Marc Raibert

speaker

243 appearances 1 recordings 1 series first heard Feb 2024 last heard Feb 2024

Marc Raibert’s voice in public audio — every appearance, attributed to the second.

Trend

recordings per month · last 12 months

No recordings in the last 12 months.Older appearances are listed below; set an alert to hear about the next one.

Appearances

newest first · ▶ plays the moment

Lex Fridman Podcast · #412 – Marc Raibert: Boston Dynamics and the Future of Robotics · 16 Feb 2024 podcast

Well, it's not obvious if you just look at the raw data. what the sequence of acts are. It's really a creative, intelligent act for you to break that down into the pieces and understand them in a way so you could say, okay, what skill do I need to accomplish each of those things?

So we're working on the front end of that kind of a problem where we observe and translate the, if it may be video, it may be live, into

a description of what we think is going on and then try and map that into skills to accomplish that and we've been developing skills as well so you know we have kind of multiple stabs at the pieces of of doing that and this is usually video of humans manipulating objects with their hands kind of thing we're starting out with bicycle repair some simple bicycle repair oh no that seems complicated that seems really complicated it is but but there's some parts of it that aren't

Like putting the seat in, you know, into the, you know, you have a tube that goes inside of another tube and there's a latch. You know, that should be within range.

I think it is. And I think that's the kind of thing that people don't recognize. Let me translate it to navigation. Mm-hmm. I think the basic paradigm for navigating a space is to get some kind of sensor that tells you where an obstacle is and what's open, build a map, and then go through the space.

But if we were doing on-the-job training where I was giving you a task, I wouldn't have to say anything about the room. We came in here, all we did is adjust the chair, but we didn't say anything about the room, and we could navigate it. So I think there's opportunities to build that kind of navigation skill into robots. And we're hoping to be able to do that.

Yeah, and lack of specification.

I mean, that's what sort of intelligence is, right? Kind of dealing with, understanding a situation even though it wasn't explained.

You know, Since ChatGBT, which is a year ago, basically, there's a huge interest in that and a huge optimism about it. And I think that there's a lot of things that machine learning, that kind of machine learning. Now, of course, there's lots of different kinds of machine learning. I think there's a lot of interest and optimism about it.

I think the facts on the ground are that doing physical things with physical robots is a little bit different than language. The tokens sort of don't exist. Pixel values aren't like words. But I think that there's a lot that can be done there. We have... We have several people working on machine learning approaches.

I don't know if you know, but we opened an office in Zurich recently, and Marco Hutter, who's one of the real leaders in reinforcement learning for robots, is the director of that office. He's still half-time at ETH now.

the university there where he has an unbelievably fantastic lab and then he's half time leading will be leading off efforts in the Zurich office so we have a healthy learning component but there's part of me that still says if you look out in the world at what the most impressive performances are

They're still pretty much, I hate to use the word traditional, but that's what everybody's calling it, traditional controls, like model predictive control. The Atlas performances that you've seen are mostly model predictive control. They've started to do some learning stuff that's really incredible. I don't know if it's all been shown yet, but you'll see it over time.

And then Marco has done some great stuff and others.

I think we're going to find a mating of the two and we'll have the best of both worlds. And we're working on that at the Institute too.

Sure, technical fearlessness means being willing to take on a problem that you don't know how to solve. Study it, figure out an entry point, maybe a simplified version or a simplified solution or something. Learn from the stepping stone and go back and eventually make a solution that meets your goals. And I think that's really important.

Yeah, and you don't know how to do it. There's easier stuff to do in life. I mean, I don't know. Watch, understand, do. It's a mountain of a challenge.

Yeah. I mean, we have others like that. We have one called Inspect, Diagnose, Fix. You call up the Maytag repairman. Okay, he's the one who you don't have to call, but you call up the dishwasher repair person, and they come to your house, and they look at your machine,

It's already been actually figured out that something doesn't work, but they have to kind of examine it and figure out what's wrong and then fix it. And I think robots should be able to do that. Boston Dynamics already has spot robots collecting data on machines, things like thermal data, reading the gauges, listening to them, getting sounds.

And that data are used to determine whether they're healthy or not. But the interpretation isn't done by the robots yet. And certainly the fixing, the diagnosing and the fixing isn't done yet. But I think it could be. And that's bringing the AI and combining it with the physical skills to do it.

Showing 141–160 of 243 · page 8 of 13