You are on page 1of 8

Okay, so maybe start with you sharing your name and your affiliation?

Daniela Rus. CSAIL MIT.

Excellent. Daniela, can you tell me a little bit about how your work in AI began?

I was a computer science and math major in college, and one day I attended a talk given by
John Hopcroft. And in this talk, he said that classical computer science was solved. By this he
meant that many of the algorithms that solved the original graph theoretic problems that were
introduced in the community in the 70s had solutions. And then John said it was time for the
grand applications. And robotics was one of them. And so, since I had grown up fascinated by
science fiction books and John proposed that I work on robotics with him. I thought it was
fantastic. And I decided to go to Cornell to work on robots with John.

Excellent. So, can you tell me a little bit about that early work?

[...] Yes, in fact, John said to me “Well what if we had robots that could make coffee? That could
make and fetch coffee.” And so, I started thinking, what would it take for us to create machines
that are able to do the kind of fine manipulation tasks involved in making a perfect cup of
coffee? Well I couldn’t solve the coffee challenge as a graduate student, but instead I ended up
working on a number of algorithms for fine motion planning for dexterous manipulation. And at
the time, most of the robots we had looked like big, had puma arms, they were industrial
manipulators. They were not very delicate. But human-inspired dexterous robot manipulators
were just being introduced. And it was really exciting to think about the control and planning
aspects of these new types of machines.

Can you tell me a little bit about your present work?

So, at the moment, I have this great dream of making pervasive robots. Creating a world where
machines can help us with cognitive and physical work, just like computers help us with
computational work. I mean if you think about it, twenty years ago computation was a task
reserved for the expert few because computers were large and expensive, and you really
needed expertise in order to know what to do with them. But all of that changed with the
introduction of the smart phone and today everyone computes. And in fact, we depend on
computations so much we actually don’t even notice it around us. And this raises a very
interesting question, in this world so changed by computation, what might it look like with
machines helping us with physical and cognitive work? And towards this end, I’m interested in
developing the science and engineering of autonomy and the science and engineering of
intelligence. And here we have to understand that a machine is made out of a body and a brain.
For any task we want that machine to do, the machine has to have the body capable of doing
the task and the brain capable of controlling the body to do the task. And then, the machine has
to be easy to interact with. And so, to me, advancing the science and engineering of autonomy
means making progress on robot bodies or machine bodies. Making progress on machine
brains. And making progress on the interaction between machines and people. And let me say,
just a little bit about each one of them. With respect to machine bodies, most of today’s robots
are inspired by the human form. They are either industrial manipulators or humanoids or they
are machines on wheels. And these machines have been around for nearly sixty years. And
these machines do extraordinary feats on assembly lines in factories. Yet these machines
remain isolated from people because they are big and heavy and difficult to work with. And this
raises an interesting question. What should the machines of the future that are safe to be
around, that are adaptive, that can exist in human-centered environments, what should they
look like? And so, I think it is a good time for us to challenge what our idea of a robot and a
machine is. And to advance new types of approaches to making robot bodies to look at new
materials. We can make robots out of hard metals and plastics, but also out of silicon and other
soft materials. Out of paper, out of food. We have a wide range of possibilities and also going
back to the idea that the shape influences function, what should all the shapes be? Can we
imagine creating machines that are inspired by all-natural shapes? And even by shapes we see
in the built environments. We can in fact imagine breathing life into many of the objects that
surround us to help us with monitoring and with small activity. With small physical tasks. So, I
would like to see the next sixty years of robotics focused on creating machines out of a wider
body of materials, in a wider body of shapes. And I really imagine a Cambrian explosion in the
introduction of machines in our lives. And altogether this will take us to a world with machines
supporting us with physical work.

I’m really struck by this notion of delicacy and diversity as you’re speaking, and I’m wondering if
there are specific niche areas where you imagine the systems that you’re working on presently
assisting people? Are you thinking about things like elderly care in particular or other ways in
which we can either support on tasks that are particularly intellectually or physically difficult? Or
are there other, you know, ways in which you imagine tending to this problem? Are they case
based or are you really working on more the algorithmic and the, the embodiment questions
broadly conceived?

Well look I believe that AI and robotic technologies can have so many applications. Can
contribute to so many aspects of our lives. And here I would like to make a distinction between
three different but interrelated fields. One is robotics. And robotics puts computation in motion.
Gives machines the ability to move in the world. The other one is artificial intelligence. And with
artificial intelligence, we give machines the ability to reason. We give machines human-like
characteristics. Maybe to see, to hear, to reason, to play games like humans. And machine
learning cuts across robotics and artificial intelligence and today with machine learning we are
able to give machines the ability to use data to make predictions and to find patterns that the
human eye cannot see. Together these three technologies have huge potential to empower us.
With these technologies, we can ensure that there will be no road accidents. We will be able to
transport people and goods much faster. We will be able to better diagnose, monitor and treat
disease. We will be able to provide people with universal communication, in other words, we will
be able to provide people to, with the ability to communicate instantly with each other, no matter
what language they speak. We will be able to provide universal democratized education. We will
keep people’s information private and safe. Altogether I imagine a world where machines can
take care of the low-level routine tasks while giving people the ability to focus on high-level
strategic, critical thinking.

Okay. You mentioned that when you were younger, you were actually influenced by science
fiction. Can you speak a little bit towards that? I mean you clearly positioned yourself within
computation in a very sophisticated way, but I’m wondering if you can speak a little bit to those
literary tropes that influenced your thinking or inspired you a bit?

So, I used to love the books of Jules Verne [not sure of name @ 10:02, 1st video] And I also
loved the movies of Jacques Cousteau. Okay, but I actually didn’t think, didn’t even dream
about making my own machine until years later. And years later, now when I became a
professor, I met a student who also loved the books of Jules Verne (check name) and the
movies of Jacques Cousteau. And then we decided to make a robot together. Our version of the
undersea exploratory. And we called this robot AMOUR. And AMOUR stood for Autonomous
Modular Optical Underwater Robot. But in fact, making this robot was really about love. This
robot had a lot of very innovative technologies embedded in it. It was the first helicopter-like,
hovering robot. It was a small robot that you could pack in a suitcase and take it with you on
field trips. This robot could take beautiful pictures, color accurate pictures which were enabled
by a very innovative approach to imaging in water. And with this robot we could also transmit
HDTV quality video. And we were really excited about this robot. So, since you asked, let me tell
you the rest of the story.

Yeah.

Okay, so it took us about three years to make this robot. [...] So, one very hot summer day, we
were in Singapore and we had just finished a research trial for our sponsor. And then we
decided to really see how far and how strong this robot was. And in fact, there was a team from
Singapore who was on the same boat with us, and we thought we might play Hunt for Red
October with our two robots. Well, we decided to take down all the safety features of our robot
to give it more juice and we giggled when the Singapore team attached their robot to a fishing
line. So, we sent off these two robots and then we waited and waited and waited, but our robot
never came back. We lost AMOUR and we were devastated. And then, we realized that in
science usually great failures teach you great lessons. And also, is the case that in science, the
first copy of an invention takes much longer.

Yes.

Than subsequent copies. So, this was in May. It was after ICRA (International Robotics
conference) and we went back to the lab, we grounded ourselves for the summer, and by the
end of the summer we had three improved copies of this robot and interestingly as we watched
this robot do the flips and move in water, we felt like we wanted more. We wanted an even
better, even more agile robot. And that was actually a bridge from our hard-bodied AMOUR
robot to the soft, underwater robots we have built the soft robot fish. And what’s really
interesting is to remember that initially it took three years to build AMOUR. Then, it took us three
months to build three improved copies of AMOUR. And today in the lab, we can build the soft
robot fish in a matter of a few days.

Wow.

And this kind of, this rapid change of technology, this rapid pace of technology is very similar to
what we have seen with computation. And this is why I’m so excited about a future with
machines for human-centered environments. Machines that are much more capable. Machines
that are delicate and soft and that can help support people. Now with these machines, we can
go to places where humans cannot go. And we can peak on the secret lives of animals. Of
natural systems. But with these machines we can also go into human environments, into your
apartment and I like to think of a time when such robots can help us with cooking. They already
help us with cleaning, but maybe they help us with cleaning in more substantive ways. In other
words, it’s back to this idea of empowering these machines to take care of the routine tasks that
take time and we do not like to do. With these machines, we can also improve healthcare. So,
for instance, at MIT we are working on an autonomous wheelchair. And we are very excited
about the possibility of using this wheelchair for rehabilitation. In our conversations with
hospitals, we learned that a large fraction of the time a rehabilitation doctor has with the patient
is spent moving the patient in the wheelchair. First to bring to the gym and then to send back to
the hospital room. Now imagine if we had an autonomous wheelchair and all of that time that is
spent in transition can now be spent by the doctor applying expertise. And by the patient
improving his or her well-being. So, I think there’s so many opportunities. It’s super exciting.
Yeah. I’m really struck by the ways in which your vision focuses on a social function with,
between humans and, and machines. The example of the wheelchair is one. I’m wondering if
you can speak a little bit on examples that you think are quite useful or precise ways in which
technologists who are developing robotic systems and AI systems communicate those systems
and their capabilities or their limitations to publics who may or may not be as conversant in the
intricacies of computation that’s needed to achieve these functions in these machines.

I think it is very important to communicate the possibilities of technology to the general public.
Technology can truly be a vector for positive good. And it could truly help everyone. But we
have to get those stories out and right now, we as technologists spend our time developing the
technology and I think it’s important for us to begin to get out and start explaining to people how
technology can help them. There are so many ways in which people can benefit in every
profession where there are low-level tasks where there are data tasks, technology can help. If
the professions have low-level, simple physical tasks, technology can help. However,
technology is not yet ready to help with expertise, with deep communications, nor with complex
tasks in unpredictable environments. It is much easier to send a robot to Mars today than it is to
get that robot to clear your tabletop.

Yes.

But we are working on improving those capabilities. So, for instance, with the use of technology,
we can imagine teachers getting help grading papers and therefore spending more time at
interacting with the students and providing more personalized education. We can imagine our
cities providing better support and security and safety for people. I like to imagine a future of
autonomous transportation which would, which will ensure that there will be no road fatalities.
And furthermore, we can ensure that our parents and grandparents have a higher quality of life
in their retirement. And all of us will have the ability to go anywhere, anytime. I don’t believe it’s
a matter of if, I believe it’s a matter of when. And, I think we will all benefit. Transportation is
indeed on its way to become like a utility, available anywhere, anytime.

I’m really struck by this idea of the swapping of tasks and how we use our time. This notion of
machines integrated into our everyday work as well as our domestic setting to help take over
some of these mundane tasks that, as you said, not all people like doing. There’s a way in
which your advocating for a use of that, that time for example, the teacher using the time that
used to be spent grading and offering assessment and evaluation of work to continued what
sounds to be like more personal instruction, right? A deeper engaging of human-to-human
relationships by virtue of removing some of the mundane tasks by virtue of the machine’s
integration into that relationship. You teach and you speak publicly and you’re developing these
technologies yourself, so what responsibility do you take on in, in communicating to the publics?
Or how do you envision your responsibility as a communicator on these systems?

I am very excited about communicating with the public about the benefits of technology. I really
see technology as a vector for the greater good. And I really see technology as being able to
support everyone. And I’m concerned when I read in the popular press stories about negative
effects of technology and dramatic negative impacts. Many stories we read about today instill
fear rather than dreaming on people. And I would love to do my share to contribute to how
people dream about how their lives can be made better by technology. There are things we can
do today but there is a lot of hopeful, wishful thinking for the future, however, I believe that a
future with machines supporting us more extensively with cognitive and physical work is not that
far off. You see, usually when I tell people what I do, I get two types of reaction. So, some
people get nervous and start cracking jokes about Skynet and ask me when they’re going to
lose their job. And other people get really excited and say, “When is my car going to be self-
driving?” Because it’s not this year.

News flash.

So, I think that it is important to understand the fears of that first group of people. And provide
alternatives for how to see things differently. And this starts with understanding that everything
we do, AI, robotics, machine learning, is about creating new tools. So, all these tools are just
tools by the people for the people. And like any other tools, they are not inherently good or bad.
They are what we choose to do with them. And I believe we can choose to do incredible things.

Good. So, I’m wondering if you can speak a little bit to the way that you see AI systems
changing the way that people are working up until the present moment? Up until now.

We have had so much impact from technologies. If you think about it today, doctors connect
with patients, and students connect with teachers that are thousands of miles away. We have
robots that help packing in factories. We have 3D printing that creates customized goods. We
have network centers, sensors that monitor facilities. We are really surrounded by a world of
opportunities. And these opportunities are only going to grow bigger when we consider the
impact of the new wave of technologies. And I really like to think of a world where routine tasks
are taken off your plate. I like to imagine robots delivering fresh produce at my doorsteps on a
daily basis. I like to think about the possibility of garbage cans that take themselves out and
automated infrastructure that supports removing them. I like to think about intelligent assistance
that may be embodied or not, that support us with all sorts of tasks to ensure that we work
efficiently, and we live well.

So, we are in this boundary space presently where to your point, the autonomous vehicles will
not be driving us around in the next year. And we have, you know, vacuuming systems that are
autonomous but are not finely tuned to work as well human workers presently within these
domestic or service economies. I’m wondering in regards to your vision in the coming twenty
years and these visions of optimism for the future, how you expect the tensions associated with
human dignity associated with work in, becoming complicated in this boundary space as we’re
moving towards the prospect of trash cans that walk themselves to the curb. I think I would like
that too. But the complexity then as well of municipalities hiring individuals who may not have a
higher-level education degree to actually collect that trash as part of this sanitary services
associated with municipality.

There are many jobs that exist today that people don’t want. There are so many job openings,
and there are openings in places like doing recycling. Very important for the planet. Handling
cold foods. Even fulfillment centers and shipping centers have trouble hiring enough people.
And there are a lot of openings. Today, there is a mismatch between where there are job
openings, what people’s skills are, and what people want to do. And I believe that there are jobs
that should be done by machines because they’re not attractive to people. And then we have to
think about how do we match people’s training with the machines. But you see today, and I
believe this will continue to be in the future, it’s not about people or machines doing a job. It’s
about people and machines supporting each other.

Back to this notion of the tool.


The machine is a tool. It’s an intelligent tool. And the machine can do some things much better
than people. Machines have chips and speed and strengths. But people have hearts and
wisdom and delicacy. And so, if you think about a world where it’s people and machines working
together with machines doing what they’re best at and people doing what they’re best at, that
will be a better world. Now, with respect to jobs, it is true that the jobs are changing. And there is
no going around this. And this is bringing a certain level of uncertainty and fear to people. And
this fear is meaningful. But on the other hand, there is a lot of reason for optimism. So, for
instance, the world economic forum released a report that talked about the possibility of some
seventy-five million jobs going away in the next few years. But a hundred thirty-three million jobs
coming. And I think we as people have, we are much better at thinking about what might go
away then imagining what could come. And if you think about where we were in, in this world
ten years ago. So, ten years ago is a short amount of time. I remember that time very clearly.
Well ten years ago we had no smart phones. We had no Cloud. We had no social networking.
And so, let us think about all the different jobs that were enabled because of the advancements
in these technologies. And these are jobs at all levels. And so, I think when you consider a
profession, it is better to think about which tasks within that profession will be automated rather
than which professions will go away. Because in every profession we spend time doing different
tasks. We spend time applying expertise. We spend time managing others. We spend time
doing stakeholder interactions and communicating. We spend time doing data collection and
data analysis. We spend time doing predictable physical work and unpredictable physical work.
And the technologies we have today are good at automating the predictable physical work and
the data tasks. If those tasks were to receive support from machines, we would have much
more time to devote to the higher-level tasks.

So, in this relationship between the human user and a system or tool, I’m wondering if there’s
an example in your own research or that you’ve identified in other’s work where power for
decision making perhaps has been transferred from the human user to that system? And I’m
wondering how you see that configuration affecting human power relationships.

So, self-driving cars is an excellent example. So, in autonomous driving, the idea is that we yield
our, the control power for the vehicle to the machine. And this is really kind of exciting to me.
But it’s better to think about this transfer of power as having a friend that is supporting you in the
car. Maybe watching your back on a treacherous piece of the highway or maybe in the future,
the cars will even keep track of our personal life and know that maybe we forgot to call our
mother because it was her birthday. And by driving, by taking over control of the vehicle, the car
would make the placement of that call very natural and easy. So, of course, level five
autonomous driving is some years away. There are, however, intermediate technologies that we
can imagine introducing in the meantime. In our work at MIT, we are exploring something we
call parallel autonomy or a guardian approach to driving. Where in fact a human remains in
control of the car, but a parallel autonomy system watches over the human’s shoulder and has a
great sense of everything that is happening in the world around the car with much greater
precision and much wider scope of view than what we’re able to do with a naked eye. And so,
such a system could intervene much like anti-lock brakes intervene today. But such a system
could intervene in logical ways to prevent road accidents. And, so while we are developing the
technologies that take us towards level five autonomy, we can use such parallel autonomy
systems with a spirit of do no harm. Which is the same spirit that we have for instance in
medicine.

This trope of the guardianship. It sounds to me that this is a philosophy that empowers the
individual human by virtue of enhancements that are afforded through the tool.
Exactly.

Is that a fair interpretation do you think?

Exactly. Exactly.

Are there other examples of tools that you think will empower an individual human or groups by
virtue of this motif or this guiding principle of do no harm?

Many examples. The one I really like is a recent study from medicine. So, in the study, doctors
and machines were asked to look at scans of lymph node cells and classify them as cancer or
not cancer. Human doctors made 3.5% error, as compared to the AI systems which made 7.5%
error. But when the doctors and the machine worked together, the error dropped down to 0.5%
This is 80% improvement in the performance of the system.

And are you able to disambiguate whether or not the system and the doctors were making
different errors?

Actually, I don’t know the answer to that question because it’s not my study.

Yeah, okay.

But what is exciting is to see how humans and machines working in tandem.

Sure.

Can complement each other. And today, these systems are deployed in the most advanced
hospital centers in the world. But imagine a world where all doctors have access to these
systems. Imagine a world where an over worked doctor or a rural doctor who do not have time
to keep up with the latest and greatest clinical trials has access to such systems that can
provide them on demand, just in time, with the most relevant information for the patient they are
working with. With these systems, the patients will benefit from the world’s greatest discoveries,
especially the ones that are most suitable to their condition.

So, in your own work or when you’re asked to offer your expertise as a practitioner in this area,
how do you speak to the potential pitfalls of machine autonomy. Or are there particular
examples that you gravitate to illustrate this concept of do no harm, so that individuals are
instructed on what the parameters might be for harm. How you characterize that.

So, here’s how I think about this. The deployment of any kind of autonomous system, whether it
a robot or a software system, that deployment requires a certain checks and balances to ensure
consumer confidence in it. And right now, we don’t really know how to approach this question.
But one idea is to consider certain attributes that are important and critical to assess the
performance of the system. So, what might those attributes be? Well we want the decision of
the system to be interpretable and explainable. We want the system to be fair. We want, if the
system is rooted in data, we want to be able to track the data provenance to ensure that the
data is good, and we want to be able to assess in some sense the bias in the data because with
today’s techniques, if you train something with a body of data that has bias, the system will
operate with the same bias.

Or scale it.
So, there are certain attributes that are general and maybe we can create a driving test for each
of these attributes for the system’s we want to deploy. But now here’s the thing, in each industry
that might adopt the use of AI, machine learning and robotics, we would have to instantiate
those attributes to the industry. So, for instance, for interpretability and explainability in finance
you might have to be able to provide an explanation for why the loan was not granted. In
medicine, you might have to provide an explanation for why the diagnosis. In transportation, if
there is an accident, what was the sequence of evaluations that led to that particular behavior
from the vehicle? So, there are a lot of super important and super exciting ideas to consider and
what I firmly believe is that this requires conversation between multiple stakeholders. Between
the technologists who are creating the technology and understanding its scope and limitations,
industry leaders and business leaders who want to use the technology, and policy makers who
know how to create the social norms that enable the safe use of technology.

Excellent. So, I think we’ll close with one last question that we’ve asked everyone. And that’s
your thoughts on the development of general artificial intelligence?

A machine that has the same level of intelligence as a human.

Perhaps. Everyone interprets this a little bit differently, but.

Well, today, we are making a lot of progress on what we call narrow AI. In other words, we have
point solutions to problems. We are celebrating the feat of AlphaGo who beat the Grand Go
Master which it’s truly a technological achievement. But I will say that the AlphaGo program will
not be able to play poker or even GoFish with you. Whereas, if I explain the rules to you, you
would just pick them up in no time. We are, in my opinion, very far from general artificial
intelligence. In order to understand the path towards more capability in our machines, we really
have to develop the science and engineering of intelligence. And today we don’t really know
how the brain works. But, by bringing computer scientists, artificial intelligence researchers,
neuroscience and brain and cognitive science researchers together, we will have an opportunity
to advance our understanding of the brain which means our understanding of life and of each
other. And perhaps that understanding could translate into new types of algorithms that will
render our machines more capable. There are significant limitations in the solutions we have for
our machines today. We have systems that require millions and millions of manually labeled
examples in order to get trained and to work. But we don’t need millions of examples in order to
pick up a new concept, to pick up abstraction in order to connect our reasoning about the world
to symbols and abstraction. It will take a long time to get there.

You might also like