Neural nets handle soft packages that coded robots struggled with

Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

1. 2:03:43

   It'd be weird. Do you think that they will bring... Do you think manufacturing will come back to the U.S.? Yeah, we're going to bring back. Because of this? My view is that we don't want to bring back manufacturing that's already overseas. We don't want to make shoes, make toys, things like that. I don't think we have the will to do this. I don't think we have the know-how to do this, as well as some of the Asian manufacturing groups.

2. 2:04:10

   When I'm overseas, so I've walked a lot of the high-volume consumer electronics lines and stuff overseas. Some of the most impressive things I've ever seen in my life. it's like you walk these lines and they're just shipping electronics like crazy and every line they have this box of automation inside of it called a little tiny robot inside of there that's moving some, whatever, a phone enclosure or something like that.

3. 2:04:33

   And it's doing it through an automated way and moving it around a little conveyor and it's moving to the next station, maybe a human's doing something and it's going down the line, it's going to the next station that's got a robotic system in there, completely customized and different from what you just saw and they have lines and lines in floors and floors of this and then buildings and buildings.

4. 2:04:49

   And you're like, holy shit, each one of those boxes is like a figure style complexity. And they have like hundreds of them. And they need to run them at high rate. It's just like, it's unbelievable, actually. It's not trivial, it's very complex. And they've been doing it for several decades on these lines. So I think one is like, I don't think that stuff we want to move back. I think we want to move back the high-end robotic stuff that's going to be like super transformative for us in the future.

5. 2:05:22

   All the futuristic tech. Yeah, we want to bring back flying cars. I want to bring back like humanoid robots, like the stuff that's like highly dynamic, very intelligent systems, like the next generation, like manufacturing 2.0 stuff. Gotcha. So we're doing that right now in California on our campus. We have a fairly large campus in the Bay Area and we manufacture right now, like, whenever 90 minutes or so.

6. 2:05:45

   And we'll continue to spin that up, and then we'll put, you know, we'll talk more about it, but we'll put more investment here into U.S. manufacturing for the future. Right on. So we're going to design humanoids here. So these are all manufactured? We manufacture those in California. Right on, man. Yeah, man. They walk off the lines, they walk over. It's like, 90 days ago, We're like making a little bit, but now we make like, there's like seven robots that are all doing like end of line checkout by themselves for like an hour and a half.

7. 2:06:17

   They do their own burn-ins, all OEOL checks. So they're self-looking at each other, self-calibrating. They're doing like burpees. other shit to make sure they like they're okay. If they fail, they go into a triage place. We understand why to fail, like that shouldn't happen. We should always fix that and it should not fail again. Like how do we fix the manufacturing process so the next one doesn't come out and ever have that failure.

8. 2:06:37

   And now we've gotten that process really dialed, I mean, dialed in. We still have issues, but like it's fairly dialed in. And so the robots come out, do a couple hour check and when they're done, they just walk over. And at some point we'd love to get, like for them to get inside their own box and another one, like get it ready to go and put it on a pallet and we can just start shipping them out. It will get in its own box and another one will throw it on the pallet and ship it out.

9. 2:07:01

   For sure. Yeah. That's not hard things though. These are like, that's not, you know what I mean? Yeah, it's just interesting to think about. I feel like it comes off the line, gets in its own box, gets loaded on by another robot and then shipped off. The scary thing for me is those are like very rigid body things like cardboard and moving boxes and maybe using machines and stuff like those are like the scary stuff a couple years ago was like laundry that like literally moves it's like really never in the same spot it's like when you touch it it's like actually moving or we do like these like packages on this manufacturing conveyor system that like you you grab it it's literally moving white it's moving because the conveyor is moving down and then the packages are squishing each other and then the package itself is moving because it's

10. 2:07:44

    plastic when you're grabbing it those are the hard things that are compliant that are like really difficult for robotics because they're not like they're not like stationary when you touch them. So those are things that were like, man, that's going to be really tough to fold laundry. And for with code, it's been impossible. The reason you haven't seen like package logistics and stuff, some of the stuff automated is because like these bags are just like heart, they're compliant. They're just tough. You can't model them. And now we have like, we put it all in a neural net. They basically instantly worked. When we were working with our, we have a logistics customer we're working with. They do like soft packages and we sign them. They're like, we want to move these packages on the conveyor system and we've put videos out about it and stuff

11. 2:08:20

    the first month we um signed them uh inga who runs you know accounts was like we need to we need to uh we need to do we need to do this for them or they're gonna be really unhappy and i was like i was like dan that's like a compliant material that uh is moving while you're touching some of them touch and they're there's something hard inside some of them are squishy they're like there's tons of we're gonna move every three seconds we got to find the barcode put it down and put it in the middle of the conveyor every three seconds of package.

12. 2:08:47

    I was like, it was 50-50 shots works. And it's got to be with a neural net. And we got a bunch of data, trained it, policy, and right away it worked. And I was like, holy shit, this is like, it worked really good. And for some reason, the neural nets do extremely well under those high veritability environments that's extremely diverse. They can learn the representations extremely well across a wider distribution.

13. 2:09:13

    and um they just love it folding t-shirts towels like packages like no problem wow stuff that would like you know you're replanning very fast as you're as these things are all moving it's doing that in real time it's just like it just works deep deep deep learning just works on humanoid