WEBVTT 00:00.000 --> 00:08.000 I think we can start, and people are still coming in. 00:08.000 --> 00:14.280 So basically this talk is going to be about how can we, as an open source community, be able 00:14.280 --> 00:20.200 to build a manoried and especially general purpose, you manoried, right, that can do kind 00:20.200 --> 00:21.200 of anything. 00:21.200 --> 00:25.800 I'm going to highlight draw, which is a project I'm building, and feel free to test 00:25.800 --> 00:26.800 it out. 00:26.800 --> 00:28.200 You don't even have to test it out. 00:28.200 --> 00:32.720 I hope that this talk inspire you enough to give it a try. 00:32.720 --> 00:37.720 So you may remember me from first-end last year, I came with a small robot that was able 00:37.720 --> 00:40.080 to cut itself, to do stuff. 00:40.080 --> 00:45.040 And the cool thing is that after this talk, there was the school founder of hugging 00:45.040 --> 00:49.680 faith that gave me a call, who was like, that's really impressive, that's really cool. 00:49.680 --> 00:54.480 You should join us to do human rights, and so that's why I'm doing this talk right now. 00:54.480 --> 00:58.480 That he was not really like enthusiastic about the robot coding himself, it was kind 00:58.480 --> 01:01.720 of like, yeah, maybe we should just do end-to-end learning, so that's kind of what we're 01:01.720 --> 01:03.320 going to talk today. 01:03.320 --> 01:06.000 And for me, I'm just a full-time open source developers. 01:06.000 --> 01:11.040 I have my own company, and you can hire me if you want, but yeah. 01:11.040 --> 01:19.160 So maybe just for start, why do we need humanoids, why can't I have all the things like 01:19.160 --> 01:20.160 all the type of robots? 01:20.160 --> 01:26.280 The thing is AI strangers award, if you probably realize with Chargipity, and Robotics 01:26.280 --> 01:29.320 have suffered for years from not being able to generalize, right? 01:29.320 --> 01:35.280 So if you have very good robots in factories, very capable, but they will never be able 01:35.280 --> 01:40.680 to generalize now, if you want him to do pancake, he'll never be able to do. 01:40.680 --> 01:46.400 But the thing is that we kind of know how to generalize with artificial intelligence when 01:46.400 --> 01:48.560 it's about mimicking humans, right? 01:48.560 --> 01:52.080 So Chargipity, you give him a lot of text, and then he starts speaking. 01:52.080 --> 01:55.040 We don't really know why, but he starts doing this. 01:55.040 --> 01:58.240 And so we kind of want to do the same thing with robots. 01:58.240 --> 02:02.640 And the way to do it with robots is we probably need to have a robot in the human shape, so 02:02.640 --> 02:06.160 that we can show him how to do stuff, and then he's going to be able to do so. 02:06.160 --> 02:09.800 It's kind of our hope, and nobody should work, but through this, we don't actually know 02:09.800 --> 02:15.240 how AI works, and so this is kind of like, how it is. 02:15.240 --> 02:20.720 So let's maybe define what's open source general purpose seminars. 02:20.720 --> 02:24.120 There's kind of like three big components that we need to pay attention. 02:24.120 --> 02:25.640 First one is AI, right? 02:25.640 --> 02:28.880 We need AI to be general purpose. 02:28.880 --> 02:34.000 We need hardware, we need open source hardware, hopefully, so that we can really think 02:34.000 --> 02:36.960 around as a hardware, and then we need software. 02:36.960 --> 02:41.040 And if we have those three components that is open source, it means that we have full control 02:41.080 --> 02:45.400 of the human rights, and we can repair it, and maintain it, which is actually the hardest thing, 02:45.400 --> 02:46.400 right? 02:46.400 --> 02:50.640 It's really easy to buy, it's really easy to build, it's very hard to maintain, and it's 02:50.640 --> 02:54.680 true for AI, true for software, it's true for hardware. 02:54.680 --> 03:00.120 Like if you go to this adventure, you will break things, if the AI is a hardware, and being 03:00.120 --> 03:04.720 able to just really point apart, it's going to go a long way. 03:04.720 --> 03:09.320 And if in the future we want to have children to live on Tatooine and be able to build 03:09.320 --> 03:17.880 this 3PO, this is the way to go, there's no way to do it with closed source hardware. 03:17.880 --> 03:19.280 But is it really attributable? 03:19.280 --> 03:23.640 Well, the thing is, there's a huge community of people doing open source humanoids now, 03:23.640 --> 03:27.200 it's kind of like a booming thing. 03:27.200 --> 03:33.400 I work with Hagenphase, they have all kind of a naxumbly kit from scratch, and they will 03:33.400 --> 03:38.840 also teach you how to print 3D parts, build it, connect several motors, and how to connect 03:38.880 --> 03:41.480 it to computers to be able to control it. 03:41.480 --> 03:44.520 And in this way, you can also build your own humanoids, right? 03:44.520 --> 03:48.520 You don't even have to follow someone, you can just take maybe a template and modify it. 03:48.520 --> 03:56.040 So I would deeply, if you into hardware and deeply calling you to try it out, I think 03:56.040 --> 04:01.160 it's there, it could be very fun, we can project there. 04:01.160 --> 04:06.960 And also, like yeah, it's a couple of movements, and then AI is also having a moment, 04:07.000 --> 04:11.280 especially open source AI, from I probably heard about deep seek, 04:11.280 --> 04:16.360 mistrol, Kuan Vial, Hagenphase, there's a lot of open source motors that is popping up, 04:16.360 --> 04:21.000 but also about the training algorithm behind it, because at the end of the, 04:21.000 --> 04:25.360 a huge chunk is going to be about training, and there's even more specialized, 04:25.360 --> 04:27.560 especially for robotics, that is popping up. 04:27.560 --> 04:32.160 So if you into AI, I would highly recommend to check those things out as well. 04:32.160 --> 04:39.600 And so this seems to be kind of a solve issue, or at least there's a huge community around it already. 04:39.600 --> 04:45.200 And so the missing part is kind of the software, from my opinion, right? 04:45.200 --> 04:47.920 And this is what I'm trying to solve with straw. 04:47.920 --> 04:51.440 The idea is that you have hardware on one side, you have AI on the other side, 04:51.440 --> 04:57.200 and you're really trying to make those two connect in a way that is very efficient and very fast. 04:57.200 --> 05:01.760 So if you only have AI, what you typically do is kind of request reply part, 05:01.760 --> 05:04.800 and where you have an API, you send a request to whatever you want, 05:04.800 --> 05:08.400 and then you get a response, it doesn't really work for robotics. 05:08.400 --> 05:11.440 What you want is kind of a pub sub type of system, 05:11.440 --> 05:16.160 where you would be able to publish a data from, for example, a camera or microphone, 05:16.160 --> 05:18.880 and then the AI is going to be able to subscribe to them, 05:18.880 --> 05:23.440 and then get the latest data, and then be able to inference or whatever, 05:23.440 --> 05:25.520 and reply back. 05:25.520 --> 05:27.440 And so that's what we're trying to do. 05:27.760 --> 05:30.720 It's actually not very complicated to do pubs, so if you into software, 05:32.240 --> 05:35.920 what is actually very hard to do is how do you distribute the code from 05:37.760 --> 05:42.240 on different type of hardware, different type of requirement from the software side. 05:42.240 --> 05:44.400 You might have code that is written in Python, if you say, 05:44.400 --> 05:47.280 you might have code written in Rust, if it's for robotics. 05:47.280 --> 05:52.400 And so how do you make it very easy and accessible for anyone to start on the weekend, 05:52.400 --> 05:57.200 and not have to install a million compilers and be able to figure out, 05:57.200 --> 06:01.440 if it's Linux, if it's Nvidia, if it's different type of hardware acceleration. 06:02.000 --> 06:06.000 And so what we try to do is to push everything on tip, 06:06.000 --> 06:12.080 which is Python, the Python package manager, and then make it just peep installable. 06:12.080 --> 06:14.720 So everything from there should be peep installable, 06:14.720 --> 06:17.520 and this is kind of the way we think to source a problem, 06:17.520 --> 06:22.160 people will be able to automatically detect which OS using, 06:22.160 --> 06:26.960 which architecture using, and hopefully we can even push it more with 06:27.040 --> 06:28.880 a which acceleration you're using as well. 06:28.880 --> 06:34.480 And we can also package Rust or CC++ code into Python as well, 06:34.480 --> 06:35.920 and then people can just use it from there. 06:37.120 --> 06:42.880 All right, so maybe I can just show you very quickly how do our works, 06:42.880 --> 06:45.760 what hard work is that you define a YAMO graph. 06:45.760 --> 06:48.000 So you have kind of a YAMO configuration, 06:48.000 --> 06:50.400 what I say, okay, I want a camera, I want a narrow model, 06:50.400 --> 06:51.600 and I want a visualization. 06:52.240 --> 06:54.400 This is the input, this is the output, 06:54.720 --> 06:59.680 so the camera is going to send a flow of images, 06:59.680 --> 07:02.800 and then we will have a narrow model on the vision, 07:02.800 --> 07:06.480 so it's just going to subscribe to it with an input keyword, right? 07:06.480 --> 07:09.920 And then you can have environment variables to specify it. 07:09.920 --> 07:12.320 And so I can show you quickly how it looks like, 07:12.320 --> 07:14.720 if you're using your own computer, 07:14.720 --> 07:19.200 so it's kind of the repository from Gidham, 07:19.200 --> 07:22.240 and in tight-door repository, you have an example space, 07:22.240 --> 07:24.960 where you will have all the examples that we've built, 07:24.960 --> 07:27.440 so one we're going to look at today is a VLM1. 07:27.440 --> 07:29.440 So VLM sent for vision language model, 07:29.440 --> 07:32.960 so one that's able to react to text and image at the same time. 07:32.960 --> 07:38.080 And so I'm going to show you one that is very easy, 07:38.080 --> 07:41.440 that I just show you, and it's just like three nodes, 07:41.440 --> 07:44.640 so three, three separate process, right? 07:44.640 --> 07:47.840 And each node is a separate process. 07:47.840 --> 07:49.360 I don't know if I mentioned this. 07:49.360 --> 07:51.600 And so I say, oh, here's three nodes. 07:51.600 --> 07:54.800 I have one output from the camera, which is an image. 07:54.800 --> 07:59.520 I have a ticker because I want to coin VL model, 07:59.520 --> 08:04.560 so coin VL is an AI model to do inference on the timer, 08:04.560 --> 08:06.720 and then we have a clock, right? 08:06.720 --> 08:10.960 So if I do this, I have to first install everything, right? 08:10.960 --> 08:14.000 My computer maybe he doesn't have to install everything. 08:14.000 --> 08:16.320 I already installed the right, it's very easy. 08:16.320 --> 08:18.880 And when I have to install nodes, I hadn't just say, 08:18.880 --> 08:24.240 do our build, and then he's going to figure out maybe too fast. 08:24.240 --> 08:26.720 So let's say I want to start from scratch. 08:26.720 --> 08:28.160 I want to create a new environment. 08:28.160 --> 08:32.080 Let's say I build a new environment randomly, 3.11. 08:32.080 --> 08:36.080 What it's going to do is I can use UV, which is a Python package manager, 08:36.080 --> 08:38.560 build also in Rust, so it's safe. 08:38.560 --> 08:41.120 And then I say, OK, I want to use this one. 08:41.120 --> 08:42.960 And now I want to install everything from this. 08:42.960 --> 08:47.840 And then we use UV, it's going to pull everything has to pull, 08:47.840 --> 08:50.560 and then it's going to install everything has to install, 08:50.560 --> 08:52.240 and then it's just going to work. 08:52.240 --> 08:55.200 And hopefully, finger crossed, it's just work from scratch. 08:55.200 --> 08:58.640 But the idea that everything from there is to confirm by the developers, 08:58.640 --> 09:01.600 so you kind of put all the specification within 09:01.600 --> 09:06.480 Pyproject, Pypy, until in a way that it should work for you on your computer. 09:06.480 --> 09:09.040 When I want to run it, I just have to say, do our run. 09:11.840 --> 09:14.720 OK, I think I missed the slide, but it's OK. 09:14.720 --> 09:17.360 And then it's going to be able to pop it. 09:17.360 --> 09:19.440 And now it's going to take a bit of time, 09:19.440 --> 09:23.280 because we have to download the model from the internet. 09:23.280 --> 09:24.720 We can also have local models. 09:24.720 --> 09:26.320 You can have adapters. 09:26.320 --> 09:29.120 That's what's been shown later. 09:29.120 --> 09:32.320 But one, OK, I'll see if they're my phone is fine. 09:32.320 --> 09:37.680 And then, say, maybe I just show you what I just showed you before, 09:37.680 --> 09:38.720 which is this. 09:38.720 --> 09:40.880 So Pypy install is just a Pypy install. 09:40.880 --> 09:41.840 There's no compiler. 09:41.840 --> 09:43.040 It's compile of free. 09:43.040 --> 09:46.000 The idea is that a lot of people, especially if you're starting up with the buildings, 09:46.000 --> 09:49.600 there's so many, so many things to learn that you might not want to have 09:49.600 --> 09:52.000 to install compilers and stuff. 09:52.000 --> 09:56.560 And when you run it, it's separate process, linked with chat memory. 09:56.560 --> 10:01.040 So that's a very big thing to have a performance system, 10:01.040 --> 10:02.560 especially for war weeks. 10:02.560 --> 10:04.560 And so now everything is running. 10:04.560 --> 10:09.920 And so now I will have the VR model, like it may be smaller. 10:10.880 --> 10:12.640 Can I zoom in? 10:12.640 --> 10:13.440 No. 10:13.440 --> 10:15.040 All right, I'm sorry. 10:15.040 --> 10:20.080 Just zoom, zoom, maybe? 10:20.080 --> 10:21.840 OK, zoom in, perfect. 10:21.840 --> 10:26.320 So you can see that the model is just describing. 10:26.320 --> 10:28.720 But he's able to have, is it too small, still? 10:28.720 --> 10:30.000 Yeah, maybe. 10:30.000 --> 10:31.120 OK, sorry. 10:31.120 --> 10:33.040 Sorry, sorry. 10:33.040 --> 10:35.360 Zoom in. 10:35.360 --> 10:37.360 Zoom in. 10:37.360 --> 10:40.480 OK, is it better? 10:40.480 --> 10:42.240 Is it OK? 10:42.240 --> 10:43.280 Oh, yeah. 10:43.280 --> 10:45.920 So now he's just describing what he's seen. 10:45.920 --> 10:49.840 And if he's seen me, he's going to say, hey, there's a guy looking at him. 10:49.840 --> 10:52.640 So I try to do peace sign. 10:52.640 --> 10:54.480 He's going to say, oh, you're making a peace sign. 10:54.480 --> 10:57.760 And so he's very descriptive, right? 10:57.760 --> 11:01.120 All right, he's carrying, OK, that's fine. 11:01.120 --> 11:03.440 And so it's very capable, right? 11:03.440 --> 11:06.720 Everything is running locally on my MacBook laptop. 11:06.800 --> 11:10.720 And now, like, OK, it's really cool, it's really nice. 11:10.720 --> 11:14.000 But it's actually very able to see stuff, right? 11:14.000 --> 11:16.800 If it is built, it's right with it, it's the built to see stuff. 11:16.800 --> 11:20.720 And so yeah, so that's super easy stuff. 11:20.720 --> 11:22.800 And so you can do cool demos. 11:22.800 --> 11:28.320 And so that's kind of like a very quick getting started with Dora. 11:28.320 --> 11:34.000 And yeah, so I just said general is important. 11:34.000 --> 11:36.480 And we tried it basically. 11:36.480 --> 11:39.360 And we said that, oh, if we show him, it's very fast. 11:39.360 --> 11:44.320 And compared to CCP, or HTTP, if you do API calls, especially when you 11:44.320 --> 11:45.680 have more and more data, right? 11:45.680 --> 11:50.720 So if you have HTTP or you have multiple cameras on the reachy, 11:50.720 --> 11:53.680 currently we're going to do a demo with this three. 11:53.680 --> 11:55.360 So it can very take some times. 11:55.360 --> 11:57.840 And if you really want to have a quick feedback, 11:57.840 --> 12:00.240 it's going to be very important. 12:00.240 --> 12:04.400 And also, we realize that not only is CPU kind of 12:04.400 --> 12:06.480 sharing very important, you also have the V-run, 12:06.480 --> 12:11.920 and you might want to do like Shed V-run between nodes, right? 12:11.920 --> 12:14.800 And this is also possible with Cuda, hopefully in the future, 12:14.800 --> 12:16.240 we can do it with more. 12:16.240 --> 12:18.400 And it's also safe a lot of time because you don't have to 12:18.400 --> 12:21.440 copy from CPU to CPU, CPU to GPU. 12:21.440 --> 12:22.720 So yeah, all right. 12:22.720 --> 12:26.000 And this is all possible because we're using Apache Arrow, 12:26.000 --> 12:29.040 which is a unified memory format. 12:29.040 --> 12:31.600 And basically they say, OK, an array, it's like this. 12:31.600 --> 12:33.040 F-16 arrays like this. 12:33.040 --> 12:35.680 And then you can use it from Python, C-C++, Rust. 12:35.680 --> 12:39.600 And there's already a lot of libraries that can use 12:39.600 --> 12:43.840 this unified memory format from new MPHPANA to PyTorch. 12:43.840 --> 12:48.480 And the good thing is that you can use it on CPU, but not on GPU. 12:48.480 --> 12:53.120 So the latest benchmark on GPU is actually using Arrow GPU. 12:53.120 --> 12:57.200 And so it's kind of not just for Cuda, but it will be able to 12:57.200 --> 13:02.000 scale on, for example, macOS or other types of platform. 13:02.000 --> 13:05.200 And you will be able to share CPU memory from Python to Rust, 13:05.200 --> 13:05.760 hopefully. 13:05.760 --> 13:08.640 So that's kind of the idea, there's a talk, I haven't, like, 13:08.640 --> 13:12.000 develop this, there's a talk in this for them from Arrow. 13:12.000 --> 13:14.400 If you want on Thunder, I think. 13:14.400 --> 13:15.600 Yeah. 13:15.600 --> 13:18.400 And so if you want, OK, so that's very cool. 13:18.400 --> 13:23.040 If you want, we can do a last demo, which is with the richy. 13:23.040 --> 13:26.080 And so what we need to do, if you want to make it even more capable, 13:26.160 --> 13:30.000 it's just to connect the output of the VLM into a robot, 13:30.000 --> 13:32.320 and then we have a moving robot, right? 13:32.320 --> 13:37.440 We just saw that it was very able to describe stuff in the world. 13:37.440 --> 13:41.520 And so yeah. 13:41.520 --> 13:48.640 And yeah, yeah, yeah, yeah, yeah, yeah, yeah, yeah, yeah. 13:48.640 --> 13:52.000 It's such an amazing part, yeah, okay. 13:52.000 --> 13:54.880 So yeah, so what you do is you take the VLM output, 13:54.880 --> 13:59.200 and then whatever he says, you're going to map it to an action. 13:59.200 --> 14:03.680 And so for example, if I try to say hello richy, 14:03.680 --> 14:05.600 can you do a hand wave? 14:05.600 --> 14:07.440 He's going to be able to hand wave me. 14:07.440 --> 14:11.040 And so basically, he's able to use this VLM description, right? 14:11.040 --> 14:12.480 And to do an action. 14:12.480 --> 14:14.720 And if I give him stuff, he's able to give stuff, 14:14.720 --> 14:16.640 he's able to go left, right? 14:16.640 --> 14:20.400 Depending on if we're sending on the left, on the right side of the camera. 14:20.400 --> 14:26.640 And also what we can do is have a speech to text model. 14:26.640 --> 14:31.120 And so what is going to be able is that now we have a fixed prompt, 14:31.120 --> 14:35.120 but we can also have a modified prompt, okay? 14:35.120 --> 14:39.920 So yeah, this is good. 14:39.920 --> 14:44.480 And so for example, we can get a napkin. 14:46.480 --> 14:47.680 Yeah, okay, that's fine. 14:48.640 --> 14:51.280 And so for example, with the speech to text, 14:51.280 --> 14:55.360 we just add a no, no, the separate process is going to run a whisper model. 14:55.360 --> 15:00.240 And then he's going to be able to translate whatever we say to a text, 15:00.240 --> 15:02.800 and then we're going to put the text into the prompt, 15:02.800 --> 15:05.760 and then you're going to be able to modify his VLM, right? 15:05.760 --> 15:08.800 Instead of just having us, he's going to be able to, 15:08.800 --> 15:11.200 I think he's kind of stuck because it's mine, okay? 15:11.200 --> 15:13.200 And then he's going to, yeah. 15:17.680 --> 15:19.680 Yeah, yeah, yeah, yeah, yeah. 15:23.360 --> 15:24.640 Yeah, that's kind of the idea. 15:24.640 --> 15:27.440 It's a very, very simple stuff, nothing to complicated, 15:27.440 --> 15:31.600 but you can be like, but wait, wait, he used to describe stuff. 15:31.600 --> 15:36.560 Now he's moving this kind of weird, the thing is like we train it, basically. 15:36.560 --> 15:39.600 And how we train it is that we collect a lot of data, 15:39.600 --> 15:42.400 and then we say, okay, so when I do this, 15:42.400 --> 15:44.640 you're going to wait, and you're going to say hand wave. 15:44.640 --> 15:47.840 And when I do this, you're going to take a napkin, 15:47.840 --> 15:51.600 and then he's going to be able, then we're going to have a training pipeline, 15:51.600 --> 15:56.240 that's going to be able to map the VLM output to those very keywords, right? 15:57.200 --> 16:00.000 But this is just one way to do training, 16:00.000 --> 16:02.080 and hopefully there's multiple ways. 16:02.080 --> 16:04.640 We don't want to fit to just one AI model, right? 16:04.640 --> 16:06.640 We don't want to fit to any type of AI model, 16:06.640 --> 16:09.920 you might have different type of stuff you want to do, right? 16:10.480 --> 16:14.320 And so yeah, but this is all possible with storage, 16:15.360 --> 16:20.080 we have a couple of training pipelines, and so there's kind of the idea. 16:23.360 --> 16:24.480 Okay, what's next? 16:24.480 --> 16:28.400 Then basically what we're going to do is add even more models. 16:28.400 --> 16:33.440 So as you can see now, the robot is only doing pretty fine stuff. 16:33.440 --> 16:39.040 And the thing is like, we currently don't know how to do AI infertile action, 16:39.120 --> 16:41.120 in an environment that he never seen. 16:41.120 --> 16:42.480 It's kind of like extremely hard. 16:42.480 --> 16:45.600 If we in the lab, and I know exactly where stuff is, 16:45.600 --> 16:48.480 and the light is always constant, I will be able to have 16:48.480 --> 16:51.040 predefined movement, and he's going to be able to grab stuff 16:51.040 --> 16:54.320 in a way that he's kind of generalistic, but we haven't figured out 16:54.320 --> 16:56.800 this on the completely unseen environment, 16:56.800 --> 16:58.560 so that's why I can't really show it. 16:58.560 --> 17:00.640 But we're working really hard, okay? 17:00.640 --> 17:03.600 And we're getting there, and hopefully next year, 17:03.600 --> 17:07.200 we could have something like this, and we also want more robots. 17:07.360 --> 17:09.600 This one is a rich sheet too. 17:09.600 --> 17:11.280 There's some part of this open source, and some part 17:11.280 --> 17:13.680 is not open source, and I know that we at first 17:13.680 --> 17:17.440 them, so hopefully we can have even more open source stuff. 17:17.440 --> 17:22.480 And we can directly from the paper installation, 17:22.480 --> 17:26.480 it just work on your computer as well, and add even more demos 17:26.480 --> 17:28.240 from speech to text to text to speech. 17:28.240 --> 17:29.440 I said text to speech. 17:29.440 --> 17:32.080 I can maybe show you later, text to speech demo. 17:32.080 --> 17:35.840 And we also have, yeah, latency optimization, 17:35.840 --> 17:38.560 and make draw more mature. 17:38.560 --> 17:40.000 So thanks for listening. 17:40.000 --> 17:42.160 I think I may be early a bit, yeah, okay. 17:42.160 --> 17:46.320 And so yeah, so this is first them 2024. 17:46.320 --> 17:48.880 It really helped us a lot. 17:48.880 --> 17:52.240 So we're very, super happy to be at first them 2025. 17:52.240 --> 17:54.560 And if you want, I can do just one last demo. 17:57.840 --> 18:04.160 And I can show you kind of a speech to speech demo, hopefully. 18:04.160 --> 18:08.400 And so it might, yeah, my screen is sharing, okay. 18:08.400 --> 18:16.720 And so, so the thing is, the thing is, we have kind of 18:16.720 --> 18:20.800 dumb down the robot to be able to move, so he's not able to 18:20.800 --> 18:24.160 to talk to us anymore, but we also trying to fix this out. 18:24.160 --> 18:25.680 If we, basically, if you talk to the road, 18:25.680 --> 18:28.000 it's just going to say wait or like, hand away. 18:28.000 --> 18:30.560 And so that's not really what we want. 18:30.560 --> 18:36.240 But I can show you with Quen Vial, if I say hello, how are you? 18:36.240 --> 18:51.600 And so basically, I can have, is working. 18:51.600 --> 18:52.320 What do you see? 18:52.320 --> 18:58.640 Stan's in front of a chokeboard with two written on it. 18:58.800 --> 19:02.720 A person stands in front of a chokeboard with two hundred and two written on it. 19:06.720 --> 19:10.720 Yeah, so I'm kind of, I wasn't really too prepared. 19:10.720 --> 19:13.040 But basically, the idea is that you can ask him stuff. 19:13.040 --> 19:14.880 And I think there's a lot of noise. 19:14.880 --> 19:17.840 And then he's going to be able to reply to you in a way that it's 19:17.840 --> 19:18.720 like, comprehensive. 19:18.720 --> 19:22.720 And you can put in a speech deck to speech node at the very end, 19:22.720 --> 19:24.320 so that he makes noises. 19:24.320 --> 19:27.120 But like, he's kind of like looping on himself now. 19:27.120 --> 19:29.440 And kind of not really working. 19:33.840 --> 19:35.120 OK. 19:35.120 --> 19:36.960 What is the men doing with his hands? 19:45.120 --> 19:47.440 Man's holding up two hearts with his fingers. 19:47.440 --> 19:50.720 So yes, that's kind of the idea. 19:50.720 --> 19:52.880 All right, that's all I have for today. 19:52.880 --> 19:54.320 Do you guys have any questions? 19:54.320 --> 19:56.960 Yes, we're standing in front of a chokeboard with a number two 19:57.920 --> 19:59.920 right now, right now, right now. 20:07.920 --> 20:11.920 So what is the hardware board that's processing this? 20:11.920 --> 20:13.600 Is it hardware, go? 20:13.600 --> 20:17.840 The board on this computer on. 20:17.840 --> 20:21.120 Also, you have a small computer on this, 20:21.120 --> 20:23.040 but just from the manufacturer itself. 20:23.040 --> 20:26.800 But the idea is that the software and the AI can run on any computers. 20:26.880 --> 20:29.360 Hopefully, if the computer is able to run it. 20:29.360 --> 20:33.440 So if it's Linux, Mac, Apple, or whatever, 20:33.440 --> 20:36.560 we're trying to make it run anywhere. 20:36.560 --> 20:38.640 Yeah, yeah, for the software sign. 20:38.640 --> 20:39.840 Yeah. 20:39.840 --> 20:41.440 Yeah. 20:41.440 --> 20:43.200 That's good for FB. 20:43.200 --> 20:46.800 Your project is trying to stop a similar problem. 20:46.800 --> 20:49.600 FB is trying to keep forward FB. 20:49.600 --> 20:53.520 Yes, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so, so you 20:53.520 --> 21:00.080 question is about, is am I trying to kind of do the same thing as Andre did for phones 21:00.080 --> 21:04.880 on robots, kind of, right, I wouldn't like say this because, you know, 21:04.880 --> 21:09.040 but Andre is a bit pretty big, but the idea is like, at the very least, 21:09.040 --> 21:13.680 yeah, we we really want to make it, so that it's very easy for anyone 21:13.680 --> 21:19.760 to pipeline stuff and make it run and be be able to people install it. 21:19.760 --> 21:22.800 Hopefully, everything is open, so, so, so, so, so, so, so, so, so, so, so, so, so, so, 21:23.360 --> 21:25.760 If people want to have, like private closest thing, feel free 21:25.760 --> 21:29.120 that's totally fine, but yeah, that's kind of the idea, and everything is 21:29.680 --> 21:32.800 is, but the big thing is, like, you have to have good communication, 21:32.800 --> 21:39.780 everything is communicating with 잘� nochering, yeah. 21:40.640 --> 21:41.860 Should I? 21:41.880 --> 21:45.040 Do you have the list of open source open source, you've�rian projects? 21:45.040 --> 21:49.040 I don't actually, I could—I was, yeah, but I don't… 21:49.040 --> 21:52.640 There's probably, like, a potato for this probably, yeah. 21:52.640 --> 22:09.640 How did you come around with the performance differences with Ross, and especially how do you feel like this project is going to close the gap with something like Ross? 22:09.640 --> 22:17.640 Because I found Ross to be a very frustrating system to use, but I think a lot of people use it just because it has all these simulation tools and visualization tools. 22:18.640 --> 22:24.640 Completely a bit. So actually I put Ross to you here. It's kind of like if you use Ross to. 22:24.640 --> 22:30.640 So Ross to has really been built for C++ and if you use in Python it's very slow. 22:30.640 --> 22:37.640 So this is kind of a graph of like what the difference isn't in Python with like no, like from Vacy, Ross to with Dua. 22:37.640 --> 22:44.640 And so normally Dua should be very fast compared to Ross to in Python on CC++. They also use trial memory. 22:44.640 --> 22:46.640 So like the performance should be similar. 22:46.640 --> 22:53.640 And but they they have some form of like GPU stuff, but it's apparently it's very hard as well. 22:53.640 --> 22:59.640 The idea that we like I think Ross to was built in a way that it's okay for robotics for robotic people. 22:59.640 --> 23:02.640 And what I'm trying to do is just doing robotics for AI people. 23:02.640 --> 23:11.640 And yeah, but I think they have a very good stuff as well on many different fronts that I may never be able to to bridge. 23:11.640 --> 23:17.640 But the idea is like if we just can do a very good work on AI, I think it's good enough for me. So that's very good. 23:25.640 --> 23:29.640 So I was curious about how you found 23:30.640 --> 23:37.640 As curious how you found your interactions with like the rest of hugging face and the level of robot project. 23:37.640 --> 23:38.640 Do they like call people's work? 23:38.640 --> 23:44.640 No, we were we closely the thing is basically there's another guy working with me called Philip, but he's been. 23:44.640 --> 23:49.640 He has a family thing of good family thing, but he's not working at the moment. 23:49.640 --> 23:56.640 And so I'm just trying to to to make Dua progress, but hopefully in the future I can work more and more with hugging face have more and more integration. 23:56.640 --> 24:06.640 There's a lot of things that's very common. I know there's a lot of things that we also can do with Dua that we can't do with like hugging face ecosystem at the moment. 24:06.640 --> 24:14.640 As as it is in the way that we like for me, for example, more and more with move forward what we want to do is thrust on Python, right. 24:14.640 --> 24:20.640 So for example, the visualization that you saw it seemed to rerun visualization is fully. 24:20.640 --> 24:25.640 This is fully rust based. It's just packaged on Python, right. 24:25.640 --> 24:31.640 And what we want is Dua is to be able to bring all the rust ecosystem into AI people, right. 24:31.640 --> 24:34.640 In a way that's very efficient in some way as well. 24:34.640 --> 24:39.640 And so that's why we really want to integrate both, but like hugging face house or AI model or thing. 24:39.640 --> 24:42.640 So how do we make those two? I combine, you know. 24:42.640 --> 24:46.640 That's kind of what we want to do and we work really closely and hopefully we we can have more. 24:46.640 --> 24:51.640 But as a moment, I'm kind of like not enough time to do it or so that's kind of how it is. 24:51.640 --> 24:54.640 Yeah, we're going to reply. Okay. 24:54.640 --> 24:56.640 All right, so thank you very much for listening. 24:56.640 --> 24:59.640 Very happy to tell you all.