Anca Dragan: Human-Robot Interaction and Reward Engineering
Guests: Anca Dragan
Full transcript · Read the full conversation in order, with its original context.
Contents11 chapters
§01Transcript · 00:00–09:09§02Transcript · 09:09–18:17§03Transcript · 18:17–26:54§04Transcript · 26:54–35:24§05Transcript · 35:24–44:09§06Transcript · 44:09–53:19§07Transcript · 53:19–1:02:38§08Transcript · 1:02:38–1:12:10§09Transcript · 1:12:10–1:20:53§10Transcript · 1:20:53–1:29:38§11Transcript · 1:29:38–1:38:46§01Transcript · 00:00–09:09
The following is a conversation with Anca Drogon,
a professor at Berkeley working on human robot interaction,
algorithms that look beyond the robot's function
in isolation and generate robot behavior
that accounts for interaction
and coordination with human beings.
She also consults at Waymo, the autonomous vehicle company,
but in this conversation,
she is 100% wearing her Berkeley hat.
She is one of the most brilliant and fun roboticists
in the world to talk with.
I had a tough and crazy day leading up to this conversation,
so I was a bit tired, even more so than usual,
but almost immediately as she walked in,
her energy, passion, and excitement
for human robot interaction was contagious.
So I had a lot of fun and really enjoyed this conversation.
This is the Artificial Intelligence Podcast.
If you enjoy it, subscribe on YouTube,
review it with five stars on Apple Podcast,
support it on Patreon,
or simply connect with me on Twitter at Lex Friedman,
spelled F R I D M A N.
As usual, I'll do one or two minutes of ads now
and never any ads in the middle
that can break the flow of the conversation.
I hope that works for you
and doesn't hurt the listening experience.
This show is presented by Cash App,
the number one finance app in the App Store.
When you get it, use code LEXPODCAST.
Cash App lets you send money to friends,
buy Bitcoin, and invest in the stock market
with as little as one dollar.
Since Cash App does fractional share trading,
let me mention that the order execution algorithm
that works behind the scenes
to create the abstraction of fractional orders
is an algorithmic marvel.
So big props to the Cash App engineers
for solving a hard problem that in the end
provides an easy interface that takes a step up
to the next layer of abstraction over the stock market,
making trading more accessible for new investors
and diversification much easier.
So again, if you get Cash App from the App Store
or Google Play and use the code LEXPODCAST,
you get $10 and Cash App will also donate $10 to FIRST,
an organization that is helping to advance robotics
and STEM education for young people around the world.
And now, here's my conversation with Anca Drogon.
When did you first fall in love with robotics?
I think it was a very gradual process
and it was somewhat accidental actually
because I first started getting into programming
when I was a kid and then into math
and then I decided computer science
was the thing I was gonna do
and then in college I got into AI
and then I applied to the Robotics Institute
at Carnegie Mellon and I was coming from this little school
in Germany that nobody had heard of
but I had spent an exchange semester at Carnegie Mellon
so I had letters from Carnegie Mellon.
So that was the only, you know, MIT said no,
Berkeley said no, Stanford said no.
That was the only place I got into
so I went there to the Robotics Institute
and I thought that robotics is a really cool way
to actually apply the stuff that I knew and loved
to like optimization so that's how I got into robotics.
I have a better story how I got into cars
which is I used to do mostly manipulation in my PhD
but now I do kind of a bit of everything application wise
including cars and I got into cars
because I was here in Berkeley
while I was a PhD student still for RSS 2014,
Peter Bill organized it and he arranged for,
it was Google at the time to give us rides
in self driving cars and I was in a robot
and it was just making decision after decision,
the right call and it was so amazing.
So it was a whole different experience, right?
Just I mean manipulation is so hard you can't do anything
and there it was.
Was it the most magical robot you've ever met?
So like for me to meet a Google self driving car
for the first time was like a transformative moment.
Like I had two moments like that,
that and Spot Mini, I don't know if you met Spot Mini
from Boston Dynamics.
I felt like I fell in love or something
like it, cause I know how a Spot Mini works, right?
It's just, I mean there's nothing truly special,
it's great engineering work but the anthropomorphism
that went on into my brain that came to life
like it had a little arm and it looked at me,
he, she looked at me, I don't know,
there's a magical connection there
and it made me realize, wow, robots can be so much more
than things that manipulate objects.
They can be things that have a human connection.
Do you have, was the self driving car the moment like,
was there a robot that truly sort of inspired you?
That was, I remember that experience very viscerally,
riding in that car and being just wowed.
I had the, they gave us a sticker that said,
I rode in a self driving car
and it had this cute little firefly on and,
or logo or something like that.
Oh, that was like the smaller one, like the firefly.
Yeah, the really cute one, yeah.
And I put it on my laptop and I had that for years
until I finally changed my laptop out and you know.
What about if we walk back, you mentioned optimization,
like what beautiful ideas inspired you in math,
computer science early on?
Like why get into this field?
It seems like a cold and boring field of math.
Like what was exciting to you about it?
The thing is I liked math from very early on,
from fifth grade is when I got into the math Olympiad
and all of that.
Oh, you competed too?
Yeah, this, it Romania is like our national sport too,
you gotta understand.
So I got into that fairly early
and it was a little, maybe too just theory
with no kind of, I didn't kind of had a,
didn't really have a goal.
And other than understanding, which was cool,
I always liked learning and understanding,
but there was no, okay,
what am I applying this understanding to?
And so I think that's how I got into,
more heavily into computer science
because it was kind of math meets something
you can do tangibly in the world.
Do you remember like the first program you've written?
Okay, the first program I've written with,
I kind of do, it was in Cubasic in fourth grade.
Wow.
And it was drawing like a circle.
Graphics.
Yeah, that was, I don't know how to do that anymore,
but in fourth grade,
that's the first thing that they taught me.
I was like, you could take a special,
I wouldn't say it was an extracurricular,
it's in the sense an extracurricular,
so you could sign up for dance or music or programming.
And I did the programming thing
and my mom was like, what, why?
Did you compete in programming?
Like these days, Romania probably,
that's like a big thing.
There's a programming competition.
Was that, did that touch you at all?
I did a little bit of the computer science Olympian,
but not as seriously as I did the math Olympian.
So it was programming.
Yeah, it's basically,
here's a hard math problem,
solve it with a computer is kind of the deal.
Yeah, it's more like algorithm.
Exactly, it's always algorithmic.
So again, you kind of mentioned the Google self driving car,
but outside of that,
what's like who or what is your favorite robot,
real or fictional that like captivated
your imagination throughout?
I mean, I guess you kind of alluded
to the Google self drive,
the Firefly was a magical moment,
but is there something else?
It wasn't the Firefly there,
I think there was the Lexus by the way.
This was back then.
But yeah, so good question.
Okay, my favorite fictional robot is WALLI.
And I love how amazingly expressive it is.
I'm personally thinks a little bit
about expressive motion kinds of things you're saying with,
you can do this and it's a head and it's the manipulator
and what does it all mean?
I like to think about that stuff.
I love Pixar, I love animation.
WALLI has two big eyes, I think, or no?
Yeah, it has these cameras and they move.
So yeah, it goes and then it's super cute.
Yeah, the way it moves is just so expressive,
the timing of that motion,
what it's doing with its arms
and what it's doing with these lenses is amazing.
And so I've really liked that from the start.
And then on top of that, sometimes I share this,
it's a personal story I share with people
or when I teach about AI or whatnot.
My husband proposed to me by building a WALLI
and he actuated it.
§02Transcript · 09:09–18:17
So it's seven degrees of freedom, including the lens thing.
And it kind of came in and it had the,
he made it have like the belly box opening thing.
So it just did that.
And then it spewed out this box made out of Legos
that open slowly and then bam, yeah.
Yeah, it was quite, it set a bar.
That could be like the most impressive thing I've ever heard.
Okay.
That was special connection to WALLI, long story short.
I like WALLI because I like animation and I like robots
and I like the fact that this was,
we still have this robot to this day.
How hard is that problem,
do you think of the expressivity of robots?
Like with the Boston Dynamics, I never talked to those folks
about this particular element.
I've talked to them a lot,
but it seems to be like almost an accidental side effect
for them that they weren't,
I don't know if they're faking it.
They weren't trying to, okay.
They do say that the gripper,
it was not intended to be a face.
I don't know if that's a honest statement,
but I think they're legitimate.
Probably yes. And so do we automatically just
anthropomorphize anything we can see about a robot?
So like the question is,
how hard is it to create a WALLI type robot
that connects so deeply with us humans?
What do you think?
It's really hard, right?
So it depends on what setting.
So if you wanna do it in this very particular narrow setting
where it does only one thing and it's expressive,
then you can get an animator, you know,
you can have Pixar on call come in,
design some trajectories.
There was a, Anki had a robot called Cosmo
where they put in some of these animations.
That part is easy, right?
The hard part is doing it not via these
kind of handcrafted behaviors,
but doing it generally autonomously.
Like I want robots, I don't work on,
just to clarify, I don't, I used to work a lot on this.
I don't work on that quite as much these days,
but the notion of having robots that, you know,
when they pick something up and put it in a place,
they can do that with various forms of style,
or you can say, well, this robot is, you know,
succeeding at this task and is confident
versus it's hesitant versus, you know,
maybe it's happy or it's, you know,
disappointed about something, some failure that it had.
I think that when robots move,
they can communicate so much about internal states
or perceived internal states that they have.
And I think that's really useful
and an element that we'll want in the future
because I was reading this article
about how kids are,
kids are being rude to Alexa
because they can be rude to it
and it doesn't really get angry, right?
It doesn't reply in any way, it just says the same thing.
So I think there's, at least for that,
for the correct development of children,
it's important that these things,
you kind of react differently.
I also think, you know, you walk in your home
and you have a personal robot and if you're really pissed,
presumably the robot should kind of behave
slightly differently than when you're super happy
and excited, but it's really hard because it's,
I don't know, you know, the way I would think about it
and the way I thought about it when it came to
expressing goals or intentions for robots,
it's, well, what's really happening is that
instead of doing robotics where you have your state
and you have your action space and you have your space,
the reward function that you're trying to optimize,
now you kind of have to expand the notion of state
to include this human internal state.
What is the person actually perceiving?
What do they think about the robots?
Something or rather,
and then you have to optimize in that system.
And so that means that you have to understand
how your motion, your actions end up sort of influencing
the observer's kind of perception of you.
And it's very hard to write math about that.
Right, so when you start to think about
incorporating the human into the state model,
apologize for the philosophical question,
but how complicated are human beings, do you think?
Like, can they be reduced to a kind of
almost like an object that moves
and maybe has some basic intents?
Or is there something, do we have to model things like mood
and general aggressiveness and time?
I mean, all these kinds of human qualities
or like game theoretic qualities, like what's your sense?
How complicated is...
How hard is the problem of human robot interaction?
Yeah, should we talk about
what the problem of human robot interaction is?
Yeah, what is human robot interaction?
And then talk about how that, yeah.
So, and by the way, I'm gonna talk about
this very particular view of human robot interaction, right?
Which is not so much on the social side
or on the side of how do you have a good conversation
with the robot, what should the robot's appearance be?
It turns out that if you make robots taller versus shorter,
this has an effect on how people act with them.
So I'm not talking about that.
But I'm talking about this very kind of narrow thing,
which is you take, if you wanna take a task
that a robot can do in isolation,
in a lab out there in the world, but in isolation,
and now you're asking what does it mean for the robot
to be able to do this task for,
presumably what its actually end goal is,
which is to help some person.
That ends up changing the problem in two ways.
The first way it changes the problem is that
the robot is no longer the single agent acting.
That you have humans who also take actions
in that same space.
Cars navigating around people, robots around an office,
navigating around the people in that office.
If I send the robot over there in the cafeteria
to get me a coffee, then there's probably other people
reaching for stuff in the same space.
And so now you have your robot and you're in charge
of the actions that the robot is taking.
Then you have these people who are also making decisions
and taking actions in that same space.
And even if, you know, the robot knows what it should do
and all of that, just coexisting with these people, right?
Kind of getting the actions to gel well,
to mesh well together.
That's sort of the kind of problem number one.
And then there's problem number two,
which is, goes back to this notion of if I'm a programmer,
I can specify some objective for the robot
to go off and optimize and specify the task.
But if I put the robot in your home,
presumably you might have your own opinions about,
well, okay, I want my house clean,
but how do I want it cleaned?
And how should robot move, how close to me it should come
and all of that.
And so I think those are the two differences that you have.
You're acting around people and what you should be
optimizing for should satisfy the preferences
of that end user, not of your programmer who programmed you.
Yeah, and the preferences thing is tricky.
So figuring out those preferences,
be able to interactively adjust
to understand what the human is doing.
So really it boils down to understand the humans
in order to interact with them and in order to please them.
Right.
So why is this hard?
Yeah, why is understanding humans hard?
So I think there's two tasks about understanding humans
that in my mind are very, very similar,
but not everyone agrees.
So there's the task of being able to just anticipate
what people will do.
We all know that cars need to do this, right?
We all know that, well, if I navigate around some people,
the robot has to get some notion of,
okay, where is this person gonna be?
So that's kind of the prediction side.
And then there's what you were saying,
satisfying the preferences, right?
So adapting to the person's preferences,
knowing what to optimize for,
which is more this inference side,
this what does this person want?
What is their intent? What are their preferences?
And to me, those kind of go together
because I think that at the very least,
if you can understand, if you can look at human behavior
and understand what it is that they want,
then that's sort of the key enabler
to being able to anticipate what they'll do in the future.
Because I think that we're not arbitrary.
We make these decisions that we make,
we act in the way we do
because we're trying to achieve certain things.
And so I think that's the relationship between them.
Now, how complicated do these models need to be
in order to be able to understand what people want?
So we've gotten a long way in robotics
with something called inverse reinforcement learning,
§03Transcript · 18:17–26:54
which is the notion of if someone acts,
demonstrates how they want the thing done.
What is inverse reinforcement learning?
You just briefly said it.
Right, so it's the problem of take human behavior
and infer reward function from this.
So figure out what it is
that that behavior is optimal with respect to.
And it's a great way to think
about learning human preferences
in the sense of you have a car and the person can drive it
and then you can say, well, okay,
I can actually learn what the person is optimizing for.
I can learn their driving style,
or you can have people demonstrate
how they want the house clean.
And then you can say, okay, this is,
I'm getting the trade offs that they're making.
I'm getting the preferences that they want out of this.
And so we've been successful in robotics somewhat with this.
And it's based on a very simple model of human behavior.
It was remarkably simple,
which is that human behavior is optimal
with respect to whatever it is that people want, right?
So you make that assumption
and now you can kind of inverse through.
That's why it's called inverse,
well, really optimal control,
but also inverse reinforcement learning.
So this is based on utility maximization in economics.
Back in the forties, von Neumann and Morgenstern
were like, okay, people are making choices
by maximizing utility, go.
And then in the late fifties,
we had Luce and Shepherd come in and say,
people are a little bit noisy and approximate in that process.
So they might choose something kind of stochastically
with probability proportional to
how much utility something has.
So there's a bit of noise in there.
This has translated into robotics
and something that we call Boltzmann rationality.
So it's a kind of an evolution
of inverse reinforcement learning
that accounts for human noise.
And we've had some success with that too,
for these tasks where it turns out
people act noisily enough that you can't just do vanilla,
the vanilla version.
You can account for noise
and still infer what they seem to want based on this.
Then now we're hitting tasks where that's not enough.
And because...
What are examples of spatial tasks?
So imagine you're trying to control some robot,
that's fairly complicated.
You're trying to control a robot arm
because maybe you're a patient with a motor impairment
and you have this wheelchair mounted arm
and you're trying to control it around.
Or one task that we've looked at with Sergei is,
and our students did, is a lunar lander.
So I don't know if you know this Atari game,
it's called Lunar Lander.
It's really hard.
People really suck at landing the thing.
Mostly they just crash it left and right.
Okay, so this is the kind of task we imagine
you're trying to provide some assistance
to a person operating such a robot
where you want the kind of the autonomy to kick in,
figure out what it is that you're trying to do
and help you do it.
It's really hard to do that for, say, Lunar Lander
because people are all over the place.
And so they seem much more noisy than really irrational.
That's an example of a task
where these models are kind of failing us.
And it's not surprising because
we're talking about the 40s, utility, late 50s,
sort of noisy.
Then the 70s came and behavioral economics
started being a thing where people were like,
no, no, no, no, no, people are not rational.
People are messy and emotional and irrational
and have all sorts of heuristics
that might be domain specific.
And they're just a mess.
The mess.
So what does my robot do to understand
what you want?
And it's a very, it's very, that's why it's complicated.
It's, you know, for the most part,
we get away with pretty simple models until we don't.
And then the question is, what do you do then?
And I had days when I wanted to, you know,
pack my bags and go home and switch jobs
because it's just, it feels really daunting
to make sense of human behavior enough
that you can reliably understand what people want,
especially as, you know,
robot capabilities will continue to get developed.
You'll get these systems that are more and more capable
of all sorts of things.
And then you really want to make sure
that you're telling them the right thing to do.
What is that thing?
Well, read it in human behavior.
So if I just sat here quietly
and tried to understand something about you
by listening to you talk,
it would be harder than if I got to say something
and ask you and interact and control.
Can you, can the robot help its understanding of the human
by influencing the behavior by actually acting?
Yeah, absolutely.
So one of the things that's been exciting to me lately
is this notion that when you try to,
that when you try to think of the robotics problem as,
okay, I have a robot and it needs to optimize
for whatever it is that a person wants it to optimize
as opposed to maybe what a programmer said.
That problem we think of as a human robot
collaboration problem in which both agents get to act
in which the robot knows less than the human
because the human actually has access to,
you know, at least implicitly to what it is that they want.
They can't write it down, but they can talk about it.
They can give all sorts of signals.
They can demonstrate and,
but the robot doesn't need to sit there
and passively observe human behavior
and try to make sense of it.
The robot can act too.
And so there's these information gathering actions
that the robot can take to sort of solicit responses
that are actually informative.
So for instance, this is not for the purpose
of assisting people, but with kind of back to coordinating
with people in cars and all of that.
One thing that Dorsa did was,
so we were looking at cars being able to navigate
around people and you might not know exactly
the driving style of a particular individual
that's next to you,
but you wanna change lanes in front of them.
Navigating around other humans inside cars.
Yeah, good, good clarification question.
So you have an autonomous car and it's trying to navigate
the road around human driven vehicles.
Similar things ideas apply to pedestrians as well,
but let's just take human driven vehicles.
So now you're trying to change a lane.
Well, you could be trying to infer the driving style
of this person next to you.
You'd like to know if they're in particular,
if they're sort of aggressive or defensive,
if they're gonna let you kind of go in
or if they're gonna not.
And it's very difficult to just,
if you think that if you wanna hedge your bets
and say, ah, maybe they're actually pretty aggressive,
I shouldn't try this.
You kind of end up driving next to them
and driving next to them, right?
And then you don't know
because you're not actually getting the observations
that you're getting away.
Someone drives when they're next to you
and they just need to go straight.
It's kind of the same
regardless if they're aggressive or defensive.
And so you need to enable the robot
to reason about how it might actually be able
to gather information by changing the actions
that it's taking.
And then the robot comes up with these cool things
where it kind of nudges towards you
and then sees if you're gonna slow down or not.
Then if you slow down,
it sort of updates its model of you
and says, oh, okay, you're more on the defensive side.
So now I can actually like.
That's a fascinating dance.
That's so cool that you could use your own actions
to gather information.
That feels like a totally open,
exciting new world of robotics.
I mean, how many people are even thinking
about that kind of thing?
A handful of us, I'd say.
It's rare because it's actually leveraging human.
I mean, most roboticists,
I've talked to a lot of colleagues and so on,
are kind of, being honest, kind of afraid of humans.
Because they're messy and complicated, right?
I understand.
Going back to what we were talking about earlier,
right now we're kind of in this dilemma of, okay,
there are tasks that we can just assume
§04Transcript · 26:54–35:24
people are approximately rational for
and we can figure out what they want.
We can figure out their goals.
We can figure out their driving styles, whatever.
Cool.
There are these tasks that we can't.
So what do we do, right?
Do we pack our bags and go home?
And this one, I've had a little bit of hope recently.
And I'm kind of doubting myself
because what do I know that, you know,
50 years of behavioral economics hasn't figured out.
But maybe it's not really in contradiction
with the way that field is headed.
But basically one thing that we've been thinking about is,
instead of kind of giving up and saying
people are too crazy and irrational
for us to make sense of them,
maybe we can give them a bit the benefit of the doubt.
And maybe we can think of them
as actually being relatively rational,
but just under different assumptions about the world,
about how the world works, about, you know,
they don't have, when we think about rationality,
implicit assumption is, oh, they're rational,
and they're all the same assumptions and constraints
as the robot, right?
What, if this is the state of the world,
that's what they know.
This is the transition function, that's what they know.
This is the horizon, that's what they know.
But maybe the kind of this difference,
the way, the reason they can seem a little messy
and hectic, especially to robots,
is that perhaps they just make different assumptions
or have different beliefs.
Yeah, I mean, that's another fascinating idea
that this, our kind of anecdotal desire
to say that humans are irrational,
perhaps grounded in behavioral economics,
is that we just don't understand the constraints
and the rewards under which they operate.
And so our goal shouldn't be to throw our hands up
and say they're irrational,
it's to say, let's try to understand
what are the constraints.
What it is that they must be assuming
that makes this behavior make sense.
Good life lesson, right?
Good life lesson.
That's true, it's just outside of robotics.
That's just good to, that's communicating with humans.
That's just a good assume
that you just don't, sort of empathy, right?
It's a...
This is maybe there's something you're missing
and it's, you know, it especially happens to robots
cause they're kind of dumb and they don't know things.
And oftentimes people are sort of supra rational
and that they actually know a lot of things
that robots don't.
Sometimes like with the lunar lander,
the robot, you know, knows much more.
So it turns out that if you try to say,
look, maybe people are operating this thing
but assuming a much more simplified physics model
cause they don't get the complexity of this kind of craft
or the robot arm with seven degrees of freedom
with these inertias and whatever.
So maybe they have this intuitive physics model
which is not, you know, this notion of intuitive physics
is something that you studied actually in cognitive science
was like Josh Denenbaum, Tom Griffith's work on this stuff.
And what we found is that you can actually try
to figure out what physics model
kind of best explains human actions.
And then you can use that to sort of correct what it is
that they're commanding the craft to do.
So they might, you know, be sending the craft somewhere
but instead of executing that action,
you can sort of take a step back and say,
according to their intuitive,
if the world worked according to their intuitive physics
model, where do they think that the craft is going?
Where are they trying to send it to?
And then you can use the real physics, right?
The inverse of that to actually figure out
what you should do so that you do that
instead of where they were actually sending you
in the real world.
And I kid you not at work people land the damn thing
and you know, in between the two flags and all that.
So it's not conclusive in any way
but I'd say it's evidence that yeah,
maybe we're kind of underestimating humans in some ways
when we're giving up and saying,
yeah, they're just crazy noisy.
So then you try to explicitly try to model
the kind of worldview that they have.
That they have, that's right.
That's right.
And it's not too, I mean,
there's things in behavior economics too
that for instance have touched upon the planning horizon.
So there's this idea that there's bounded rationality
essentially and the idea that, well,
maybe we work under computational constraints.
And I think kind of our view recently has been
take the Bellman update in AI
and just break it in all sorts of ways by saying state,
no, no, no, the person doesn't get to see the real state.
Maybe they're estimating somehow.
Transition function, no, no, no, no, no.
Even the actual reward evaluation,
maybe they're still learning
about what it is that they want.
Like, you know, when you watch Netflix
and you know, you have all the things
and then you have to pick something,
imagine that, you know, the AI system interpreted
that choice as this is the thing you prefer to see.
Like, how are you going to know?
You're still trying to figure out what you like,
what you don't like, et cetera.
So I think it's important to also account for that.
So it's not irrationality,
because they're doing the right thing
under the things that they know.
Yeah, that's brilliant.
You mentioned recommender systems.
What kind of, and we were talking
about human robot interaction,
what kind of problem spaces are you thinking about?
So is it robots, like wheeled robots
with autonomous vehicles?
Is it object manipulation?
Like when you think
about human robot interaction in your mind,
and maybe I'm sure you can speak
for the entire community of human robot interaction.
But like, what are the problems of interest here?
And does it, you know, I kind of think
of open domain dialogue as human robot interaction,
and that happens not in the physical space,
but it could just happen in the virtual space.
So where's the boundaries of this field for you
when you're thinking about the things
we've been talking about?
Yeah, so I try to find kind of underlying,
I don't know what to even call them.
I try to work on, you know, I might call what I do,
the kind of working on the foundations
of algorithmic human robot interaction
and trying to make contributions there.
And it's important to me that whatever we do
is actually somewhat domain agnostic when it comes to,
is it about, you know, autonomous cars
or is it about quadrotors or is it about,
is this sort of the same underlying principles apply?
Of course, when you're trying to get
a particular domain to work,
you usually have to do some extra work
to adapt that to that particular domain.
But these things that we were talking about around,
well, you know, how do you model humans?
It turns out that a lot of systems need
to core benefit from a better understanding
of how human behavior relates to what people want
and need to predict human behavior,
physical robots of all sorts and beyond that.
And so I used to do manipulation.
I used to be, you know, picking up stuff
and then I was picking up stuff with people around.
And now it's sort of very broad
when it comes to the application level,
but in a sense, very focused on, okay,
how does the problem need to change?
How do the algorithms need to change
when we're not doing a robot by itself?
You know, emptying the dishwasher,
but we're stepping outside of that.
I thought that popped into my head just now.
On the game theoretic side,
I think you said this really interesting idea
of using actions to gain more information.
But if we think of sort of game theory,
the humans that are interacting with you,
with you, the robot?
Wow, I'm thinking the identity of the robot.
Yeah, I do that all the time.
Yeah, is they also have a world model of you
and you can manipulate that.
I mean, if we look at autonomous vehicles,
people have a certain viewpoint.
You said with the kids, people see Alexa in a certain way.
Is there some value in trying to also optimize
how people see you as a robot?
Or is that a little too far away from the specifics
of what we can solve right now?
So, well, both, right?
§05Transcript · 35:24–44:09
So it's really interesting.
And we've seen a little bit of progress on this problem,
on pieces of this problem.
So you can, again, it kind of comes down
to how complicated does the human model need to be?
But in one piece of work that we were looking at,
we just said, okay, there's these parameters
that are internal to the robot
and what the robot is about to do,
or maybe what objective,
what driving style the robot has or something like that.
And what we're gonna do is we're gonna set up a system
where part of the state is the person's belief
over those parameters.
And now when the robot acts,
that the person gets new evidence
about this robot internal state.
And so they're updating their mental model of the robot.
So if they see a car that sort of cuts someone off,
they're like, oh, that's an aggressive car.
They know more.
If they see sort of a robot head towards a particular door,
they're like, oh yeah, the robot's trying to get
to that door.
So this thing that we have to do with humans
to try and understand their goals and intentions,
humans are inevitably gonna do that to robots.
And then that raises this interesting question
that you asked, which is, can we do something about that?
This is gonna happen inevitably,
but we can sort of be more confusing
or less confusing to people.
And it turns out you can optimize
for being more informative and less confusing
if you have an understanding of how your actions
are being interpreted by the human,
and how they're using these actions to update their belief.
And honestly, all we did is just Bayes rule.
Basically, okay, the person has a belief,
they see an action, they make some assumptions
about how the robot generates its actions,
presumably as being rational,
because robots are rational.
It's reasonable to assume that about them.
And then they incorporate that new piece of evidence
in the Bayesian sense in their belief,
and they obtain a posterior.
And now the robot is trying to figure out
what actions to take such that it steers
the person's belief to put as much probability mass
as possible on the correct parameters.
So that's kind of a mathematical formalization of that.
But my worry, and I don't know if you wanna go there
with me, but I talk about this quite a bit.
The kids talking to Alexa disrespectfully worries me.
I worry in general about human nature.
Like I said, I grew up in Soviet Union, World War II,
I'm a Jew too, so with the Holocaust and everything.
I just worry about how we humans sometimes treat the other,
the group that we call the other, whatever it is.
Through human history, the group that's the other
has been changed faces.
But it seems like the robot will be the other, the other,
the next other.
And one thing is it feels to me
that robots don't get no respect.
They get shoved around.
Shoved around, and is there, one, at the shallow level,
for a better experience, it seems that robots
need to talk back a little bit.
Like my intuition says, I mean, most companies
from sort of Roomba, autonomous vehicle companies
might not be so happy with the idea that a robot
has a little bit of an attitude.
But I feel, it feels to me that that's necessary
to create a compelling experience.
Like we humans don't seem to respect anything
that doesn't give us some attitude.
That, or like a mix of mystery and attitude and anger
and that threatens us subtly, maybe passive aggressively.
I don't know.
It seems like we humans, yeah, need that.
Do you, what are your, is there something,
you have thoughts on this?
All right, I'll give you two thoughts on this.
Okay, sure.
One is, one is, it's, we respond to, you know,
someone being assertive, but we also respond
to someone being vulnerable.
So I think robots, my first thought is that
robots get shoved around and bullied a lot
because they're sort of, you know, tempting
and they're sort of showing off
or they appear to be showing off.
And so I think going back to these things
we were talking about in the beginning
of making robots a little more, a little more expressive,
a little bit more like, eh, that wasn't cool to do.
And now I'm bummed, right?
I think that that can actually help
because people can't help but anthropomorphize
and respond to that.
Even that though, the emotion being communicated
is not in any way a real thing.
And people know that it's not a real thing
because they know it's just a machine.
We're still interpreting, you know, we watch,
there's this famous psychology experiment
with little triangles and kind of dots on a screen
and a triangle is chasing the square
and you get really angry at the darn triangle
because why is it not leaving the square alone?
So that's, yeah, we can't help.
So that was the first thought.
The vulnerability, that's really interesting that,
I think of like being, pushing back, being assertive
as the only mechanism of getting,
of forming a connection, of getting respect,
but perhaps vulnerability,
perhaps there's other mechanisms that are less threatening.
Yeah.
Is there?
Well, I think, well, a little bit, yes,
but then this other thing that we can think about is,
it goes back to what you were saying,
that interaction is really game theoretic, right?
So the moment you're taking actions in a space,
the humans are taking actions in that same space,
but you have your own objective, which is, you know,
you're a car, you need to get your passenger
to the destination.
And then the human nearby has their own objective,
which somewhat overlaps with you, but not entirely.
You're not interested in getting into an accident
with each other, but you have different destinations
and you wanna get home faster
and they wanna get home faster.
And that's a general sum game at that point.
And so that's, I think that's what,
treating it as such is kind of a way we can step outside
of this kind of mode that,
where you try to anticipate what people do
and you don't realize you have any influence over it
while still protecting yourself
because you're understanding that people also understand
that they can influence you.
And it's just kind of back and forth is this negotiation,
which is really talking about different equilibria
of a game.
The very basic way to solve coordination
is to just make predictions about what people will do
and then stay out of their way.
And that's hard for the reasons we talked about,
which is how you have to understand people's intentions
implicitly, explicitly, who knows,
but somehow you have to get enough of an understanding
of that to be able to anticipate what happens next.
And so that's challenging.
But then it's further challenged by the fact
that people change what they do based on what you do
because they don't plan in isolation either, right?
So when you see cars trying to merge on a highway
and not succeeding, one of the reasons this can be
is because they look at traffic that keeps coming,
they predict what these people are planning on doing,
which is to just keep going,
and then they stay out of the way
because there's no feasible plan, right?
Any plan would actually intersect
with one of these other people.
So that's bad, so you get stuck there.
So now kind of if you start thinking about it as no, no, no,
actually these people change what they do
depending on what the car does.
Like if the car actually tries to kind of inch itself forward,
they might actually slow down and let the car in.
And now taking advantage of that,
well, that's kind of the next level.
We call this like this underactuated system idea
where it's kind of underactuated system robotics,
but it's kind of, you're influenced
these other degrees of freedom,
but you don't get to decide what they do.
I've somewhere seen you mention it,
the human element in this picture as underactuated.
So you understand underactuated robotics
is that you can't fully control the system.
You can't go in arbitrary directions
in the configuration space.
Under your control.
Yeah, it's a very simple way of underactuation
where basically there's literally these degrees of freedom
that you can control,
and these degrees of freedom that you can't,
but you influence them.
And I think that's the important part
is that they don't do whatever, regardless of what you do,
that what you do influences what they end up doing.
I just also like the poetry of calling human robot
interaction an underactuated robotics problem.
§06Transcript · 44:09–53:19
And you also mentioned sort of nudging.
It seems that they're, I don't know.
I think about this a lot in the case of pedestrians
I've collected hundreds of hours of videos.
I like to just watch pedestrians.
And it seems that.
It's a funny hobby.
Yeah, it's weird.
Cause I learn a lot.
I learned a lot about myself,
about our human behavior, from watching pedestrians,
watching people in their environment.
Basically crossing the street
is like you're putting your life on the line.
I don't know, tens of millions of time in America every day
is people are just like playing this weird game of chicken
when they cross the street,
especially when there's some ambiguity
about the right of way.
That has to do either with the rules of the road
or with the general personality of the intersection
based on the time of day and so on.
And this nudging idea,
it seems that people don't even nudge.
They just aggressively take, make a decision.
Somebody, there's a runner that gave me this advice.
I sometimes run in the street,
not in the street, on the sidewalk.
And he said that if you don't make eye contact with people
when you're running, they will all move out of your way.
It's called civil inattention.
Civil inattention, that's a thing.
Oh wow, I need to look this up, but it works.
What is that?
My sense was if you communicate like confidence
in your actions that you're unlikely to deviate
from the action that you're following,
that's a really powerful signal to others
that they need to plan around your actions.
As opposed to nudging where you're sort of hesitantly,
then the hesitation might communicate
that you're still in the dance and the game
that they can influence with their own actions.
I've recently had a conversation with Jim Keller,
who's a sort of this legendary chip architect,
but he also led the autopilot team for a while.
And his intuition that driving is fundamentally
still like a ballistics problem.
Like you can ignore the human element
that is just not hitting things.
And you can kind of learn the right dynamics
required to do the merger and all those kinds of things.
And then my sense is, and I don't know if I can provide
sort of definitive proof of this,
but my sense is like an order of magnitude
are more difficult when humans are involved.
Like it's not simply object collision avoidance problem.
Where does your intuition,
of course, nobody knows the right answer here,
but where does your intuition fall on the difficulty,
fundamental difficulty of the driving problem
when humans are involved?
Yeah, good question.
I have many opinions on this.
Imagine downtown San Francisco.
Yeah, it's crazy, busy, everything.
Okay, now take all the humans out.
No pedestrians, no human driven vehicles,
no cyclists, no people on little electric scooters
zipping around, nothing.
I think we're done.
I think driving at that point is done.
We're done.
There's nothing really that still needs
to be solved about that.
Well, let's pause there.
I think I agree with you and I think a lot of people
that will hear will agree with that,
but we need to sort of internalize that idea.
So what's the problem there?
Cause we might not quite yet be done with that.
Cause a lot of people kind of focus
on the perception problem.
A lot of people kind of map autonomous driving
into how close are we to solving,
being able to detect all the, you know,
the drivable area, the objects in the scene.
Do you see that as a, how hard is that problem?
So your intuition there behind your statement
was we might have not solved it yet,
but we're close to solving basically the perception problem.
I think the perception problem, I mean,
and by the way, a bunch of years ago,
this would not have been true.
And a lot of issues in the space were coming
from the fact that, oh, we don't really, you know,
we don't know what's where.
But I think it's fairly safe to say that at this point,
although you could always improve on things
and all of that, you can drive through downtown San Francisco
if there are no people around.
There's no really perception issues
standing in your way there.
I think perception is hard, but yeah, it's, we've made
a lot of progress on the perception,
so I had to undermine the difficulty of the problem.
I think everything about robotics is really difficult,
of course, I think that, you know, the planning problem,
the control problem, all very difficult,
but I think what's, what makes it really kind of, yeah.
It might be, I mean, you know,
and I picked downtown San Francisco,
it's adapting to, well, now it's snowing,
now it's no longer snowing, now it's slippery in this way,
now it's the dynamics part could,
I could imagine being still somewhat challenging, but.
No, the thing that I think worries us,
and our intuition's not good there,
is the perception problem at the edge cases.
Sort of downtown San Francisco, the nice thing,
it's not actually, it may not be a good example because.
Because you know what you're getting from,
well, there's like crazy construction zones
and all of that. Yeah, but the thing is,
you're traveling at slow speeds,
so like it doesn't feel dangerous.
To me, what feels dangerous is highway speeds,
when everything is, to us humans, super clear.
Yeah, I'm assuming LiDAR here, by the way.
I think it's kind of irresponsible to not use LiDAR.
That's just my personal opinion.
That's, I mean, depending on your use case,
but I think like, you know, if you have the opportunity
to use LiDAR, in a lot of cases, you might not.
Good, your intuition makes more sense now.
So you don't think vision.
I really just don't know enough to say,
well, vision alone, what, you know, what's like,
there's a lot of, how many cameras do you have?
Is it, how are you using them?
I don't know. There's details.
There's all, there's all sorts of details.
I imagine there's stuff that's really hard
to actually see, you know, how do you deal with glare,
exactly what you were saying,
stuff that people would see that you don't.
I think I have, more of my intuition comes from systems
that can actually use LiDAR as well.
Yeah, and until we know for sure,
it makes sense to be using LiDAR.
That's kind of the safety focus.
But then the sort of the,
I also sympathize with the Elon Musk statement
of LiDAR is a crutch.
It's a fun notion to think that the things that work today
is a crutch for the invention of the things
that will work tomorrow, right?
Like it, it's kind of true in the sense that if,
you know, we want to stick to the comfort zone,
you see this in academic and research settings
all the time, the things that work force you
to not explore outside, think outside the box.
I mean, that happens all the time.
The problem is in the safety critical systems,
you kind of want to stick with the things that work.
So it's an interesting and difficult trade off
in the case of real world sort of safety critical
robotic systems, but so your intuition is,
just to clarify, how, I mean,
how hard is this human element for,
like how hard is driving
when this human element is involved?
Are we years, decades away from solving it?
But perhaps actually the year isn't the thing I'm asking.
It doesn't matter what the timeline is,
but do you think we're, how many breakthroughs
are we away from in solving
the human robotic interaction problem
to get this, to get this right?
I think it, in a sense, it really depends.
I think that, you know, we were talking about how,
well, look, it's really hard
because anticipate what people do is hard.
And on top of that, playing the game is hard.
But I think we sort of have the fundamental,
some of the fundamental understanding for that.
And then you already see that these systems
are being deployed in the real world,
you know, even driverless.
Like there's, I think now a few companies
that don't have a driver in the car in some small areas.
I got a chance to, I went to Phoenix and I,
I shot a video with Waymo and I needed to get
that video out.
People have been giving me slack,
but there's incredible engineering work being done there.
And it's one of those other seminal moments
for me in my life to be able to, it sounds silly,
but to be able to drive without a ride, sorry,
without a driver in the seat.
§07Transcript · 53:19–1:02:38
I mean, that was an incredible robotics.
I was driven by a robot without being able to take over,
without being able to take the steering wheel.
That's a magical, that's a magical moment.
So in that regard, in those domains,
at least for like Waymo, they're solving that human,
there's, I mean, they're going, I mean, it felt fast
because you're like freaking out at first.
That was, this is my first experience,
but it's going like the speed limit, right?
30, 40, whatever it is.
And there's humans and it deals with them quite well.
It detects them, it negotiates the intersections,
the left turns and all of that.
So at least in those domains, it's solving them.
The open question for me is like, how quickly can we expand?
You know, that's the, you know,
outside of the weather conditions,
all of those kinds of things,
how quickly can we expand to like cities like San Francisco?
Yeah, and I wouldn't say that it's just, you know,
now it's just pure engineering and it's probably the,
I mean, and by the way,
I'm speaking kind of very generally here as hypothesizing,
but I think that there are successes
and yet no one is everywhere out there.
So that seems to suggest that things can be expanded
and can be scaled and we know how to do a lot of things,
but there's still probably, you know,
new algorithms or modified algorithms
that you still need to put in there
as you learn more and more about new challenges
that you get faced with.
How much of this problem do you think can be learned
through end to end?
Is it the success of machine learning
and reinforcement learning?
How much of it can be learned from sort of data
from scratch and how much,
which most of the success of autonomous vehicle systems
have a lot of heuristics and rule based stuff on top,
like human expertise injected forced into the system
to make it work.
What's your sense?
How much, what will be the role of learning
in the near term and long term?
I think on the one hand that learning is inevitable here,
right?
I think on the other hand that when people characterize
the problem as it's a bunch of rules
that some people wrote down,
versus it's an end to end RL system or imitation learning,
then maybe there's kind of something missing
from maybe that's more.
So for instance, I think a very, very useful tool
in this sort of problem,
both in how to generate the car's behavior
and robots in general and how to model human beings
is actually planning, search optimization, right?
So robotics is the sequential decision making problem.
And when a robot can figure out on its own
how to achieve its goal without hitting stuff
and all that stuff, right?
All the good stuff for motion planning 101,
I think of that as very much AI,
not this is some rule or something.
There's nothing rule based around that, right?
It's just you're searching through a space
and figuring out are you optimizing through a space
and figure out what seems to be the right thing to do.
And I think it's hard to just do that
because you need to learn models of the world.
And I think it's hard to just do the learning part
where you don't bother with any of that,
because then you're saying, well, I could do imitation,
but then when I go off distribution, I'm really screwed.
Or you can say, I can do reinforcement learning,
which adds a lot of robustness,
but then you have to do either reinforcement learning
in the real world, which sounds a little challenging
or that trial and error, you know,
or you have to do reinforcement learning in simulation.
And then that means, well, guess what?
You need to model things, at least to model people,
model the world enough that whatever policy you get of that
is actually fine to roll out in the world
and do some additional learning there.
So. Do you think simulation, by the way, just a quick tangent
has a role in the human robot interaction space?
Like, is it useful?
It seems like humans, everything we've been talking about
are difficult to model and simulate.
Do you think simulation has a role in this space?
I do.
I think so because you can take models
and train with them ahead of time, for instance.
You can.
But the models, sorry to interrupt,
the models are sort of human constructed or learned?
I think they have to be a combination
because if you get some human data and then you say,
this is how, this is gonna be my model of the person.
What are for simulation and training
or for just deployment time?
And that's what I'm planning with
as my model of how people work.
Regardless, if you take some data
and you don't assume anything else and you just say,
okay, this is some data that I've collected.
Let me fit a policy to how people work based on that.
What tends to happen is you collected some data
and some distribution, and then now your robot
sort of computes a best response to that, right?
It's sort of like, what should I do
if this is how people work?
And easily goes off of distribution
where that model that you've built of the human
completely sucks because out of distribution,
you have no idea, right?
If you think of all the possible policies
and then you take only the ones that are consistent
with the human data that you've observed,
that still leads a lot of, a lot of things could happen
outside of that distribution where you're confident
then you know what's going on.
By the way, that's, I mean, I've gotten used
to this terminology of not a distribution,
but it's such a machine learning terminology
because it kind of assumes,
so distribution is referring to the data
that you've seen.
The set of states that you encounter
at training time. They've encountered so far
at training time. Yeah.
But it kind of also implies that there's a nice
like statistical model that represents that data.
So out of distribution feels like, I don't know,
it raises to me philosophical questions
of how we humans reason out of distribution,
reason about things that are completely,
we haven't seen before.
And so, and what we're talking about here is
how do we reason about what other people do
in situations where we haven't seen them?
And somehow we just magically navigate that.
I can anticipate what will happen in situations
that are even novel in many ways.
And I have a pretty good intuition for,
I don't always get it right, but you know,
and I might be a little uncertain and so on.
But I think it's this that if you just rely on data,
you know, there's just too many possibilities,
there's too many policies out there that fit the data.
And by the way, it's not just state,
it's really kind of history of state,
cause to really be able to anticipate
what the person will do,
it kind of depends on what they've been doing so far,
cause that's the information you need to kind of,
at least implicitly sort of say,
oh, this is the kind of person that this is,
this is probably what they're trying to do.
So anyway, it's like you're trying to map history of states
to actions, there's many mappings.
And history meaning like the last few seconds
or the last few minutes or the last few months.
Who knows, who knows how much you need, right?
In terms of if your state is really like the positions
of everything or whatnot and velocities,
who knows how much you need.
And then there's so many mappings.
And so now you're talking about
how do you regularize that space?
What priors do you impose or what's the inductive bias?
So, you know, there's all very related things
to think about it.
Basically, what are assumptions that we should be making
such that these models actually generalize
outside of the data that we've seen?
And now you're talking about, well, I don't know,
what can you assume?
Maybe you can assume that people like actually
have intentions and that's what drives their actions.
Maybe that's, you know, the right thing to do
when you haven't seen data very nearby
that tells you otherwise.
I don't know, it's a very open question.
Do you think sort of that one of the dreams
of artificial intelligence was to solve
common sense reasoning, whatever the heck that means.
Do you think something like common sense reasoning
has to be solved in part to be able to solve this dance
of human robot interaction, the driving space
or human robot interaction in general?
Do you have to be able to reason about these kinds
of common sense concepts of physics,
of, you know, all the things we've been talking about
humans, I don't even know how to express them with words,
but the basics of human behavior, a fear of death.
So like, to me, it's really important to encode
§08Transcript · 1:02:38–1:12:10
in some kind of sense, maybe not, maybe it's implicit,
but it feels that it's important to explicitly encode
the fear of death, that people don't wanna die.
Because it seems silly, but like the game of chicken
that involves with the pedestrian crossing the street
is playing with the idea of mortality.
Like we really don't wanna die.
It's not just like a negative reward.
I don't know, it just feels like all these human concepts
have to be encoded.
Do you share that sense or is this a lot simpler
than I'm making out to be?
I think it might be simpler.
And I'm the person who likes to complicate things.
I think it might be simpler than that.
Because it turns out, for instance,
if you say model people in the very,
I'll call it traditional, I don't know if it's fair
to look at it as a traditional way,
but you know, calling people as,
okay, they're rational somehow,
the utilitarian perspective.
Well, in that, once you say that,
you automatically capture that they have an incentive
to keep on being.
You know, Stuart likes to say,
you can't fetch the coffee if you're dead.
Stuart Russell, by the way.
That's a good line.
So when you're sort of treating agents
as having these objectives, these incentives,
humans or artificial, you're kind of implicitly modeling
that they'd like to stick around
so that they can accomplish those goals.
So I think in a sense,
maybe that's what draws me so much
to the rationality framework,
even though it's so broken,
we've been able to, it's been such a useful perspective.
And like we were talking about earlier,
what's the alternative?
I give up and go home or, you know,
I just use complete black boxes,
but then I don't know what to assume out of distribution
that come back to this.
It's just, it's been a very fruitful way
to think about the problem
in a very more positive way, right?
People aren't just crazy.
Maybe they make more sense than we think.
But I think we also have to somehow be ready for it
to be wrong, be able to detect
when these assumptions aren't holding,
be all of that stuff.
Let me ask sort of another small side of this
that we've been talking about
the pure autonomous driving problem,
but there's also relatively successful systems
already deployed out there in what you may call
like level two autonomy or semi autonomous vehicles,
whether that's Tesla Autopilot,
work quite a bit with Cadillac SuperGuru system,
which has a driver facing camera that detects your state.
There's a bunch of basically lane centering systems.
What's your sense about this kind of way of dealing
with the human robot interaction problem
by having a really dumb robot
and relying on the human to help the robot out
to keep them both alive?
Is that from the research perspective,
how difficult is that problem?
And from a practical deployment perspective,
is that a fruitful way to approach
this human robot interaction problem?
I think what we have to be careful about there
is to not, it seems like some of these systems,
not all are making this underlying assumption
that if, so I'm a driver and I'm now really not driving,
but supervising and my job is to intervene, right?
And so we have to be careful with this assumption
that when I'm, if I'm supervising,
I will be just as safe as when I'm driving.
That I will, if I wouldn't get into some kind of accident,
if I'm driving, I will be able to avoid that accident
when I'm supervising too.
And I think I'm concerned about this assumption
from a few perspectives.
So from a technical perspective,
it's that when you let something kind of take control
and do its thing, and it depends on what that thing is,
obviously, and how much it's taking control
and how, what things are you trusting it to do.
But if you let it do its thing and take control,
it will go to what we might call off policy
from the person's perspective state.
So states that the person wouldn't actually
find themselves in if they were the ones driving.
And the assumption that the person functions
just as well there as they function in the states
that they would normally encounter
is a little questionable.
Now, another part is the kind of the human factor side
of this, which is that I don't know about you,
but I think I definitely feel like I'm experiencing things
very differently when I'm actively engaged in the task
versus when I'm a passive observer.
Like even if I try to stay engaged, right?
It's very different than when I'm actually
actively making decisions.
And you see this in life in general.
Like you see students who are actively trying
to come up with the answer, learn this thing better
than when they're passively told the answer.
I think that's somewhat related.
And I think people have studied this in human factors
for airplanes.
And I think it's actually fairly established
that these two are not the same.
So.
On that point, because I've gotten a huge amount
of heat on this and I stand by it.
Okay.
Because I know the human factors community well
and the work here is really strong.
And there's many decades of work showing exactly
what you're saying.
Nevertheless, I've been continuously surprised
that much of the predictions of that work has been wrong
in what I've seen.
So what we have to do,
I still agree with everything you said,
but we have to be a little bit more open minded.
So the, I'll tell you, there's a few surprising things
that supervise, like everything you said to the word
is actually exactly correct.
But it doesn't say, what you didn't say
is that these systems are,
you said you can't assume a bunch of things,
but we don't know if these systems are fundamentally unsafe.
That's still unknown.
There's a lot of interesting things,
like I'm surprised by the fact, not the fact,
that what seems to be anecdotally from,
well, from large data collection that we've done,
but also from just talking to a lot of people,
when in the supervisory role of semi autonomous systems
that are sufficiently dumb, at least,
which is, that might be the key element,
is the systems have to be dumb.
The people are actually more energized as observers.
So they're actually better,
they're better at observing the situation.
So there might be cases in systems,
if you get the interaction right,
where you, as a supervisor,
will do a better job with the system together.
I agree, I think that is actually really possible.
I guess mainly I'm pointing out that if you do it naively,
you're implicitly assuming something,
that assumption might actually really be wrong.
But I do think that if you explicitly think about
what the agent should do
so that the person still stays engaged.
What the, so that you essentially empower the person
to do more than they could,
that's really the goal, right?
Is you still have a driver,
so you wanna empower them to be so much better
than they would be by themselves.
And that's different, it's a very different mindset
than I want them to basically not drive, right?
And, but be ready to sort of take over.
So one of the interesting things we've been talking about
is the rewards, that they seem to be fundamental too,
the way robots behaves.
So broadly speaking,
we've been talking about utility functions and so on,
but could you comment on how do we approach
the design of reward functions?
Like, how do we come up with good reward functions?
Well, really good question,
because the answer is we don't.
This was, you know, I used to think,
I used to think about how,
well, it's actually really hard to specify rewards
for interaction because it's really supposed to be
what the people want, and then you really, you know,
we talked about how you have to customize
what you wanna do to the end user.
But I kind of realized that even if you take
the interactive component away,
it's still really hard to design reward functions.
So what do I mean by that?
I mean, if we assume this sort of AI paradigm
in which there's an agent and his job is to optimize
some objectives, some reward, utility, loss, whatever, cost,
if you write it out, maybe it's a set,
depending on the situation or whatever it is,
if you write that out and then you deploy the agent,
you'd wanna make sure that whatever you specified
§09Transcript · 1:12:10–1:20:53
incentivizes the behavior you want from the agent
in any situation that the agent will be faced with, right?
So I do motion planning on my robot arm,
I specify some cost function like, you know,
this is how far away you should try to stay,
so much it matters to stay away from people,
and this is how much it matters to be able to be efficient
and blah, blah, blah, right?
I need to make sure that whatever I specified,
those constraints or trade offs or whatever they are,
that when the robot goes and solves that problem
in every new situation,
that behavior is the behavior that I wanna see.
And what I've been finding is
that we have no idea how to do that.
Basically, what I can do is I can sample,
I can think of some situations
that I think are representative of what the robot will face,
and I can tune and add and tune some reward function
until the optimal behavior is what I want
on those situations,
which first of all is super frustrating
because, you know, through the miracle of AI,
we've taken, we don't have to specify rules
for behavior anymore, right?
The, who were saying before,
the robot comes up with the right thing to do,
you plug in this situation,
it optimizes right in that situation, it optimizes,
but you have to spend still a lot of time
on actually defining what it is
that that criteria should be,
making sure you didn't forget
about 50 bazillion things that are important
and how they all should be combining together
to tell the robot what's good and what's bad
and how good and how bad.
And so I think this is a lesson that I don't know,
kind of, I guess I close my eyes to it for a while
cause I've been, you know,
tuning cost functions for 10 years now,
but it's really strikes me that,
yeah, we've moved the tuning
and the like designing of features or whatever
from the behavior side into the reward side.
And yes, I agree that there's way less of it,
but it still seems really hard
to anticipate any possible situation
and make sure you specify a reward function
that when optimized will work well
in every possible situation.
So you're kind of referring to unintended consequences
or just in general, any kind of suboptimal behavior
that emerges outside of the things you said,
out of distribution.
Suboptimal behavior that is, you know, actually optimal.
I mean, this, I guess the idea of unintended consequences,
you know, it's optimal respect to what you specified,
but it's not what you want.
And there's a difference between those.
But that's not fundamentally a robotics problem, right?
That's a human problem.
So like. That's the thing, right?
So there's this thing called Goodhart's law,
which is you set a metric for an organization
and the moment it becomes a target
that people actually optimize for,
it's no longer a good metric.
What's it called?
Goodhart's law.
Goodhart's law.
So the moment you specify a metric,
it stops doing its job.
Yeah, it stops doing its job.
So there's, yeah, there's such a thing
as optimizing for things and, you know,
failing to think ahead of time
of all the possible things that might be important.
And so that's, so that's interesting
because Historia works a lot on reward learning
from the perspective of customizing to the end user,
but it really seems like it's not just the interaction
with the end user that's a problem of the human
and the robot collaborating
so that the robot can do what the human wants, right?
This kind of back and forth, the robot probing,
the person being informative, all of that stuff
might be actually just as applicable
to this kind of maybe new form of human robot interaction,
which is the interaction between the robot
and the expert programmer, roboticist designer
in charge of actually specifying
what the heck the robot should do,
specifying the task for the robot.
That's fascinating.
That's so cool, like collaborating on the reward design.
Right, collaborating on the reward design.
And so what does it mean, right?
What does it, when we think about the problem,
not as someone specifies all of your job is to optimize,
and we start thinking about you're in this interaction
and this collaboration.
And the first thing that comes up is
when the person specifies a reward, it's not, you know,
gospel, it's not like the letter of the law.
It's not the definition of the reward function
you should be optimizing,
because they're doing their best,
but they're not some magic perfect oracle.
And the sooner we start understanding that,
I think the sooner we'll get to more robust robots
that function better in different situations.
And then you have kind of say, okay, well,
it's almost like robots are over learning,
over putting too much weight on the reward specified
by definition, and maybe leaving a lot of other information
on the table, like what are other things we could do
to actually communicate to the robot
about what we want them to do besides attempting
to specify a reward function.
Yeah, you have this awesome,
and again, I love the poetry of it, of leaked information.
So you mentioned humans leak information
about what they want, you know,
leak reward signal for the robot.
So how do we detect these leaks?
What is that?
Yeah, what are these leaks?
Whether it just, I don't know,
those were just recently saw it, read it,
I don't know where from you,
and it's gonna stick with me for a while for some reason,
because it's not explicitly expressed.
It kind of leaks indirectly from our behavior.
From what we do, yeah, absolutely.
So I think maybe some surprising bits, right?
So we were talking before about, I'm a robot arm,
it needs to move around people, carry stuff,
put stuff away, all of that.
And now imagine that, you know,
the robot has some initial objective
that the programmer gave it
so they can do all these things functionally.
It's capable of doing that.
And now I noticed that it's doing something
and maybe it's coming too close to me, right?
And maybe I'm the designer,
maybe I'm the end user and this robot is now in my home.
And I push it away.
So I push away because, you know,
it's a reaction to what the robot is currently doing.
And this is what we call physical human robot interaction.
And now there's a lot of interesting work
on how the heck do you respond to physical human
robot interaction?
What should the robot do if such an event occurs?
And there's sort of different schools of thought.
Well, you know, you can sort of treat it
the control theoretic way and say,
this is a disturbance that you must reject.
You can sort of treat it more kind of heuristically
and say, I'm gonna go into some like gravity compensation
mode so that I'm easily maneuverable around.
I'm gonna go in the direction that the person pushed me.
And to us, part of realization has been
that that is signal that communicates about the reward.
Because if my robot was moving in an optimal way
and I intervened, that means that I disagree
with his notion of optimality, right?
Whatever it thinks is optimal is not actually optimal.
And sort of optimization problems aside,
that means that the cost function,
the reward function is incorrect,
or at least is not what I want it to be.
How difficult is that signal to interpret
and make actionable?
So like, cause this connects
to our autonomous vehicle discussion
where they're in the semi autonomous vehicle
or autonomous vehicle when a safety driver
disengages the car, like,
but they could have disengaged it for a million reasons.
Yeah, so that's true.
Again, it comes back to, can you structure a little bit
your assumptions about how human behavior
relates to what they want?
And you can, one thing that we've done is
literally just treated this external torque
that they applied as, when you take that
and you add it with what the torque
the robot was already applying,
that overall action is probably relatively optimal
in respect to whatever it is that the person wants.
And then that gives you information
about what it is that they want.
So you can learn that people want you
to stay further away from them.
Now you're right that there might be many things
that explain just that one signal
and that you might need much more data than that
§10Transcript · 1:20:53–1:29:38
for the person to be able to shape
your reward function over time.
You can also do this info gathering stuff
that we were talking about.
Not that we've done that in that context,
just to clarify, but it's definitely something
we thought about where you can have the robot
start acting in a way, like if there's
a bunch of different explanations, right?
It moves in a way where it sees if you correct it
in some other way or not,
and then kind of actually plans its motion
so that it can disambiguate
and collect information about what you want.
Anyway, so that's one way,
that's kind of sort of leaked information,
maybe even more subtle leaked information
is if I just press the E stop, right?
I just, I'm doing it out of panic
because the robot is about to do something bad.
There's again, information there, right?
Okay, the robot should definitely stop,
but it should also figure out
that whatever it was about to do was not good.
And in fact, it was so not good
that stopping and remaining stopped for a while
was a better trajectory for it
than whatever it is that it was about to do.
And that again is information about
what are my preferences, what do I want?
Speaking of E stops, what are your expert opinions
on the three laws of robotics from Isaac Asimov
that don't harm humans, obey orders, protect yourself?
I mean, it's such a silly notion,
but I speak to so many people these days,
just regular folks, just, I don't know,
my parents and so on about robotics.
And they kind of operate in that space of,
you know, imagining our future with robots
and thinking what are the ethical,
how do we get that dance right?
I know the three laws might be a silly notion,
but do you think about like
what universal reward functions that might be
that we should enforce on the robots of the future?
Or is that a little too far out and it doesn't,
or is the mechanism that you just described,
it shouldn't be three laws,
it should be constantly adjusting kind of thing.
I think it should constantly be adjusting kind of thing.
You know, the issue with the laws is,
I don't even, you know, they're words
and I have to write math
and have to translate them into math.
What does it mean to?
What does harm mean?
What is, it's not math.
Obey what, right?
Cause we just talked about how
you try to say what you want,
but you don't always get it right.
And you want these machines to do what you want,
not necessarily exactly what you literally,
so you don't want them to take you literally.
You wanna take what you say and interpret it in context.
And that's what we do with the specified rewards.
We don't take them literally anymore from the designer.
We, not we as a community, we as, you know,
some members of my group, we,
and some of our collaborators like Peter Beal
and Stuart Russell, we sort of say,
okay, the designer specified this thing,
but I'm gonna interpret it not as,
this is the universal reward function
that I shall always optimize always and forever,
but as this is good evidence about what the person wants.
And I should interpret that evidence
in the context of these situations that it was specified for.
Cause ultimately that's what the designer thought about.
That's what they had in mind.
And really them specifying reward function
that works for me in all these situations
is really kind of telling me that whatever behavior
that incentivizes must be good behavior
with respect to the thing
that I should actually be optimizing for.
And so now the robot kind of has uncertainty
about what it is that it should be,
what its reward function is.
And then there's all these additional signals
that we've been finding that it can kind of continually
learn from and adapt its understanding of what people want.
Every time the person corrects it, maybe they demonstrate,
maybe they stop, hopefully not, right?
One really, really crazy one is the environment itself.
Like our world, you don't, it's not, you know,
you observe our world and the state of it.
And it's not that you're seeing behavior
and you're saying, oh, people are making decisions
that are rational, blah, blah, blah.
It's, but our world is something that we've been acting with
according to our preferences.
So I have this example where like,
the robot walks into my home and my shoes are laid down
on the floor kind of in a line, right?
It took effort to do that.
So even though the robot doesn't see me doing this,
you know, actually aligning the shoes,
it should still be able to figure out
that I want the shoes aligned
because there's no way for them to have magically,
you know, be instantiated themselves in that way.
Someone must have actually taken the time to do that.
So it must be important.
So the environment actually tells, the environment is.
Leaks information.
It leaks information.
I mean, the environment is the way it is
because humans somehow manipulated it.
So you have to kind of reverse engineer the narrative
that happened to create the environment as it is
and that leaks the preference information.
Yeah, and you have to be careful, right?
Because people don't have the bandwidth to do everything.
So just because, you know, my house is messy
doesn't mean that I want it to be messy, right?
But that just, you know, I didn't put the effort into that.
I put the effort into something else.
So the robot should figure out,
well, that something else was more important,
but it doesn't mean that, you know,
the house being messy is not.
So it's a little subtle, but yeah, we really think of it.
The state itself is kind of like a choice
that people implicitly made about how they want their world.
What book or books, technical or fiction or philosophical,
when you like look back, you know, life had a big impact,
maybe it was a turning point, it was inspiring in some way.
Maybe we're talking about some silly book
that nobody in their right mind would want to read.
Or maybe it's a book that you would recommend
to others to read.
Or maybe those could be two different recommendations
of books that could be useful for people on their journey.
When I was in, it's kind of a personal story.
When I was in 12th grade,
I got my hands on a PDF copy in Romania
of Russell Norvig, AI modern approach.
I didn't know anything about AI at that point.
I was, you know, I had watched the movie,
The Matrix was my exposure.
And so I started going through this thing
and, you know, you were asking in the beginning,
what are, you know, it's math and it's algorithms,
what's interesting.
It was so captivating.
This notion that you could just have a goal
and figure out your way through
kind of a messy, complicated situation.
So what sequence of decisions you should make
to autonomously to achieve that goal.
That was so cool.
I'm, you know, I'm biased, but that's a cool book to look at.
You can convert, you know, the goal of intelligence,
the process of intelligence and mechanize it.
I had the same experience.
I was really interested in psychiatry
and trying to understand human behavior.
And then AI modern approach is like, wait,
you can just reduce it all to.
You can write math about human behavior, right?
Yeah.
So that's, and I think that stuck with me
because, you know, a lot of what I do, a lot of what we do
in my lab is write math about human behavior,
combine it with data and learning, put it all together,
give it to robots to plan with, and, you know,
hope that instead of writing rules for the robots,
writing heuristics, designing behavior,
they can actually autonomously come up with the right thing
to do around people.
That's kind of our, you know, that's our signature move.
We wrote some math and then instead of kind of hand crafting
this and that and that and the robot figuring stuff out
and isn't that cool.
And I think that is the same enthusiasm that I got from
the robot figured out how to reach that goal in that graph.
Isn't that cool?
So apologize for the romanticized questions,
but, and the silly ones,
if a doctor gave you five years to live,
sort of emphasizing the finiteness of our existence,
what would you try to accomplish?
It's like my biggest nightmare, by the way.
I really like living.
So I'm actually, I really don't like the idea of being told
that I'm going to die.
Sorry to linger on that for a second.
Do you, I mean, do you meditate or ponder on your mortality
or human, the fact that this thing ends,
§11Transcript · 1:29:38–1:38:46
it seems to be a fundamental feature.
Do you think of it as a feature or a bug too?
Is it, you said you don't like the idea of dying,
but if I were to give you a choice of living forever,
like you're not allowed to die.
Now I'll say that I want to live forever,
but I watched this show.
It's very silly.
It's called The Good Place and they reflect a lot on this.
And you know, the,
the moral of the story is that you have to make the afterlife
be a finite too.
Cause otherwise people just kind of, it's like Wally.
It's like, ah, whatever.
So, so I think the finiteness helps, but,
but yeah, it's just, you know, I don't, I don't,
I'm not a religious person.
I don't think that there's something after.
And so I think it just ends and you stop existing.
And I really like existing.
It's just, it's such a great privilege to exist that,
that yeah, it's just, I think that's the scary part.
I still think that we like existing so much because it ends.
And that's so sad.
Like it's so sad to me every time.
Like I find almost everything about this life beautiful.
Like the silliest, most mundane things are just beautiful.
And I think I'm cognizant of the fact that I find it beautiful
because it ends like it.
And it's so, I don't know.
I don't know how to feel about that.
I also feel like there's a lesson in there for robotics
and AI that is not like the finiteness of things seems
to be a fundamental nature of human existence.
I think some people sort of accuse me of just being Russian
and melancholic and romantic or something,
but that seems to be a fundamental nature of our existence
that should be incorporated in our reward functions.
But anyway, if you were speaking of reward functions,
if you only had five years, what would you try to accomplish?
This is the thing.
I'm thinking about this question and have a pretty joyous moment
because I don't know that I would change much.
I'm trying to make some contributions to how we understand
human AI interaction.
I don't think I would change that.
Maybe I'll take more trips to the Caribbean or something,
but I tried some of that already from time to time.
So, yeah, I try to do the things that bring me joy
and thinking about these things bring me joy is the Marie Kondo thing.
Don't do stuff that doesn't spark joy.
For the most part, I do things that spark joy.
Maybe I'll do less service in the department or something.
I'm not dealing with admissions anymore.
But no, I think I have amazing colleagues and amazing students
and amazing family and friends and spending time in some balance
with all of them is what I do and that's what I'm doing already.
So, I don't know that I would really change anything.
So, on the spirit of positiveness, what small act of kindness,
if one pops to mind, were you once shown that you will never forget?
When I was in high school, my friends, my classmates did some tutoring.
We were gearing up for our baccalaureate exam
and they did some tutoring on, well, some on math, some on whatever.
I was comfortable enough with some of those subjects,
but physics was something that I hadn't focused on in a while.
And so, they were all working with this one teacher
and I started working with that teacher.
Her name is Nicole Beccano.
And she was the one who kind of opened up this whole world for me
because she sort of told me that I should take the SATs
and apply to go to college abroad and do better on my English and all of that.
And when it came to, well, financially I couldn't,
my parents couldn't really afford to do all these things,
she started tutoring me on physics for free
and on top of that sitting down with me to kind of train me for SATs
and all that jazz that she had experience with.
Wow. And obviously that has taken you to be here today,
sort of one of the world experts in robotics.
It's funny those little... For no reason really.
Just out of karma.
Wanting to support someone, yeah.
Yeah. So, we talked a ton about reward functions.
Let me talk about the most ridiculous big question.
What is the meaning of life?
What's the reward function under which we humans operate?
Like what, maybe to your life, maybe broader to human life in general,
what do you think...
What gives life fulfillment, purpose, happiness, meaning?
You can't even ask that question with a straight face.
That's how ridiculous this is.
I can't, I can't.
Okay. So, you know...
You're going to try to answer it anyway, aren't you?
So, I was in a planetarium once.
Yes.
And, you know, they show you the thing and then they zoom out and zoom out
and this whole, like, you're a speck of dust kind of thing.
I think I was conceptualizing that we're kind of, you know, what are humans?
We're just on this little planet, whatever.
We don't matter much in the grand scheme of things.
And then my mind got really blown because they talked about this multiverse theory
where they kind of zoomed out and were like, this is our universe.
And then, like, there's a bazillion other ones and they just pop in and out of existence.
So, like, our whole thing that we can't even fathom how big it is was like a blimp that went in and out.
And at that point, I was like, okay, like, I'm done.
This is not, there is no meaning.
And clearly what we should be doing is try to impact whatever local thing we can impact,
our communities, leave a little bit behind there, our friends, our family, our local communities,
and just try to be there for other humans because I just, everything beyond that seems ridiculous.
I mean, are you, like, how do you make sense of these multiverses?
Like, are you inspired by the immensity of it?
Do you, I mean, is there, like, is it amazing to you or is it almost paralyzing in the mystery of it?
It's frustrating.
I'm frustrated by my inability to comprehend.
It just feels very frustrating.
It's like there's some stuff that, you know, we should time, blah, blah, blah, that we should really be understanding.
And I definitely don't understand it.
But, you know, the amazing physicists of the world have a much better understanding than me.
But it still seems epsilon in the grand scheme of things.
So, it's very frustrating.
It just, it sort of feels like our brain don't have some fundamental capacity yet, well, yet or ever.
I don't know.
Well, that's one of the dreams of artificial intelligence is to create systems that will aid,
expand our cognitive capacity in order to understand, build the theory of everything with the physics
and understand what the heck these multiverses are.
So, I think there's no better way to end it than talking about the meaning of life and the fundamental nature of the universe and the multiverses.
And the multiverse.
So, Anca, it is a huge honor.
One of my favorite conversations I've had.
I really, really appreciate your time.
Thank you for talking today.
Thank you for coming.
Come back again.
Thanks for listening to this conversation with Anca Dragan.
And thank you to our presenting sponsor, Cash App.
Please consider supporting the podcast by downloading Cash App and using code LexPodcast.
If you enjoy this podcast, subscribe on YouTube, review it with 5 stars on Apple Podcast,
support it on Patreon, or simply connect with me on Twitter at LexFriedman.
And now, let me leave you with some words from Isaac Asimov.
Your assumptions are your windows in the world.
Scrub them off every once in a while or the light won't come in.
Thank you for listening and hope to see you next time.