00:00:00 · Michael Cho
[00:00:00] S0: Hello, guys. Welcome to another episode of RoboPapers. Today, we are really glad to have Xiaojian and Siyuan. Xiaojian is the CEO, and Siyuan is the chief scientist of a new startup called Delta Intelligence, and they've just released their model, Delta-0, and they've written a pretty detailed technical blog post along with it. So very glad to have both of you coming on to share your work. Maybe as a start, both of you can share a little bit about yourself first. Maybe Xiaojian can start first.
00:00:26 · Xiaojian Ma
[00:00:26] S1: Sure, sure, sure. Yeah, thanks Michael and Chris for the kind invitation. And so I'm Xiaojian, I'm the CEO and founder of Delta Intelligence. So Delta is a company that's dedicated on solving general intelligence for humanoid robots. So we're focusing on bipedal robots, this form of robot, because we do think that humanoid is the future and it's a general form of physical intelligence as well. And so the company is building this intelligence, building this foundation model, including the brain, which is the large models and also the controller. And also we have these four position hybrid control for our human robot as well. So that's basically what companies are doing. Delta-0 is basically our first foundation model that we're trying to showcase the capability we're having right now on these Unitree G1 robots. But we're actually expanding our adaptation not just to Unitree, but also to some other robots. RoboVender in China and other countries as well. So just to, you know, showing that our model has this cross involvement capability at this point. So maybe a little bit more about myself. Before founding Delta, I'm actually a researcher in machine learning and robotics. I've been doing this robotic research for a couple of years. Before I was returning to China, I joined BIGAI as a research scientist on Embody AI. I was a PhD student at UCLA, and during my PhD, I was doing a lot of internships in some major AI companies like Google, their robotics team, like NVIDIA Research, pretty much with Yuke, Yuke and Jim. And also, but way before they were founding the Jailab, And also a very brief internship at Google DeepMind as well. So all these three experiences are with our full-time research internship. So my project is mostly about building these large models and multi-model models for re-reasoning and embodied decision-making as well. So yeah, so basically that's what I'm coming from.
00:02:28 · Siyuan Huang, Michael Cho
[00:02:28] S2: Okay. Okay. Yes. Thank you for the invitation, Chris, Michael, we're old friends. And actually with this, I'm sharing about papers with my students for several times. And this is the first time I've shared about a blog about the Delta-0, our new foundation model. And I'm the chief scientist of Delta. And we released this blog from the end of the last month. And right now we have raised some discussions about how to approach the real criminal intelligence and how to build up the foundation model for the whole body local manipulation. I think we did a lot of posts at the RoboPapers with Chris and Michael. Pretty excited about the development of humanoid foundation models. And I think this blog is a summary of how we can really achieve a better humanoid foundation model and how to address the coupled controller hardware and model problems in this blog. Yeah. Yeah. So let's go over the demo first. So basically it's an indoor thing and we have two robots. And one robot is doing the, we call it a whole body dexterous manipulations. So actually the hardware is a Unitree G1 robot with two WUJI hands, two WUJI hands. And the first thing we show, the robot is doing a, is doing a turntable and putting the record into the turntable. And he would do the dance too. Yeah, can you hear the sound from the video? [00:03:58] S0: Yep, yep, we can.
00:03:58 · Michael Cho, Siyuan Huang, Chris Paxton, Xiaojian Ma
[00:03:58] S0: Yep, yep, we can. [00:03:59] S2: So our expectation is that in the future we can have a robot in our home and then can really do the same thing like humans. They can enjoy themselves, like putting on the record. And then listen to the music, do some little dance, and then take a rest, of course. And then the robot, of course, they can do the everyday housework for human beings. So they can arrange the bed, picking up a joystick from the bed, and then arrange the So these tasks are called long-horizon local ventilations. So basically it requires the whole body to do the coordinations and finish the whole body everyday tasks, like picking up a toy from the ground and put it on the bed. And then the robot can also interact with this kind of articulated object using certain force to open the door and then close the door, opening the dishwasher. Using the left hand to stitch to the ground and the robot can have a larger force. [00:05:03] S3: That's really cool. It was bracing itself against the fridge and pulling with the other hand. [00:05:08] S1: It's like using your core to adjust your body and trying to pull out the larger force. It's very human, very fluent. [00:05:17] S2: And check about the food in the refrigerator and take a rest as well. This is so cool, man. [00:05:23] S0: This is really long horizon. [00:05:25] S3: Also, I want to note how few cuts there were in that, right? Like there was just one, right? When you cut from the first robot to the second one, right? There are actually two cuts. [00:05:33] S0: The second one is continuous.
00:05:33 · Michael Cho, Siyuan Huang
[00:05:33] S0: The second one is continuous. [00:05:34] S2: The first one and the second one. [00:05:35] S0: Yep. That's great. I mean, all this opening of DOS is non-trivial. Like, you know, it's that thing. Very, very impressive. [00:05:45] S2: Okay, cool. [00:05:46] S0: Please take us through.
00:05:46 · Siyuan Huang
[00:05:46] S2: See ya. Okay. Okay. So basically the video is composed by two parts. The first part is the robot doing the very dexterous manipulation task. And second part is the robot doing very long horizon local manipulation. So basically it's about three minutes of uncut video and there are like nine atomic tests and more than like 30 steps for the robot to finish all these tests. So it's pretty challenging. For us, we have seen companies like Figure, they show a video like this, but it's really hard. As Michael also said, open, close this kind of task requires a lot of interaction with the articulated objects. It's pretty hard to connect those tasks because it should start from the end of the last atomic task. And this is difficult because we don't use any like SLAM based navigation still in this video. So the robot is purely driven by the model to do the navigation, to do the upper body interaction and lower body interactions, all of this stuff from one single model. So it's pretty challenging. And the blog introduces more details about our models. So first we want to introduce about how we like treated the problem of humanoid foundation model and why we think is really important for the, for us to tackle the problem of humanoid impetages. So basically, the humanoid form offers the possibility of working wherever people do. So we think that humanoid is a general form. It's better than a real robot, so it can do the works everywhere. And it shows that humanoid Human intelligence is what turns that possibility into useful and reliable actions. So basically we have human intelligence, we can drive all this human robot to do the useful and reliable actions. Right now we have seen a lot of demos and in the future, we want to turn these capabilities into deployments. And this is the vision of our companies, the data intelligence. And I think there are some challenges in Tackling this kind of female intelligence. So first we show the long-term local manipulations, right? So we expect the robot to do the long-term actions, not just some single pickup like actions in front of the table. So basically the challenge is like in long horizon tasks, like every So basically the surrounding will change, and the location will change, and the interaction mode will change, and there are a lot of dynamics.