00:29:20 · Siyuan Huang
And then we predict the actions. The actions are the motion commands. So it's not just only the motion skeleton, it's the motion command. It indicates how the robot should control its bodies and then try to control the robot. And then the controller converts the motion command into the whole body control signal. And then this can convert the command and the images and instructions into the graphite or into the actions for the robot. And during the inference, we can actually have the human interventions we introduced before we use a data action design. So the humans can use a PICO or a teleoperation system, the motion capture, they can intervene when they think the robot has made some mistakes. So this is critical for the future development and the human in the loop learning. So this is the basic structures of the model. And in the training time, we have designed certain dynamics, like the model in Cosmos, like inverse dynamic models. So we designed this kind of learning dynamics to have a like more like fully, fully exploration of the training data. In this version, we have about, um, you know, 10,000 hours of egocentric and human data to train the Delta-0 Foundation models. And during inference, we can, do, similar to the training we can use the multimodal self-attention to attend the images and instructions, and then we can generate the action output, convert them into the command and control then the GYN robot. So this is the design. And here are more breakdowns for the design principle for the controller. So actually we discussed about control a lot during the last several months. Especially for the development of humanoid models. First from the controller, right? Do this kind of dance, performance, and have amazing behaviors from humanoid foundation models. So we have some Certain principles, design principles. The first is that the control infrastructures should be designed for the brain learning. And then the second is that we can try to make the body movement, body's capabilities available to the policy. So this, this try to make sure that the capability can be connected between the controller and the brainer. And in detail, we have certain requirements for this kind of controller. So the first is the controller can provide an interface for the intent and intervention. So this motivates DeltaBit, DeltaController.
00:32:17 · Siyuan Huang
And second, the controller is a whole body. So we have seen controller that separate the upper body and lower body. That is easier for the VLA models or WAN models. But we really need a whole body controller for what So this is critical for a human robot and critical for the deployment. So this is critical for a human or robot and it's critical for the deployment. So we have seen demos from papers, actually We have a lot of papers as well. But from these papers, we didn't observe a very reliable performance from the controller, which means that the robot can have some jittering performance. They can have turning performance. So this kind of performance, although you can hide this kind of performance during the demo, right? We all know that the robotics demo is tricky, but actually if you want to turn that capability into the deployment, into production, This kind of stability is unacceptable for the humanoid foundation model. So we want the controller to have really reliable and stable performance throughout different tasks and throughout the tasks that requires compliance. This is pretty important for the humanoid foundation model. And the fourth, the first one is execution fidelity and responsiveness. So we, we want the controller have like a better accuracy, especially for the task that requires high, high, high That's just accuracy, like the turning, like the turntable test. And the fifth is the compliance. When the robot is trying really to use more, more forces for opening the door or for some heavy, heavy load. So we need the robot to use extra forces for the interaction. So it requires the robot to have the compliance while maintaining the balance of the whole body coordination. So these five principles are the principles within We think we need to maintain during the scaling process of the controller and it's important not only for the performing certain dance and important for the local manipulations, important for the future deployment of human robots. And here are some examples of Delta action intervention. And it's pretty interesting. We can see that in each task of the deployment, we can roll out the policy, the policy. And then if we found out some problems during the We can stop the actions, stop the policies, and we can do the interventions. And the intervention is basically on the delta position of the anti-factors.
00:35:15 · Siyuan Huang, Chris Paxton, Michael Cho
That means the robot is not following the actual trajectory of the performer. It is learning or is adjusting from the delta actions of the anti-factor. [00:35:27] S3: So you're just like nudging its hand, basically, and it's still controlling its fingers and the rest of its body. Is that right? [00:35:33] S2: Yeah. Yeah. We use the Delta as the input, not the whole body trajectory. [00:35:39] S3: But the fingers were closing. I was like, how is it set? [00:35:42] S2: Yeah. Yeah. Finger is not designed in this version. So it's not like the buttons. [00:35:48] S3: It's like the policy is still deciding, like, this looks like the time I should close my fingers. Okay. No, that's cool. Thanks. [00:35:54] S2: Yeah. We're doing it in our next version. [00:35:58] S0: Yeah, so I guess what you're saying is that because this tracks the delta or just a change, and so the human literally, let's say the, it looks like the human is taking over from the robot when the robot is already bending down. The human himself don't have to bend down to simulate that. Exactly. Okay, okay. Okay, a lot more comfortable for the human to do the intervention.
00:35:58 · Michael Cho, Siyuan Huang
[00:35:58] S0: Yeah, so I guess what you're saying is that because this tracks the delta or just a change, and so the human literally, let's say the, it looks like the human is taking over from the robot when the robot is already bending down. The human himself don't have to bend down to simulate that. Exactly. Okay, okay. Okay, a lot more comfortable for the human to do the intervention. [00:36:18] S2: Okay, yeah, sure. Yeah, we can see Xiao Zhan in front here. And during the execution, even for a challenging task, we can start from here and then we can intervene and start from the human as well. So you can sit there operating a robot that's doing one-fit stand. Yeah, it's quite intensive. Another example? Yeah, I think we are the first one to really tackle this system, the database control system for the And here you can use it as an [00:36:49] S0: interface.
00:36:52 · Siyuan Huang
[00:36:52] S2: Because the reason is that if you're using the whole body as the input or the The joint input for the trackers, you need to synchronize the actions for the human and the robot whenever you want to do the intervention. And this is basically not possible for, you can do this, but the efficiency is really, really low. So it's basically not possible for the real training, real RL training for the human robot. So this, I think this system is a must for the future of the human robot system. Yeah. And then we also compare some metrics of the trackers. We call it execution fidelity on everyday task motions. We compare with other like SOTA controller, including the heft, the Mimic light, the new version of the sonic, and the scale BFM. We compare with them in terms of the global root errors and body position errors. These are some tracking metrics. We evaluate these trackers on the daily tasks for the local manipulation. So these are not for the... Like the tracking dataset, we're using it on our daily local manipulation tasks. So it's actually proving that it can have a That has performance over this kind of daily interactions. And then we have, um, uh, more, uh, design for the, for the brain part. So basically I, I tried to go through this briefly is, um, a latent word action models and we design certain dynamics. So basically the vision And then we train the action branch and also the MMDIT, the multimodal diffusion models, tension models. And then we have certain dynamics, including the forward dynamics. Inverse dynamics, regional planning, and policy only. That corresponds to certain learning dynamics of the training according to different training rewards. We try to explore, fully explore the range of the data. So we designed these dynamics schemes. And then long-horizon execution is done by stage-by-stage instructions. So this stage can be provided by higher-level models. Currently, people are discussing whether GPT-6 can really replace the robot foundation models. But actually, first, it can serve as good high-level models. Maybe you can call it system two. So it can, it can decompose this kind of high level task into some atomic, atomic instructions for every, for more detailed sub steps. It can be a higher level models, an agent or a human. So this long horizon task can be staged.
00:39:47 · Siyuan Huang
And it can have some adaptations or can have some transitions from transition appearance or behaviors from task to another. And the exception is specified by the objectives. So delta 0 must learn how to reposition its body. So the position must be accurate before starting from the next actions. So this is the design we consider about the So the next section is about how we scaling the reliable humanoid intelligence. And this is quite interesting. We decompose this into three parts. The first part is we want to scale the data. We want learning from human data. I think from the first principle, we Doing all of this foundation model for humanoid robots is that we believe that human data is the richest source of robot learning. So we try to make use of this kind of high quality or low quality human data, including the upper body and also the lower body. We try to make use of this kind of data for the learning of the Delta Zero. And this motivates the design of the action space and motivate design of the data recipes from the Delta Zero. So actually we design our own devices of collecting such data. And here is a brief view of the collection devices. So it's basically an egocentric collection system with an extra fisheye camera aiming at the body part. So here is the fisheye cameras. So S is towards and down, towards the human body. So we can have a better view and extra view. Like this one is a fisheye camera. We can have an extra view of the human body. So this can increase the coverage. Because we have seen a lot of occlusion during the data collection system for the standard egocentric data collection system. So we want to adjust this kind of occlusion. Uh, in, in, in like, you know, in a better way. So we use this kind of visual camera. So we'd have a better coverage. We can use this view as input. We train a better model to estimate the whole body poses. I think this is one of the most promising way to scaling out the whole body data. Because right now we are using systems like motion capture systems for in-house data collections, whole body, and we use Pico, right? The motion capture systems, they are not scalable. And the Pico, honestly, is not very accurate, especially for sound. You need to do the calibration, and for the feed, it's not accurate along the way, a long time. So we designed our O-system to own...