Ken Goldberg
speaker
221 appearances
3 recordings
3 series
first heard Jan 2025
last heard 7 May
Ken Goldberg’s voice in public audio — every appearance, attributed to the second.
Trend
recordings per month · last 12 monthsRecordings per month over the last 12 months — 1 in all, peaking in May 2026 with 1.
Appearances
Multi-celled organism. You're right. The Roomba is the most successful robot of all time. So when they count robots out there, they count these Roombas where there's like 10 million of those. But that's the robot, right? And it's very simple. It's basically just random motion. And over time, it does cover your carpet. And it's pretty reliable.
But of course, it also has this problem that it gets stuck. all the time and tangled up in stuff. And so it's not ideal and it can't go upstairs. Also, a lot of people bought them as a novelty. There's a lot of them sitting in the closet somewhere.
I have two drones.
No, I need like six hours to basically figure it out. And it's sitting in the box.
Right. So everybody's impressed. The analogy, if you say, okay, you can beat the best person in the world at chess, then that means you have a very powerful machine, artificial intelligence. Now it can beat the best person at Go. And nobody can play Go or chess that well. So you think this is smarter than everybody.
That's the way people reason. But then it can't drive a car. It's a whole bunch of things it can't do. And anything physical is just picking up or opening this can that I just did that is impossible for robots.
Splatters, broken glass. Yeah. But I agree, taking things in and out of the dishwasher would be great. Just unloading and loading the dishwasher, right? And some would say that their dishwasher is a robot. It is, it's very successful. See, there's this idea that if you can use humans and robot together, that's very powerful. So that's what I call complementarity.
When if you figure out that you have a machine, you can use it, but you have the human do the parts that we're good at and then let it do the parts it's good at. Together, you have a great system. And a dishwasher is a beautiful example. And the washing machine, they do all this, but we have to load it and unload it.
In a laundry aspect, it's also that you want your clothes to be folded at this precise time, right when they come out, because then they're at the perfect stage.
No wrinkles. And if you do it too soon, they're kind of soggy. If they're too late, they get all wrinkled. So having a machine to do that would be quite good. And there's some really interesting new results that just came out about this. But we've been working on it too. And one of the ideas is you fling the clothes up and you use air to help smooth them out.
Like humans do that all the time, right? You snap, you know. Yeah. That has only been really done in robotics in the last five years.
So I'm super optimistic. I love working on this topic. And I feel like we have a lot more work to do. So that's also encouraging. I don't worry that it fails. I actually love the times when it does succeed. That's super rewarding, knowing how hard it is. You're like a fan of hockey instead of basketball.
But when I get it, boy, it's... Oh, that's interesting. I never thought of it that way. Yeah, because... In grasping, we've actually made some good progress just in picking up objects. And that was the breakthrough. So coming back to this timeline, so in 2012, there was this breakthrough in vision.
And suddenly deep learning, this new way of building these very large networks, we're using lots of data and using GPUs, graphical processing units. It's basically a new kind of computer. It has this breakthrough where suddenly machines are able to recognize images and things in images. Like it'll say that's a book and that's a cup and that's a microphone. That's part of Fei-Fei Li's work.
Exactly. So Fei-Fei Li is actually at the center of this. She builds this data. So she plays a pivotal role. When all that gets put together, suddenly it's a revolution. And that's a big moment in robotics and history. We apply it to robotics. And so her system was called ImageNet. And so the system that we designed for grasping, we called DexNet. As an homage to her. Oh, that's great.
So DexNet was our system. We worked on it for five years, and we basically applied deep learning techniques to be able to figure out where to grasp objects. And it started working better than anything had been done before. And I was so surprised because I had been trying to work on this problem, and then I suddenly was able to pick up almost everything we could put in front of it. Oh, wow.
Yeah, no, that's a really great point. There is a critical point when you get enough data and suddenly it starts working. It took a lot. It was 80 million plus images that Fei-Fei put together. Right. And in our case, we had 7 million grasp examples that we had found. And then it started to work and it was like, oh, this is so exciting.
And think of it with a very simple gripper, just a parallel pincer. So you would put a bin of objects in front of it and it would start to pick them one by one and put them out. And so we would test it by going into the basement of the garage. We'd just throw all kinds of stuff in there. And it would just pick them up consistently and clear the bin. And we would try and fool it.
He must have been elated. It was so much fun. There's a story where we got invited to show this to Jeff Bezos. And he invited us down to this event in Palm Springs. He said, bring the robot. I want to see this. We had never left a lab before, so it was a big deal to put it on a truck. And we weren't sure it was going to work. We had like 300 objects that we brought with us, got it all set up.
He came in the booth and it was working. And we were so relieved. And he was trying it with different things and it was just like it was in the lab. And everything was going great. And then his assistant was standing there and took off his shoe. And he said, well, can I try my shoe? And I remember my mouth goes dry. Because of all the things we've tried it with, we've never tried a shoe.
Showing 121–140 of 221 · page 7 of 12
← Previous
Next →