200 | Oh, man.

00:20:35.720 | You guys didn't pick up the banana peels, huh?

00:20:38.000 | (audience laughing)

00:20:41.000 | So our concept for the Spot Mini product

00:20:52.760 | is to make a platform.

00:20:54.280 | It's sort of the, we're thinking of it

00:20:56.080 | like the Android of robots.

00:20:58.400 | So with Android, there's a hardware platform,

00:21:01.360 | and then there's a software platform,

00:21:04.000 | and then developers, third party developers,

00:21:07.440 | create their own apps that use the platform.

00:21:10.200 | So we've made this spot so that there's a place

00:21:14.640 | to mount hardware on the robot,

00:21:16.560 | but there's also an API to program it through,

00:21:19.440 | and then there's a facility to have additional computing

00:21:22.480 | external to the robot, and we're working with third parties

00:21:27.040 | to develop their own applications that run on the platform.

00:21:35.760 | This is a video that we haven't been able

00:21:38.660 | to release publicly.

00:21:39.940 | Please don't tape it and show it,

00:21:42.320 | because I'll explain later if you wanna know why not,

00:21:47.160 | but this is just revealing that we do have an arm

00:21:50.000 | on the new version of Spot.

00:21:52.120 | It's using a camera in the hand to find the door handle.

00:21:58.080 | This robot doesn't weigh a lot,

00:22:04.040 | so it has to use tricks to keep the door open,

00:22:06.840 | so that's why it puts its foot in the door.

00:22:09.040 | (audience laughing)

00:22:12.040 | (audience applauding)

00:22:19.200 | And here again, we wanna show that we've made

00:22:23.800 | the solution robust to certain kinds of disturbances.

00:22:26.920 | So Andy there, Andy's sitting over here,

00:22:29.560 | is pushing on the door, pushing on the hand.

00:22:32.880 | The robot keeps track of how much progress it's made

00:22:36.840 | in doing its task.

00:22:38.220 | (audience laughing)

00:22:45.840 | It's so smart, it even kicks that shell out of the way.

00:22:50.020 | No, that was a total accident.

00:22:51.880 | (audience laughing)

00:22:54.680 | And now it's just gone back to try again.

00:23:02.640 | Okay.

00:23:03.480 | And then this is a demo of autonomy.

00:23:08.680 | Here the robot has, in a previous session,

00:23:13.400 | we've taken it around the lab, this is Boston Dynamics,

00:23:16.320 | taken it around the lab and recorded visual data

00:23:19.780 | that could be used for navigation.

00:23:21.720 | And it's using its stereo to match up features

00:23:25.020 | in the environment so that it can navigate

00:23:28.160 | and go where it had gone on the previous path.

00:23:33.160 | So there's no one driving it for this, it's all autonomous.

00:23:37.760 | That was outside my office.

00:23:42.280 | Every day around noon, the robot seems to show up

00:23:45.560 | and I hear it pausing out there.

00:23:47.200 | I don't know why it turned there.

00:23:53.240 | Sometimes it comes up with a solution that isn't,

00:23:55.820 | in here you'll see another one.

00:23:57.400 | It comes up with a solution that isn't quite

00:23:59.600 | what you'd call as an optimization,

00:24:03.360 | but it does get a solution.

00:24:06.500 | So we're pretty excited by this.

00:24:09.280 | We call this Patrol Route and we're working

00:24:11.480 | on developing a lot of software to support it,

00:24:13.800 | to make it so that other people can capture a patrol route

00:24:17.600 | and then execute them on a routine basis,

00:24:20.880 | and then do other tasks while they're on the patrol route.

00:24:27.320 | Seth, you're on.

00:24:29.360 | So now we'll do a demo of Spot Mini.

00:24:32.180 | So for this demo, Seth's got a joystick

00:24:44.760 | and he's telling it the speed to go

00:24:48.080 | in the forward direction and turning,

00:24:50.720 | but the robot's doing all its own gate selection,

00:24:55.360 | coordination of legs, balance obviously.

00:25:00.360 | So the robot has a bunch of different gates.

00:25:04.200 | It can walk.

00:25:05.080 | Here it's doing one leg at a time.

00:25:09.320 | It can trot.

00:25:14.360 | I don't know, you do whatever gates you want, Seth.

00:25:16.840 | He's gotta use a selected egg.

00:25:22.000 | So here's trotting, which is diagonal pairs of legs.

00:25:25.400 | It can do pacing, which is lateral pairs of legs

00:25:28.960 | to get working together.

00:25:30.560 | I have to tell you, in the earliest days

00:25:33.440 | of me being involved in leg and locomotion,

00:25:35.680 | I thought gate was a big deal,

00:25:37.780 | but it's really kind of a small thing.

00:25:40.360 | And I don't think it's central to what matters,

00:25:43.580 | which is support, stability,

00:25:46.080 | propulsion, and things like that.

00:25:48.940 | I'm gonna wrap up shortly.

00:25:50.180 | I just thought I'd say a couple of words

00:25:51.880 | about the mechanical side.

00:25:54.420 | Atlas is a new version of a humanoid.

00:25:59.560 | I know some of you worked

00:26:00.840 | with the DARPA Robotics Challenge humanoid,

00:26:03.820 | which was a big hulking thing that we made,

00:26:06.160 | and this is a much more svelte one.

00:26:08.580 | And the way we got there was to work on

00:26:13.020 | the elements of the mechanical design

00:26:18.420 | to take advantage of 3D printing

00:26:20.600 | and some optimization.

00:26:23.620 | And we focused on two or three different things.

00:26:26.500 | One is making some of the leg parts

00:26:29.680 | where we embed hydraulic pathways,

00:26:32.260 | hydraulic actuators, places for valve mounts

00:26:35.180 | and filters and things like that into the leg.

00:26:37.580 | And this is what that looks like.

00:26:39.220 | There's a single upper leg part

00:26:42.300 | that incorporated about 15 or 20

00:26:45.100 | different separate components in the previous design,

00:26:47.960 | which made it lighter, more compact,

00:26:50.180 | and higher strength to weight ratio.

00:26:53.140 | We also developed a hydraulic power unit,

00:26:56.880 | which takes many components.

00:27:00.620 | The thing on the left are the components as separate ones.

00:27:04.420 | And we were able to print up parts

00:27:07.140 | that integrated them so that there was a motor,

00:27:10.020 | a pump inside of a motor, an accumulator,

00:27:13.700 | a reservoir, valves, filters, and those things.

00:27:17.460 | And we shrunk it down so that the robot

00:27:20.260 | could be smaller and lighter.

00:27:22.460 | And using that approach, we went from

00:27:26.060 | about a 375 pound DRC robot to a 190 pound robot,

00:27:31.060 | and then the current one is about 165 pounds.

00:27:35.560 | Now this picture might lead you to believe

00:27:37.480 | that I'm advertising myself as only weighing 165 pounds.

00:27:41.420 | And unfortunately that's not true, but I'm working on it.

00:27:45.480 | (audience laughing)

00:27:47.520 | But it is close to my size and weight.

00:27:49.440 | And I don't know, I don't think we have this out as a video.

00:27:54.120 | Here's some robot behavior that uses whole body motion,

00:27:59.120 | meaning the mobility base plus the arms plus the torso

00:28:05.440 | are all combining in order to handle these boxes.

00:28:08.960 | It's using vision with the QR codes to simplify the task.

00:28:12.560 | Here we're trying to go at human speeds of operation,

00:28:16.120 | and so the robot searches for a box using its vision.

00:28:20.100 | (audience laughing)

00:28:23.100 | I think that was the only take we ever got

00:28:28.600 | with both robots working together.

00:28:31.320 | And one of the problems with YouTube

00:28:36.800 | is everybody's already seen what you've been up to

00:28:38.680 | by the time you go around to give a talk,

00:28:40.920 | so I imagine most of you have seen this.

00:28:43.240 | But here's a parkour robot we're working on

00:28:47.840 | where we've actually strengthened the hips

00:28:50.820 | so that it can do a little bit more jumping and--

00:28:53.320 | (audience laughing)

00:28:58.320 | And it's kind of interesting that we've been interested

00:29:10.360 | in making a robot a little bit like the humanoid

00:29:13.360 | that has less degrees of freedom,

00:29:15.740 | fewer degrees of freedom and is simpler,

00:29:17.760 | and we designed this robot,

00:29:19.400 | and the ultimate version of this will have about 10 joints,

00:29:22.920 | whereas the humanoid had 28,

00:29:25.880 | and have many of the same capabilities.

00:29:28.860 | We have some use cases for this

00:29:34.040 | that I'm not gonna talk about today,

00:29:36.320 | but this robot can lift heavy loads.

00:29:39.040 | It has a relatively small footprint

00:29:40.960 | given what its strength is.

00:29:43.420 | So the way things are done in logistics now

00:29:46.600 | is to use big robot arms

00:29:48.800 | that take up a lot of floor area or heavy,

00:29:52.000 | and we're looking at ways of using a robot like this one.

00:29:56.040 | Not exactly this, it's sort of an evolution of this design

00:30:00.100 | in order to do logistics operations.

00:30:02.720 | So I wanna make a pitch to you.

00:30:07.640 | Boston Dynamics is hiring,

00:30:09.440 | and I hope some of you will apply for a job there.

00:30:13.760 | These are, how many is it, six times three.

00:30:17.440 | These are 18 MIT alum that currently work at the company,

00:30:22.440 | many of them for many years,

00:30:26.020 | so I'm sort of making the point

00:30:27.520 | that these people are happy there,

00:30:29.240 | just like you could be,

00:30:31.740 | and I hope you'll look at our website

00:30:34.520 | and see what we're looking for and consider it.

00:30:37.760 | So I'm just gonna wrap up by talking about,

00:30:41.360 | you know, I used to be a professor here

00:30:43.440 | and at Carnegie Mellon,

00:30:44.760 | and when I was a professor,

00:30:46.320 | we used to mostly wrote papers,

00:30:48.740 | and we were excited by how many papers we could write

00:30:53.200 | and how many people cited them in their papers,

00:30:57.760 | but as a company guy,

00:30:59.280 | instead of papers, I think we count YouTube hits,

00:31:04.120 | and instead of citations,

00:31:06.840 | here I wanna tell you what this is,

00:31:08.720 | but most of you probably know.

00:31:10.640 | (audience laughing)

00:31:18.680 | (audience laughing)

00:31:21.680 | So now we count spoofs instead of citations,

00:31:40.080 | and I'm happy to say that we're doing great.

00:31:42.780 | We have about two dozen big dog spoofs.

00:31:47.780 | Here's four of them,

00:31:49.620 | and the upper left is in Akihabara, Japan.

00:31:52.580 | The upper right is a Los Angeles online television show.

00:31:55.860 | It's the Netherlands on the lower left,

00:31:58.260 | and I guess that's Appalachia on the right.

00:32:00.880 | The poor kid doesn't even have a friend

00:32:03.900 | to be in his movie. (audience laughing)

00:32:08.900 | Well, what about Atlas?

00:32:12.040 | (audience laughing)

00:32:15.040 | Can you hear that?

00:32:17.520 | I love you, box. (audience laughing)

00:32:26.680 | Goodnight, box. (audience laughing)

00:32:28.720 | Box. (audience laughing)

00:32:33.720 | Hello, box. (audience laughing)

00:32:36.360 | Do it, do it, I love you.

00:32:38.400 | No. (audience laughing)

00:32:41.520 | Here's another one.

00:32:43.040 | (audience laughing)

00:32:46.040 | (dog barking)

00:32:56.180 | (audience laughing)

00:32:59.180 | All right, where do you want to, mother?

00:33:05.060 | (audience laughing)

00:33:09.500 | (audience applauding)

00:33:12.660 | So we have a big crew working on all these projects.

00:33:23.420 | You've gotten to meet a couple of them here,

00:33:24.920 | but it's really quite a team

00:33:26.780 | and an absolute pleasure to work with.

00:33:29.240 | So anyway, thank you. (audience laughing)

00:33:31.780 | (audience applauding)

00:33:39.440 | - Thanks for the presentation, it was amazing.

00:33:42.580 | What sort of physics simulation, if any,

00:33:46.500 | do you have in your robots?

00:33:47.780 | And do you really think that with the current trend

00:33:52.340 | of neural networks, we can just do end-to-end modeling

00:33:55.340 | of these robots without any sort of notion of physics,

00:33:58.940 | but just neural networks?

00:34:00.400 | - So we have simulators that we've worked on

00:34:04.140 | for a long time, very detailed, in some cases validated.

00:34:08.340 | Validated mean compare the behavior of the simulator

00:34:11.360 | to the physics of ground truth.

00:34:14.460 | And I think they're important for our work

00:34:17.980 | and we use them frequently,

00:34:19.540 | but the end-to-end doesn't ring quite true.

00:34:22.580 | Usually when we use simulation,

00:34:25.040 | the user is knowledgeable about the trade-offs

00:34:28.680 | between doing a physical experiment

00:34:30.260 | and doing a simulated experiment.

00:34:32.180 | And they're usually getting at some specific setup question

00:34:37.580 | rather than the idea that you start at one end.

00:34:40.220 | At least in our experience,

00:34:43.020 | trying to simulate all the subtleties

00:34:46.540 | of the hydraulic actuator, backlash in gears,

00:34:50.220 | flexibility, the non-rigidity in the components,

00:34:54.900 | that's a big undertaking and usually so distracting

00:34:59.900 | that you can't really get on with what you're doing.

00:35:02.440 | So I think we use experiment for those subtleties

00:35:05.300 | and we use simulation for bigger level dynamics questions.

00:35:09.800 | - Hey, would you say mechanical concerns

00:35:16.620 | or computational capability is more of a difficulty

00:35:21.620 | in terms of determining how quickly

00:35:24.220 | you can perform tasks with the robots?

00:35:26.900 | - You know, we like to say that they're equally important.

00:35:30.860 | We now, although we didn't start out this way,

00:35:33.300 | we now have equal strength in our groups

00:35:36.780 | in the mechanical design and implementation

00:35:39.500 | and in the software and controls and sensing.

00:35:43.620 | And I think they all matter.

00:35:46.300 | I think if you try and get by

00:35:47.820 | with just marginally designed hardware,

00:35:51.540 | you don't get much experimental time in

00:35:53.180 | because the thing's broken all the time.

00:35:55.540 | So even though we are rough on our machines,

00:35:58.180 | they mostly keep working because we put a lot of attention

00:36:02.740 | to detail in how they're designed.

00:36:04.440 | But there's still, I think perception is still

00:36:09.020 | a tall pole in the tent.

00:36:11.100 | Certainly if you want to rival human perception,

00:36:14.020 | I don't think we're anywhere near there.

00:36:16.500 | I think the self-driving car stuff is helping.

00:36:20.160 | There's a lot of interesting things happen there.

00:36:21.860 | I think specialized hardware is getting

00:36:24.580 | ASICs and things that could help.

00:36:28.180 | But it's all still needed.

00:36:31.400 | - So you guys have developed various components

00:36:36.400 | that all kind of come together to build one robot.

00:36:43.520 | Have you seen applications for any

00:36:45.480 | of these separate components elsewhere?

00:36:47.000 | So organic design, for example, for the Atlas,

00:36:50.960 | maybe prosthetics or hip replacements

00:36:53.800 | or something like that because there seems to be

00:36:55.720 | a lot of development going on individually

00:36:58.120 | as well as in the big picture.

00:36:59.640 | - I mean, you're asking a very good question.

00:37:01.280 | It was a question in case people couldn't hear is,

00:37:04.320 | aside from the value to the whole robot

00:37:07.440 | of the components we're making,

00:37:09.160 | are the components useful some other way?

00:37:11.400 | And the place where we think it's probably most true

00:37:15.500 | is the specialized hydraulic components we've made,

00:37:18.440 | servo valves and the HPU.

00:37:21.900 | I'm sure we could sell them into other industry.

00:37:25.880 | As a company focus question though,

00:37:30.980 | that's really what it comes to.

00:37:32.440 | Do we really wanna be doing that?

00:37:33.760 | Will that absorb too much time and attention and personnel?

00:37:38.440 | Or do we wanna, our heart is really in building

00:37:41.880 | future generations of robots.

00:37:43.280 | So I think we're gonna probably stay there.

00:37:45.920 | - Thanks.

00:37:49.760 | I was wondering, have you done any research

00:37:52.400 | in regards to getting the robots to perform tasks

00:37:55.400 | involving direct physical contact with humans?

00:38:00.320 | - Nope.

00:38:01.160 | The only thing we've done is we've done teleoperation,

00:38:06.120 | which is not what you mean,

00:38:07.600 | where we have a human moving and the robot copying,

00:38:11.620 | which is very interesting because you can see

00:38:14.820 | that that's a way of showing how fast the robot can be

00:38:18.700 | and how coordinated it can be using a human

00:38:21.180 | for part of the computing.

00:38:22.920 | But we don't have them interacting with people.

00:38:25.160 | I guess the closest is we once did a thing

00:38:27.600 | where a person and a robot picked up a stretcher

00:38:31.140 | and worked together to pick up the stretcher,

00:38:33.460 | but they weren't touching each other.

00:38:34.800 | They were going through the stretcher material.

00:38:38.160 | Do we have plans?

00:38:40.660 | We're really, to be honest, we're really struggling

00:38:45.760 | with coming up with some strong concepts for safety

00:38:49.000 | even without doing that.

00:38:50.320 | Robots, people's first reaction to a robot

00:38:56.680 | and people's first reaction

00:38:57.800 | and how you make a robot safe if there's a problem

00:39:00.240 | don't really work very well.

00:39:01.360 | You can't freeze the robot.

00:39:02.960 | You have to find some, you have to keep them going,

00:39:07.400 | find a way to get into a safer state.

00:39:09.900 | So I think having them in contact with people

00:39:12.040 | is just gonna be harder.

00:39:14.360 | So eventually we want to to help,

00:39:16.240 | you know, to carry, lift the elderly and things like that,

00:39:18.920 | but we're not there yet.

00:39:20.280 | - My question's about the relative rates of progress

00:39:25.460 | in robotics and machine intelligence.

00:39:28.160 | So an economist might maybe measure it

00:39:30.160 | by seeing how much money is going into computing hardware

00:39:33.320 | versus arms and legs, sensors and actuators,

00:39:36.000 | that kind of thing.

00:39:36.960 | So in one possible scenario,

00:39:41.400 | the machine intelligence rushes ahead

00:39:44.800 | and the robots are progressing more slowly

00:39:48.920 | because of kind of slow build test cycle, basically.

00:39:53.080 | It's the real world things.

00:39:54.520 | It's not so easy to get a rapid build test cycle

00:39:56.300 | with a robot.

00:39:57.140 | And in the other scenario,

00:40:00.420 | the robots are more advanced than the machine intelligence

00:40:03.700 | 'cause machine intelligence

00:40:04.820 | is just such a conceptually difficult problem.

00:40:06.900 | So in one scenario,

00:40:08.900 | the machines are telling the humans what to do.

00:40:10.740 | In the other scenario,

00:40:11.660 | the humans are telling the machines what to do, if you like.

00:40:14.340 | So do you have any kind of perspective on that whole issue

00:40:17.020 | of the machine intelligence folk gonna rush ahead,

00:40:20.180 | being robots, guys struggling behind,

00:40:22.380 | or the robots gonna get there before the massive problem

00:40:25.460 | of machine intelligence gets solved?

00:40:28.000 | Or maybe somewhere in the middle.

00:40:28.840 | - I think, let's see,

00:40:30.280 | I don't know exactly what you mean by machine intelligence.

00:40:33.000 | Are you talking about having Google do better search?

00:40:37.240 | - So computation in general.

00:40:39.880 | So at the start, I talked about economists measuring sensors,

00:40:43.800 | actuators, and compute hardware.

00:40:46.200 | So that's the kind of split I'm thinking about.

00:40:49.480 | Yeah, I think that it's always been a misconception

00:40:52.660 | that the hardware components by themselves

00:40:55.640 | constitute progress in intelligence or in robot behavior.

00:41:02.700 | I think they're important ingredients, but by themselves.

00:41:07.300 | You know, when I was a graduate student here,

00:41:09.500 | I can remember reading an ad

00:41:11.700 | for an optical character recognition system.

00:41:17.020 | And what the ad said was, you know, we have camera,

00:41:20.880 | we have a thing for holding the paper you're looking at,

00:41:23.960 | all you have to do is write the software.

00:41:25.820 | So it was all done except for you had to write the software.

00:41:28.380 | And you know, the whole problem was there.

00:41:33.380 | So I don't know if I'm answering your question.

00:41:37.380 | You know, robotics is hard.

00:41:40.040 | I think it feels like we're making progress.

00:41:42.260 | If you keep pushing, we keep making progress.

00:41:44.560 | It's not like there's a knee in the curve that we've hit.

00:41:48.500 | But I also think that the rest of the AI world

00:41:50.660 | is making good progress too,

00:41:52.180 | and it's fun being a part of it.

00:41:55.640 | - Hi.

00:41:59.460 | My question is mostly related to security.

00:42:05.420 | So since you are productizing your robots now,

00:42:10.060 | there has been research on the lidars mainly,

00:42:14.760 | where you could spoof a lidar

00:42:16.920 | and the sensor basically cannot see anything.

00:42:21.800 | So are you looking into that as well?

00:42:24.680 | Taking into consideration these awesome robots

00:42:29.760 | that you're building could be in, let's say,

00:42:32.720 | defense, working for the defense as well.

00:42:35.000 | So those are like really harsh environments.

00:42:36.920 | - Yeah, I mean, these are very hard problems.

00:42:39.520 | If someone, if an intelligent adversary

00:42:42.540 | wants to trick the robot,

00:42:44.460 | it's not all that hard these days.

00:42:48.060 | You know, we're working probably the other end

00:42:50.100 | of the problem, you know, trying to do the basics right now.

00:42:54.340 | I don't think, you know, I don't think robots

00:42:58.180 | are gonna be as autonomous in a hostile environment

00:43:02.300 | as people either think or fear

00:43:05.900 | because of how frail they'll still be

00:43:09.440 | until we get further along.

00:43:10.840 | - Hi there. - Hey.

00:43:15.880 | - I wanted to ask about two things

00:43:17.560 | that are going to probably play a big role in adoption.

00:43:21.200 | The first is price.

00:43:22.360 | So if you could speak to the current unit price

00:43:24.560 | of a Spot Mini and how that you think

00:43:25.960 | is going to evolve over time.

00:43:27.820 | And the second is sort of consumer psychology.

00:43:31.320 | I felt like when I saw the test at the end

00:43:35.720 | of the robots wearing, my level of comfort

00:43:37.680 | with it being in my house suddenly shot up.

00:43:40.160 | It seemed way more human.

00:43:41.760 | So I was just thinking about what kinds of experiments

00:43:44.400 | you guys have run, what you've thought about

00:43:45.820 | with respect to making people more comfortable

00:43:47.360 | with robots working around them.

00:43:49.200 | - Yeah, in terms of cost, you know,

00:43:53.800 | we're not saying what this thing costs yet,

00:43:56.720 | but we will later in the year.

00:43:58.940 | We have reduced the cost of this by about a factor of 10

00:44:02.300 | from what the first prototypes cost.

00:44:04.600 | So we're making progress.

00:44:06.000 | In terms of the psychology of robots,

00:44:10.200 | it's been very interesting to watch.

00:44:12.080 | You know, we got branded sort of as robot abusers

00:44:16.880 | because we kicked our robot.

00:44:18.980 | Really what we were doing was trying to show

00:44:21.880 | how good they were at balancing.

00:44:23.720 | And we didn't think we were abusing them.

00:44:26.080 | I have video of me pushing on my daughter

00:44:28.140 | when she's one years old and actually knocking her over,

00:44:30.800 | but that wasn't my goal.

00:44:32.200 | I wanted to kind of test out her balance.

00:44:34.640 | (audience laughing)

00:44:36.400 | I bet you, you know, if you guys have kids

00:44:38.960 | or you're at all that, you've done stuff like that.

00:44:42.040 | So, but we have adjusted a little bit.

00:44:45.720 | And so we don't usually push on the robots in our videos,

00:44:49.120 | despite the one we showed with Andy hockey-sticking

00:44:54.000 | the hand on this thing.

00:44:55.560 | That's why we had the banana peels as a way

00:44:58.360 | to have the robot crash without us being,

00:45:00.960 | having our fingerprints on it.

00:45:02.520 | You know, I guess the other data point I have

00:45:08.840 | is that if you look at the likes and dislikes

00:45:11.520 | on our YouTube videos, we found a way to get

00:45:15.120 | the likes to dislikes ratio much higher

00:45:18.520 | by partly probably by not looking like

00:45:22.200 | we're abusing the robots.

00:45:25.640 | There's probably a long way to go

00:45:27.080 | to make these things really friendly.

00:45:30.240 | And I have to admit there's a little spirit

00:45:34.160 | at our company of being kind of,

00:45:37.680 | you know, it's fun being bad boys in terms of,

00:45:40.720 | you know, just make the robot do cool stuff

00:45:42.680 | and leave the emotions to others.

00:45:46.540 | And certainly the social robots that have so much

00:45:50.080 | going into making them cute, I don't know.

00:45:54.480 | I'm sure we'll have marketing people working on that.

00:45:57.760 | I don't know what else to say.

00:45:59.160 | (audience laughing)

00:46:01.440 | - Hi, I have a general question.

00:46:03.560 | So in terms of research purpose or like practical purpose,

00:46:07.680 | so what are the reasons that we choose

00:46:09.280 | to investigate on this humanoid robot?

00:46:12.640 | So it seems like it cannot run as fast as the cheetah

00:46:16.480 | and it also cannot carry as many stuff as the big dog.

00:46:20.080 | Yeah.

00:46:22.640 | - You're basically saying that the humanoids

00:46:24.920 | don't seem to be as practical in terms of functionality?

00:46:27.200 | - Right, so is it more efficient,

00:46:28.920 | like are the humanoid robots more efficient

00:46:31.760 | than these cheetahs and the big dogs?

00:46:34.800 | - Well, you know, the, so I don't have a good answer.

00:46:39.800 | The motivation for the DRC, the DARPA Robotics Challenge,

00:46:43.780 | which was humanoid robots, was to say

00:46:46.640 | that they wanted to use robots that could go to the places

00:46:49.480 | designed for humans.

00:46:51.520 | And so that's why they used the human form.

00:46:54.680 | And I think, you know, there's an argument there.

00:46:57.680 | It is true that the human form has a lot of complexity

00:47:01.880 | to it because you have very complicated legs in the biped

00:47:05.800 | and they're supporting the weight of the body and the arms,

00:47:08.400 | whereas the quadrupeds can spread all that out.

00:47:11.200 | So I'm sympathetic to your question.

00:47:17.240 | I don't really have an answer.

00:47:18.680 | I can tell you that the public's reaction

00:47:22.220 | to a humanoid robot is off the scale

00:47:25.520 | compared to anything we've done with quadruped robots

00:47:28.600 | for what that's worth.

00:47:30.240 | So we always get a lot of viewership

00:47:33.520 | if we show a humanoid doing something.

00:47:35.420 | But I think it's a question that we will keep addressing.

00:47:39.320 | We are gonna keep pushing on getting the humanoid

00:47:41.800 | to do more and more human-like things,

00:47:45.160 | even though we probably won't commercialize them

00:47:47.200 | as soon as we commercialize the other stuff.

00:47:49.400 | - How do you specify goals?

00:47:53.840 | And although you said earlier that it's expensive

00:47:56.200 | to do simulations and stuff,

00:47:57.720 | do you have any intentions

00:47:59.200 | of doing any deep reinforcement learning?

00:48:01.200 | - What was the last thing you said?

00:48:02.200 | - Do you have any intentions

00:48:03.040 | of doing deep reinforcement learning?

00:48:04.880 | - I'll do the last one first.

00:48:10.280 | I'm sure we will use learning before too long.

00:48:15.560 | I'm not sure whether it'll be deep reinforcement learning

00:48:18.400 | or something else, but mostly we're interested

00:48:21.400 | in optimizing the complicated state space

00:48:25.720 | partitioning we do.

00:48:27.040 | Right now we use, people make very simple decisions

00:48:32.960 | as to how to divide up the space,

00:48:34.560 | and we think that these things could probably

00:48:36.200 | be really improved if we use the learning approach.

00:48:40.720 | So that's probably the first place we'll apply it.

00:48:42.720 | We do a little bit of learning here and there,

00:48:44.500 | but not much compared to how much learning

00:48:46.840 | is talked about out there.

00:48:48.440 | What was the other question?

00:48:50.040 | (man speaking off mic)

00:48:51.600 | How do we specify a goal?

00:48:53.360 | You mean to the robot, or how do we decide as a company?

00:48:56.160 | So I don't think there's any across the board answer.

00:49:03.000 | We write applications, for instance,

00:49:06.080 | for each of these uses.

00:49:07.940 | So for instance, where we were doing the patrol route,

00:49:11.680 | we have an application that has a UI

00:49:14.120 | that lets the user tell it the information it needs.

00:49:17.720 | It can tell it to go ahead and start on the patrol,

00:49:21.200 | and things like that.

00:49:22.620 | For the door, I think there's a button on the controller.

00:49:28.840 | We can show you afterwards if you want.

00:49:31.800 | And you walk the robot up to the door,

00:49:34.400 | where you're steering it, and then you press the button,

00:49:37.120 | and then it starts looking for the door handle,

00:49:38.960 | and it goes through the whole, you know,

00:49:40.600 | it goes through the door.

00:49:41.800 | (man speaking off mic)

00:49:44.920 | But I don't think these answers are fundamental.

00:49:46.640 | I think you could do it lots of different ways.

00:49:48.640 | You know, we're working on all the machinery

00:49:50.880 | coming up from the bottom to be able to do these things.

00:49:54.240 | And then, you know, in some case,

00:49:55.440 | you could have it be buttons on a UI.

00:49:57.600 | It could be an API that's accessed

00:50:00.680 | through some higher level AI.

00:50:03.240 | And we just aren't sweating that part of it at this point.

00:50:06.560 | - Hi, so aside from locomotion,

00:50:12.520 | I can use my body for like, you know,

00:50:14.720 | nonverbal communication to communicate my intentions

00:50:17.320 | and other such things, even though I'm not always aware of it.

00:50:20.560 | And I guess I'm wondering if this is something

00:50:23.880 | that you've considered for these robots.

00:50:26.560 | - I think the closest we've come

00:50:27.760 | is having the robot go like this after the flip,

00:50:30.680 | which was a way of communicating.

00:50:32.680 | We really haven't done anything along those lines.

00:50:36.120 | I'll bet you, though, that people writing code

00:50:38.720 | can interpret a lot of the subtleties of what's,

00:50:41.280 | you know, what's working and what isn't

00:50:42.680 | by looking at things like that.

00:50:44.000 | But the robot isn't trying to communicate that way.

00:50:46.800 | - I have two questions.

00:50:54.680 | How do you make the robots really fast?

00:51:01.760 | - How do we make them fast?

00:51:03.160 | - No, my question is, how did you make them fast?

00:51:10.360 | - I mean, like, the time, how?

00:51:14.040 | - We get a lot of people who are really smart

00:51:18.800 | and good at working together with each other at our lab,

00:51:23.800 | and then they make plans,

00:51:26.240 | and everybody tries to stay on the plan,

00:51:30.400 | and then, you know, pull it together.

00:51:33.280 | Sometimes it doesn't go as fast as we'd like,

00:51:36.680 | especially if we have to buy parts from someone else

00:51:39.560 | and they're slow.

00:51:40.920 | That happens a lot.

00:51:41.880 | No, honestly.

00:51:43.760 | Is that what you mean?

00:51:45.080 | So we don't make them that fast.

00:51:48.600 | You know, we're pretty fast,

00:51:49.960 | you know, usually four or five months to build a new robot,

00:51:53.960 | something like that.

00:51:57.760 | But mostly it's getting people to work together.

00:51:59.920 | What's the other question?

00:52:01.360 | - The other question is, why do the people push the robots?

00:52:07.800 | (audience laughing)

00:52:09.400 | - Why do they push?

00:52:11.000 | Why do they push?

00:52:11.920 | The robots are always balancing themselves,

00:52:18.080 | and so we wanna show that they can balance

00:52:21.320 | by showing that when you knock them,

00:52:24.600 | they still, they don't fall over,

00:52:26.480 | they stay up on their feet.

00:52:28.080 | So we're kind of showing off.

00:52:29.800 | (audience laughing)

00:52:33.280 | Are you building anything?

00:52:34.580 | Why not?

00:52:36.280 | - I don't know.

00:52:38.080 | - You should.

00:52:39.880 | - It's way off.

00:52:41.200 | - Why?

00:52:42.040 | - Way off.

00:52:44.000 | - Where?

00:52:46.400 | In the basement?

00:52:47.240 | - I'm not good at building.

00:52:48.800 | - No, yes you are.

00:52:51.040 | You might think you're not.

00:52:52.160 | - Well, in some games they are.

00:52:53.640 | - You oughta give it a try.

00:52:55.000 | And you're the right age to get started.

00:52:58.680 | - I'm six and a half.

00:53:00.080 | - Perfect.

00:53:00.920 | (audience laughing)

00:53:08.720 | - All right, with that, I think,

00:53:11.280 | please give Mark a big hand.

00:53:13.280 | Thank you very much.

00:53:14.120 | - Thank you.

00:53:14.960 | (audience applauding)

Marc Raibert: Boston Dynamics | MIT Artificial Intelligence (AI)

Chapters