Robotics Professor Answers Robot Questions | Tech Support | WIRED

0:00 I'm Aaron Ames.

0:00 I'm a professor of mechanical engineering.

0:02 I'm here today to answer your questions from the internet.

0:05 This is robotics support.

0:11 Fred Bone says, "I don't understand the benefits of these food delivery robots.

0:15 You know, this is a job that could be automated,

0:17 so let's have robots automate it.

0:18 They obviously work in some scenarios, but they're not super robust.

0:22 I think the bigger thing they're trying to solve is a proof

0:24 of concept demonstration to see if we

0:26 can solve things like automated package delivery,

0:28 last mile delivery, things like that.

0:30 A lot of these are first steps towards

0:32 trying to solve this more general delivery problem,

0:35 which is actually a really hard problem.

0:36 And by the way, putting these robots on the road has taught us

0:38 a lot about both what works and also a lot of what doesn't.

0:41 If you've seen these delivery robots, they get stuck a lot.

0:43 They can hit things.

0:44 They can fall off the road.

0:46 It does tell you how hard robots is.

0:48 And I think that's sort of something to keep in mind when we're talking

0:50 about humanoids and all these other things is look at the robots that are

0:53 actually in the field today and how they sort of still kind of fail

0:57 quite a bit and it tells you we got a lot of work to do.

0:59 And by doing that, at least on the robotics side,

1:01 we will learn a lot from Paradise Knights WTFs with dancing robots.

1:07 I thought they were supposed to be our slaves

1:09 before they enslave us at least for a bit.

1:12 They're just partying.

1:13 It's pretty good.

1:14 Yeah, WTF on multiple levels.

1:16 First and foremost, why are they having robots dance?

1:20 First, there's amazing progress that's happened in humanoids.

1:24 Among other things, back flipping, running.

1:26 I mean, really, the behaviors we've achieved is

1:28 in the last year or two has been absolutely remarkable.

1:32 We figured out a very nice pipeline in which to get robots to do this.

1:35 And the way you do it is you start with a human doing those actions.

1:38 The reason why the dancing looks so humanlike is it's a human dancing.

1:42 It's it's really just puppetry.

1:44 very beautiful and advanced engineering puppetry, but puppetry nonetheless.

1:48 So, a human puts on a mocap suit or you use cameras and the human dances,

1:52 not me because I'm a very bad dancer, but you know, a good dancing human.

1:54 And then you take that data and you get the trajectories from that data.

1:57 That is the motion of of the human over time.

1:59 And then you train a reinforcement learning

2:01 algorithm on the humanoid robot that basically mimics

2:04 or copies that human data as much as possible with the morphology of the robot.

2:07 And the end result is it dances like the person that was dancing.

2:10 And as long as everything is just the way you expected it to be,

2:13 meaning the environment's like it was when the human did the thing,

2:16 we can get robots to do that now.

2:17 But you asked about this, aren't they supposed

2:19 to be helping us sort of before they party?

2:22 And the answer to that is that we still don't know how to solve that problem.

2:26 That's sort of the not very secret secret

2:28 is that all of the hard problems getting

2:30 robots to be truly autonomous and in our homes

2:32 are still hard problems that are unsolved.

2:35 The next question is from project guy 111.

2:37 What percentage chance do you think we'll end up in a Terminator future?

2:42 So, I think there's two answers to this.

2:43 One is what's the chance that AI and learning will do bad things?

2:48 And I think that probability is actually fairly high if we're not careful.

2:51 And and now's the time to be careful.

2:52 If you trust AI to be the decision maker,

2:55 if you're not very careful about having guard rails for that AI,

2:58 it will make bad decisions.

2:59 I mean, if you've ever used chat GPT or LLMs,

3:01 you see that it can produce really nice answers sometimes and it's impressive,

3:04 but then sometimes it just it's wrong.

3:07 So, we cannot trust AI.

3:09 In my opinion, you can never trust AI, but you can use AI as a powerful tool.

3:13 Just like if you search something online on Google,

3:16 you get a lot of results back,

3:17 gives you a lot of information, but you have to verify and double check.

3:20 So, I think if we put like AI in charge of our, you know,

3:23 weapons or something silly like that, then, you know, bad things will happen.

3:27 At the same time, the second part of Terminator was it became sensient, right?

3:30 It actually learned to think on its own.

3:32 And I don't think we're anywhere near that.

3:34 So, I have no concern of sensient AI.

3:36 Right now, AI is not intelligent.

3:39 They say AI, artificial intelligence.

3:41 There's no actual intelligence.

3:42 It has no notion of what it's saying or doing.

3:45 It is simply pattern matching at a scale we've never pattern matched before.

3:49 From I got too silly.

3:51 What benefits do legged robots have over wheeled or tracked vehicles?

3:54 legs are inherently beneficial if you want to operate in environments for which

3:59 they're built for humans and more importantly where things are not flat.

4:03 So wheels are massively efficient in how you can

4:06 move around environments as long as there's not uneven terrain.

4:08 If you've ever been in a wheelchair

4:10 or wheeled someone out around in a wheelchair,

4:13 you realize very quickly how flat the world is not.

4:16 Even what you perceive as flat,

4:17 even in a city environment where there's sidewalks and everything,

4:20 you realize there's curbs that don't dip enough.

4:22 There's big brakes in them and all

4:23 those little things for wheels become big problems.

4:25 They become sort of sticking points and legs have

4:28 the inherent ability to walk more robustly over multiple terrain types.

4:32 Quadripeds being one example.

4:33 I mean where your dog or your cat can go on four legs is pretty incredible.

4:37 Bipeds of course are sort of the ultimate expression of mobility

4:40 in human environments because the environments are built for us.

4:43 So if you need to get into a small space and and get up

4:46 a small set of stairs or something like that or climb up a ladder,

4:49 only a biped can do that.

4:50 Give a ladder to a quadriped.

4:51 is not going to know what to do.

4:53 Give a ladder to a human.

4:54 As long as they're reasonably healthy, they can climb up that ladder no problem.

4:57 From Sammy 514, what are robot dogs actually being used for?

5:02 So, we've actually come a long ways in robot dogs,

5:05 which are actually technically called quadripeds because they have four legs.

5:09 It's really amazing what's happened in the last sort

5:11 of decade where we've gone from these robot dogs

5:14 being in really lab environments and research environments to being

5:18 things that you can buy at insanely low prices.

5:21 The hardware has made immense progress.

5:24 I think the practical use cases are still a little thin.

5:28 They're thinking about doing them as things like inspecting buildings,

5:31 sending robots ahead in disastrous scenarios, right?

5:33 And for that legs are definitely better than wheels.

5:36 The next question is from Levon21.

5:38 Is there any attempt to put chat GPT inside a robot?

5:41 Yeah, there's lots of attempts.

5:43 So, right now there's many humanoid robot makers

5:45 that have that as a layer in their humanoid robots.

5:48 And so, the way to think about this more

5:50 generally is that as we're making robots do things,

5:54 there's no one thing that's going to make them do all things, right?

5:56 So, imagine your body, you have a brain,

5:58 you also have a spinal cord, you have propriception.

6:01 So, you sort of intrinsically have multiple computers

6:03 running in your body at any given time.

6:05 like your spinal cord is a computer in its

6:07 own right and those computers run different algorithms.

6:10 So at the highest level, at our cognitive level,

6:12 that's where LLMs will sort of play a role if you'd like.

6:15 So you want to ask the robot a question and have

6:17 it perceive something about the environment and make a decision.

6:20 So ask the robot, you know, where is the red apple on the table?

6:23 It can use an LLM to parse that language into code.

6:26 Use computer vision then to to detect all the apples on the table

6:30 and return based on that an estimate of where it is.

6:33 When you actually want to reach for the apple,

6:34 an LLM is not going to reach for the apple.

6:37 That's where traditional robot control will

6:38 come in or where reinforcement learning will

6:40 come in, which is a whole another type of learning that's different from LLM.

6:44 And those are what's actually running on the robot, right?

6:46 Those are what are making the robot go,

6:48 just like your spinal cord is really what moves your body

6:50 most of the time without you having to think about it.

6:53 From I want to be free 10, are autonomous vehicles mobile robots?

6:57 Yeah.

6:57 I mean, if you want to ask what the most advanced robot is today,

7:01 you would take autonomous cars and mobile warehouse robots like Amazon has.

7:07 Autonomous cars, I think,

7:09 are one of the prime examples of the farthest that we've pushed autonomy.

7:14 So, yeah, it's absolutely a robot.

7:16 It's an advanced robot that's a beautiful work of engineering.

7:18 From Vite Persona, how does the Roomba know what's what?

7:23 I think what they're asking is how does the Roomba

7:25 have a semantic understanding of the environment is the way we

7:29 would technically phrase this question that is able to identify

7:32 things around the room that it sees and correctly identifying them.

7:36 So basically we've been able to take

7:38 a bunch of training data a bunch of examples

7:40 of pictures on the internet videos on the internet

7:43 and teach robots how to correctly identify those.

7:45 By teach what we mean is we can train up basically this big

7:49 neural network that takes images in and produces what's in the images out.

7:54 And as a result, we can now correctly identify most things in an image.

7:58 And so what Roomba does is simply it has

8:01 a camera that's perceiving those things in the environment.

8:03 It's checking with the internet based on these large models and then identifying

8:06 the things in the environment and using

8:08 that information to tell you what's going on.

8:10 The next question is from Fireplace Air.

8:12 Why make robots humanoid shaped when they could have six arms?

8:15 Humanoid robots are the most suited to do the most things,

8:19 even if they're not the best suited to do any given thing.

8:22 Again, the world is built for us.

8:25 We've built the robot for something of our shape and our function.

8:30 Doors, stairs, narrow corridors, all these things.

8:34 And if you want a robot that can slot into any scenario where a human can go,

8:38 which is pretty much most scenarios we care about, not all, but most.

8:41 You want somebody to start making you sandwiches at at in your kitchen or you

8:45 want a robot that can be dropped

8:46 into an existing factory and automate some tasks, right?

8:50 I want to say all that as a preface to the fact

8:52 that humanoid robots are not the best form for again a given application.

8:58 For example, right now the largest owner of robots is Amazon.

9:01 It has over a million.

9:03 And as a result, one of the largest sector of robotics is warehouse robots.

9:07 robots with two wheels that go under

9:10 these pallets and then move them around warehouses.

9:12 In particular, what Amazon does is have the robots

9:14 go pick up a pallet with the thing

9:15 you order from Amazon and it moves that over to a human who's filling the order.

9:19 So, the human can just stand there and they never have

9:21 to move and basically all of this stuff comes to them.

9:23 Now, you could theoretically have a humanoid robot do

9:25 this, like push these big pallets around the warehouse,

9:28 but the question would be why?

9:30 Like, these robots are very good at doing what they're meant to do,

9:33 and that environment is designed to work synergistically with the robot.

9:36 So, you could have six arms on a humanoid if you had

9:38 an application that determined that six arms would make a big difference.

9:42 Maybe you have a firefighting robot and you find that two

9:44 arms is not enough and you want four more arms

9:46 cuz you want to hold the hose and you want

9:47 to hold a camera and you want to hold the fire extinguisher.

9:50 That's the great thing about robots, by the way,

9:52 is we can make them any way we want.

9:53 A Reddit user asks, "When it comes to automation,

9:55 how close is Amazon to actually automating most, if not all,

9:59 warehouse work with robots?" The answer is very,

10:02 very close to parts and then other ones are further away.

10:06 So in terms of automation of warehouses,

10:08 Amazon is by far the leader in this domain.

10:11 Amazon robotics in particular has been working on warehouse robots,

10:16 specifically robots that move around warehouses

10:18 to move pallets around to move your order

10:19 to a person filling the order for, you know, 20 plus years now.

10:23 And they've really refined that process so they maximally and efficiently can

10:26 move packages to the sorter and then the sorder packages them up.

10:30 So the question becomes what remains?

10:32 And what remains is then the part that the human's currently doing

10:35 which is grabbing these things off the shelf and putting them in boxes.

10:39 And so this is much more an open-ended

10:41 work where they're testing out solutions right now.

10:44 So for that you can use things called local

10:46 manipulators meaning you have robot arms maybe on mobile bases

10:49 and maybe use suction cups or soft graspers and you

10:52 would pick up objects and put them in boxes.

10:54 And you can do this with varying success levels right now,

10:57 but you're not at the success level to truly automate that where you don't need

11:01 a person present for all objects cuz you

11:03 imagine all the objects that go in Amazon boxes,

11:05 all the different geometries and how they feel and look.

11:08 I mean, it it's it's a very complicated

11:10 problem to automatically and autonomously load all these objects.

11:13 So, that problem's sort of open-ended.

11:15 Kaleidoscope Inside asks,

11:17 if you could commission a robot to be built with no financial barriers,

11:21 what would your robot do?

11:23 So if I had no financial constraints,

11:25 I would try to sort of solve the exoskeleton problem.

11:30 Like a billion could get it done.

11:31 A billion in my mind could develop something

11:34 that would eradicate the need for wheelchairs essentially.

11:37 A Reddit user asks, "Why do most of the four-legged robots,

11:41 see Boston Dynamics, have their knees look backwards?

11:45 It's the inverted leg morphology.

11:47 There's potentially mathematical advantage to having your legs like this.

11:50 It's actually a two-fold thing.

11:52 It's not just that they're inverted, but that they're very light.

11:54 So, it turns out the way you control robots

11:56 and the way you get really stable locomotion behavior

11:59 on robots is by having very light legs

12:02 and having most of your mass centered in one spot.

12:04 There's lots of mathematical reasons for that that leverage

12:07 that as sort of an assumption.

12:08 And what that means is as you're walking, if your legs are very light,

12:11 you can place them very quickly to catch yourself as you fall.

12:14 A lot of the robots were designed based on that morphology.

12:17 And in addition, the inverted leg is

12:19 actually what birds have cuz birds actually have

12:21 this type of morphology where they have

12:23 very heavy big bodies and very light legs.

12:26 And as a result, birds have some of the most robust locomotion out there.

12:29 They can actually step down huge holes and then step up out

12:32 of the hole without ever changing how their main body is moving.

12:36 And so if you can get that kind of behavior on robots,

12:38 they'll be very robust and able to go all

12:40 the places where you want to take robots with legs.

12:41 A Reddit user asks, "What is so special about the Mars rover?" Curiosity.

12:47 I mean, what's not special about a Mars rover?

12:50 Pretty much everything.

12:51 It was on Mars.

12:52 I mean, the reality is these robots are incredible engineering feats.

12:57 First, just getting it to Mars is special.

12:59 And then the rover itself,

13:01 all the rovers that have gone, Curiosity and all of them,

13:04 they have these amazing engineering sort of gems in them that they built up.

13:08 You know, their suspension system is specifically

13:10 built so that it can handle rough terrain.

13:12 Their wheels are specifically built so they'll be robust to going over

13:16 rocks and things like that and you won't blow out a tire, right?

13:18 So, the entire structure is built to maximally

13:21 be robust to this really harsh environment.

13:23 I mean, there's the electronics which have to deal with this very

13:26 sandy and and and windswept environment where there's dust storms.

13:30 You know, there's power limitations just the power alone.

13:33 I mean, the sun doesn't come through at the same intensity.

13:36 You have to have good solar panels that actually charge the electronics.

13:39 The electronics have to be battery efficient, right?

13:41 And then after all of that, you have to do science with it.

13:43 And then, oh, of course, you have to talk back with Earth at the same

13:46 time and make sure you don't lose that signal.

13:47 I mean, the list goes on and on.

13:48 But to build a system like that, I mean,

13:50 every one of the rovers we've sent to Mars is an immense

13:53 and remarkable engineering feat that sort of just makes me happy as an engineer.

13:59 The next question is from F.

14:00 Kruski, 3411.

14:02 Wait, how do autonomous cars even work?

14:05 like are we trusting robots with our lives now?

14:09 Not sure about this.

14:10 It turns out that autonomous cars have gone through some things.

14:14 About a decade ago when when it seemed to be close,

14:16 there were all these startup companies that formed and they all tried to do

14:20 autonomous cars many different ways and then

14:22 exactly what you're worried about happened.

14:25 There were crashes, people got hurt despite the fact

14:28 that they'd already been working on making the system safe.

14:30 But they really realized and I think that the entire autonomous car industry

14:34 pushed more and more to make sure there was really rigorous safety methods.

14:38 So both in the algorithms to keep them safe

14:40 and then also protocols around that software architectures etc.

14:43 and then data to support it.

14:45 And so in that way there's a lot of evidence

14:47 to show that autonomous cars if they're currently deployed have very

14:50 rigorous safety standards and there's a lot of data to back

14:53 that up with the limited number of crashes that they've had.

14:56 I'm not saying they can't get better, but you know,

14:59 after a decade, autonomous cars have done the hard work.

15:02 The decade ago, we could have autonomous cars drive around on the street.

15:05 But the difference between driving around on the street for, you know,

15:08 a demo and actually driving around and all

15:10 of a sudden other cars pull out or, you know,

15:11 a ball bounces into the street and a kid chases after it, right?

15:14 Those corner cases kill and safety kills.

15:17 Meaning, if you're not safe, you die.

15:19 Both as a company, but also as a as a technology.

15:22 The next question is from Time is Grand Noob question.

15:25 Why does Elon Musk think LA lidar is not a good idea?

15:28 I don't know Elon Musk's mindset, but in my mind, lidar is awesome.

15:34 So, you have really two main sensing modalities in robots, autonomous cars, etc.

15:38 I mean, there's many more, but but in terms of perceiving the environment,

15:41 the two most popular things are using cameras, of course,

15:43 which is what Tesla does, and you have lidar.

15:45 Now, lidar sends out a bunch of laser pulses,

15:48 and those bounce back and tell you what's around you.

15:50 So, it's sort of 3D radar, but now it's using lasers.

15:53 Liidar cameras, of course, have a lot of information present.

15:57 They can tell you not just how far objects are,

15:59 if you can properly estimate distance,

16:01 but what objects there are and where they're at in the environment.

16:04 So, you can get semantic understanding in the environment and things like that.

16:07 LiDAR does an incredible job of precisely

16:10 identifying everything in the environment with like

16:12 a full 360 view so you know exactly where all the objects are,

16:15 but you have no idea really what the objects are.

16:17 Liidar is super effective on robots and on autonomous cars because

16:22 what you want to do is make sure you're not hitting anything.

16:24 And that means anything.

16:25 It doesn't matter if you're sort of hitting another car or hitting,

16:28 you know, the guardrails on the side of the road.

16:30 You don't want to hit any of that stuff.

16:32 And for that, LAR can work really quickly

16:34 and really robustly to do things like dynamic collision avoidance.

16:37 Cameras again give you the semantic understanding.

16:40 And so I imagine that what they're thinking at Tesla

16:43 generally speaking is if they can solve the camera

16:46 problem and make them as good as LAR then

16:47 this will extend to a lot of other application domains.

16:50 But I think what I found on robots is you

16:52 want to use every sensor you can get your hands on.

16:54 And so for that, for safety critical applications,

16:57 which Tesla cars could still improve on, to be honest,

17:01 you want LAR in the loop so you can

17:03 quickly and rapidly respond to dynamic changes in the environment

17:06 and not have to deal with the latency that's

17:08 present in a lot of perception based um representations.

17:11 We've got a Reddit user who asks,

17:13 "Why is a surgical robot better than a surgeon?"

17:16 The reality is that robots are really good at certain things.

17:20 They're very very good at precise small motions and doing

17:25 those precise motions repeatedly again and again and again.

17:28 There's also places where humans are infinitely better than robots.

17:31 Basically any place where you have to interact

17:34 with the environment in a soft and tactile way.

17:37 And so here's where actually surgical robots

17:39 get complicated because our skin is soft,

17:41 our organs are soft, our bodies are soft.

17:43 And so as this robot is moving through your body to perform surgery,

17:46 you need to be aware of those soft

17:48 and sort of compliant interactions with the human body.

17:51 And that's why a surgeon is in the mix.

17:53 The surgical robot can do those precise

17:55 motions while the surgeon operates the robot

17:57 and they get all this haptic and tactile

17:59 feedback while they're doing the robotic surgery.

18:01 So they can take advantage of the precision of the robot

18:03 while still being able to do what humans do really well,

18:06 which is operate in these complex environments.

18:10 This is from the balisand.

18:11 Why is it so difficult to make a closed clothing robot?

18:14 The problem with clothes is that clothes are really difficult to model.

18:18 They move around.

18:19 There's fabric.

18:20 I mean, it's very difficult to handle.

18:22 So, a robot has to interact with something

18:24 in the environment that we can't put in a computer easily.

18:27 And so, that's why this is a big challenge and why there's been a big push

18:30 to use things like machine learning and AI

18:32 to try to understand this clothes folding problem.

18:35 That is you train a robot with the humans folding clothes like you telly

18:39 operate the robot so that it folds the clothes a bunch of times and then

18:41 you try to teach the robot that task without it actually having a model

18:46 of the environment itself but just these reference

18:48 trajectories that humans generated through tea operation.

18:51 The next question is from surprise news 763.

18:53 Whatever happened to the neo robot?

18:56 It's interesting because there was this big announcement that you could buy like

18:59 a humanoid robot as soon as this year in fact and have in your home.

19:03 Now to give them credit,

19:05 they completely were transparent about the fact that this robot

19:10 couldn't actually do a lot without teley operation.

19:12 There's even in the app,

19:14 they show you how you will book time on the robot with a teley operator.

19:18 So this is interesting from a couple of perspectives.

19:20 One is it's a tacit at least acknowledgement that humanoid

19:24 robots are not actually ready to be deployed in your homes.

19:26 I mean that's the that's the clear and obvious ramification.

19:29 So the question is why are they going to put

19:31 a robot in the home that's not actually ready?

19:32 I mean it can't do a lot of tasks on its own very few.

19:35 The reason why is right now the idea in the robotics

19:37 industry is that the only thing that's missing is data.

19:40 So we have internet scale data and that's what made LLMs possible.

19:44 So what they're betting on is if we can get

19:46 sort of internet scale humanoid robot data we will be able

19:49 to solve all these problems just like chat GPT solved the problems

19:52 for language we'll be able to solve in this other way.

19:54 So, the idea with the Neo robot is not so much that they think they're really

19:57 going to sell all of these robots

19:58 and make money off them by having them teleyoperated,

20:01 but rather this is an amazing way to do data collection.

20:03 You put all these robots in homes and then

20:05 people agree to let you teleyoperate them in their home.

20:07 They collect all this data and then they're going to use this data to train

20:10 a large model to determine how to make the robot do these tasks automatically.

20:14 If you can get early adopters to sign up for that because it's cool,

20:17 you'll be able to generate enough data to actually train

20:20 large models to to learn how to do these tasks.

20:23 But I want to say one final thing is that I don't think it's going to work.

20:26 I don't think that there's enough data you can collect

20:30 in this way to solve this problem with just human data.

20:32 People tend to confuse LLMs and text

20:35 with robots which are fundamentally different things.

20:38 In particular, the amount of data needed to understand how a robot moves.

20:43 These trajectories, think about them as trajectories as the language

20:47 of robots and they include position, velocity, force,

20:50 all this rich dynamic information,

20:52 but they're so variable because there's so many

20:54 degrees of freedom that it makes language seem simple.

20:57 And in fact, the way to really understand this is look at evolution.

20:59 We developed language in a very small fraction of our total evolution.

21:04 Most of the time we spent evolving was to do these other basic things, right?

21:08 Things that require motion of our body.

21:10 That's why they say you only use 10% of your brain.

21:12 What they what they mean by that is that a lot of your brain

21:14 is being used for these very complicated motions that we that we do, right?

21:18 And so the point of this all being is I don't think that data is going to work.

21:21 I think that what we're missing is not more data.

21:24 Not that it can't be helpful or help to polish things,

21:26 but we need to understand physics as well.

21:29 You need to merge physics with human data and that's the only way you're

21:32 going to solve the general intelligence problem

21:35 on humanoid robots because robots are not language.

21:39 That's it.

21:40 That's all the questions.

21:41 Hope you learned something along the way.

21:43 Thanks,