杏吧原创

The hardest thing about robots? Teaching them to cope with us

Humans are spectacularly weird and unpredictable. Computer scientist Anca Dragan is working on helpful robots that understand how we get things wrong

As a child, what did you want to do when you grew up?

I got interested in computer science at an early age, but before then, it mostly depended on which movie last inspired me.

Explain what you do in one easy paragraph.

I get robots to account for humans when deciding on their actions. On the one hand, this means giving them the ability to make predictions about what we will do next and make sure they can coordinate with us, like when autonomous cars negotiate for their turn during a merge in traffic. On the other hand, this means giving them the ability to infer what it is that we want them to do. How do I want my car to drive? What help do I need from my personal home robot?

What do you love most about what you do?

I love that I get to focus on problems that I believe will be relevant to us in the long term, like how to define what we want machines to do or how to get machines to better understand us.

What鈥檚 the most exciting thing you鈥檙e working on right now?

We鈥檙e working on making machines more adaptable to the wrong model of human behaviour. For instance, we can get robots to figure out where you鈥檙e heading by assuming your actions are relatively rational. But what if you鈥檙e avoiding something the robot doesn鈥檛 know about? Or what if the robot assumes you鈥檙e thinking many more steps ahead than you actually are? Or maybe the robot assumes you鈥檙e trying to teach it something, but you鈥檙e really just doing your own thing. Can robots be more robust to such misspecification of human behaviour?

Were you good at science at school?

I was OK at physics and pretty terrible at the rest. Let me put it this way: I don鈥檛 think you would have looked at me in school and anticipated that I鈥檇 become a professor at one of the top universities worldwide. In the end, I clicked very well with research, I took as much initiative as I could and I got very lucky.

If you could send a message back to yourself as a kid, what would you say?

I definitely wouldn鈥檛 say, 鈥淒on鈥檛 worry, everything will be great鈥, because I think the worrying is what made me ambitious. Growing up in Romania, my dad told me that if I didn鈥檛 do exceptionally well in school, I鈥檇 have to go work on the farm 鈥 and I believed him!

If you could have a long conversation with any scientist, living or dead, who would it be?

I鈥檇 really love to talk to Alan Turing.

How useful will your skills be after the apocalypse?

Not at all useful. Unless our problem is that AI systems don鈥檛 know what we want, I鈥檝e got nothing.

Do you have an unexpected hobby, and if so, please will you tell us about it?

鈥淢y dad told me that if I didn鈥檛 do exceptionally well in school, I鈥檇 have to work on the farm 鈥 and I believed him!鈥

I love sailing because I get to take advantage of living by the water, despite the fact that the San Francisco Bay is so cold. I really need to focus to make sure I don鈥檛 do anything bad to the boat, so it鈥檚 almost like a way of meditating.

OK, one last thing: tell us something that will blow our minds鈥

Imagine you are learning to do a task and a robot is trying to help you. You would think that the better you are at the task, the better the outcome for you and the robot. It turns out that isn鈥檛 true: success relies a lot on whether what you do and what the robot does gel together. This is one aspect that makes assisting people hard.

Topics: Artificial intelligence / Robots