If Robots Want to Work with Us, We Must Fix Four Problems

It seems weird, I know,

but in the very near future,

you might have robotic arm handing you beer,

and that means we need to start thinking about

how exactly we're going to be interacting

with these kinds of machines.

The good news is, some very smart people

are working on this very problem,

like UC Berkeley roboticist, Anca Dragan.

So may we present to you Anca's

four biggest problems in human robot interaction.

Problem number one is enabling robots

to anticipate human behavior.

When we work together or interact as humans,

we have a pretty easy time anticipating

what's gonna happen next,

and we'd like for robots to do the same.

So say I'm working with this arm from Universal Robots.

I want it to hand me this screwdriver.

It can go about that one of two ways,

either pointy bit first or the other way.

That would actually be more natural

for me to go into my next task.

Robots can sort of plan and say,

Well, I should give it this way

because then the natural way for the person

to give it also works out.

So this is a beautiful example

of how robots can gently guide us to perform better

in the tasks that we need to perform.

Problem number two is transparency,

but interaction is a two way street,

so people will need to anticipate

the robot's behavior, as well.

If I see you carrying something,

I can figure out about how heavy it is, right?

If this were much heavier, I'd hold it in a different way.

So there's a lot of inferences

that we people make when we watch other people,

and we'd like it to be the case that people

can make the same inferences about robots.

The thing about robots is,

they don't know their own strength,

so what's clever about this robot

is it'll actually stop if it makes contact with a human.

A robot lifting a container full of water

and a container full of lead is one and the same,

so if they're going to be handing objects over to humans,

they need to be able to communicate the weight of things,

like they might have to slow down or pretend to struggle.

We've found that even something as subtle as the timing

of the robot's motion has such a big influence

in terms of what people will end up reading into the motion.

Things like hesitation, non-capability and naturalness,

and you know, the weight thing into timing.

Problem number three I'd say is customization.

Imagine you get an autonomous car.

Not every single person will want that autonomous car

to drive in the same way, right?

You ideally would want all autonomous cars to be safe,

but the trade off that they're making, right,

between how efficient they are, how fast they're driving,

how much they're breaking the rules of traffic,

how close they're willing to get to other cars,

you want these trade offs to be

kind of determined by the person.

It's really up to them.

[Matt] What's interesting here is that the autonomous cars

of the future won't be one vanilla personality.

They'll be customizable.

We've been working on ways for the robot

to elicit your preferences without actually asking you

for demonstrations of how it should behave,

so while we're exploring for instance

these comparison based queries

where the robot sits you down, maybe in a simulator,

and shows you kind of a driving scenario and says,

Hey, I could, in this situation,

I could do this or I could do this other thing.

Which one would you want?

[Matt] The issue then becomes striking a balance

between aggressive and defensive.

I mean, you have to give the people the cars they want,

but the whole point here is safety,

so human robot interaction will be pivotal

going forward in the robo-car revolution.

[Anca] Problem number four is really about

making robots better at what they do.

So say you tell a Rumba that it'll get extra

brownie points if it picks up as much dust as possible.

The optimal policy ends up being

is for the robot to collect some dust,

then not just go on and collect more,

but to just eject its current dust,

and then suck it back in again

and do this on a loop, right?

It collects a lot of brownie points that way.

Robots tend to take things literally.

This arm, for instance, follows a strict set of commands,

but as we interact more and more with robots,

perhaps they shouldn't be so rigid in their objectives,

especially if there are humans along to guide them.

We found that by making robots

a little bit more humble in a sense,

a little bit more uncertain about

what they should really be doing

and looking to the person for guidance about that,

then all of a sudden, human oversight

and human intervention is incredibly valuable.

It's something that you derive a lot of information from.

So there you have it, four fascinating problems

in human robot interactions

that roboticists will no doubt conquer,

and I will cheers to that.