00:23:25 · Ademi Adeniji, Vincent Liu, Chris Paxton
[00:23:25] S2: Yeah, I think that was pretty comprehensive. I think we should maybe look at some of the in the wild to inference time transfer results. Like I think there's some cool videos where you like fold a towel in your bedroom and it transfers. [00:23:37] S1: Yes. Let's see. We can probably find... You guys can still see my screen, right? I guess these aren't in my... Well, maybe this... Yeah, this one is. Yeah. So the nice part about this is like with just twenty minutes of data basically, because everything is super summarized, like like the... If if you have, like, this general, like, learning stack pipeline, however you wanna call it, where you go from dense video to just actions, in a way, we're, like, sort of move... We're shrinking the portion of policy, like, that policy learning is responsible for. Right? We're sort of, like, extract... We're expanding a lot of compute and, like, knowledge contained in, like, grounded vision models to simplify the learning problem. So, like, that kind of comes with some trade offs. But one of the benefits is that, you know, with just twenty minutes of data across four four different... Very different visual settings here, none of these are even in a lab. These these these three here. These sort of point information is robust across all these settings, which is why, you know, the robot's able to fold the towel as well. [00:24:38] S3: Yeah. What's the success rate on this, for example?
00:24:38 · Chris Paxton, Vincent Liu
[00:24:38] S3: Yeah. What's the success rate on this, for example? [00:24:41] S1: Everything was around... Let's see. What is it? 60 or 70%? I think 70 on average. And it definitely we noticed that, like, the performance of this of this sort of framework was more of a function of how good the three d measurements are, as opposed to how much data you have. And, like, it was relatively easy to get general behavior because you're able to do these augmentations, you know, to expand your domain input, basically. And also, like, these demonstrations aren't too complicated. It's not like there's many different ways to fold the towel that we've demonstrated. So we really saw that, you know, if you have, for example, HaMeR plus ARIA hand detection, about, like, centimeter of error, whatever. And then so you're therefore limited to sort of that level of precision. [00:25:33] S3: Yeah. Makes sense. [00:25:34] S1: Yeah. And then I guess another nice property, and this is what point policy shows, is that because, again, you're using points, which is just more like spatial geometric information, you can swap out the literal, like, training object, and you're able to, you know, fold a different color towel or, like, you know, just change the color, like, kind of variations of of your training object. [00:25:57] S3: Would it work with distractors? Like, you put another towel down, what would happen?
00:25:57 · Chris Paxton, Vincent Liu, Ademi Adeniji
[00:25:57] S3: Would it work with distractors? Like, you put another towel down, what would happen? [00:26:01] S1: That's a good point. I don't think we've really tried that. I think that my my solution might... If I had to work around that, I'd probably just engineer, like, some higher level thing that sort of sorts this out. [00:26:11] S3: Yeah. Just use like Sam or something. Yeah. That makes [00:26:13] S1: sense. Yeah. Exactly. Yeah. Exactly. [00:26:14] S2: I think the... We have something kind of similar with the fruit bowl experiment. [00:26:17] S1: Yes. That's right. So that's kind of what we do for the fruit bowl or the... Yeah. The... Put the fruit in the bowl. You sort of exploit the fact that you can sort of... Like these fruits are about the same size and so if you label like points on them, they sort of mean the same thing to the model. And therefore, like depending on how you prompt this SAM model or like the correspondence model, you're able to have it go target different fruits, even though sort of like we just have one model trained on pick and place or like pick and place of like different fruits. [00:26:48] S2: Yeah. So basically it's language conditioned, right? Like the segmentation and the actual inference time point selection is kind of guided by how you specify the task. But the actual model is sort of more modular in some sense because, like, the geometric reasoning is exactly the same.
00:26:48 · Ademi Adeniji, Vincent Liu, Chris Paxton
[00:26:48] S2: Yeah. So basically it's language conditioned, right? Like the segmentation and the actual inference time point selection is kind of guided by how you specify the task. But the actual model is sort of more modular in some sense because, like, the geometric reasoning is exactly the same. [00:27:09] S1: Yeah. So this is, like, benefits and... Or there's trade offs to this. Right? So another instance of why we have, like, error rates at all is that we noticed that the highest source of error comes from... Comes in inference from the correspondence model. So it's basically this, like, you just use some old stable diffusion model, you query it for a step, you take out, like, a feature map from its activations. And then you do this for your training image, and you do this for your inference image. And then you just sort of like try to match, like, the cosine similarity for each point that you want to map onto this. And so... [00:27:44] S3: Go ahead. Yeah.
00:27:44 · Chris Paxton, Vincent Liu
[00:27:44] S3: Go ahead. Yeah. [00:27:46] S1: Yeah. And so like another, like, side note is, you know, we collect data with these Aria glasses, which is fisheye, and they're kind of yellowy or, you know, a little look grainy looking. And in inference, we have the iPhone because we just wanna unproject with depth. And so diff is really not a strong model at all or like a algorithm at all. And we basically help it out by using grounding dyno to segment out the object. So this is what ADEMI means when it's language conditioned. Like, the system is sort of language conditioned. So we segment out the object with DINO, which works pretty well. And even when... Even, like, when you run diff on the segmentation, sometimes it gets the points all wrong. Right? So it's like jumbling up the correspondence information or basically the orientation of the object. Yeah. And then sometimes with the iPhone depth map, which is very low resolution, if you... If your points are near the edges of of the object, maybe the depth map flickers. And so you... All of a sudden, you're reading a point that's, like, 30 centimeters further. And so it's just like this kind of, like, system level thing, which is maybe in a way a drawback of having, like, a not fully end to end setting. [00:28:53] S3: Yeah. That makes sense. Well, you could always, like, train your own point correspondence model or something like that. I guess that obviously out of scope for this paper, but is that the... Do you think that's, like, the way forward to improve... To best improve performance then?
00:28:53 · Chris Paxton, Vincent Liu, Ademi Adeniji
[00:28:53] S3: Yeah. That makes sense. Well, you could always, like, train your own point correspondence model or something like that. I guess that obviously out of scope for this paper, but is that the... Do you think that's, like, the way forward to improve... To best improve performance then? [00:29:05] S1: I... Again, it's, like, one of those things where that might be optimizing or, like, trying to push a local optima. [00:29:11] S2: Sure. [00:29:11] S1: Yeah. Me too, if you wanna speak to... Like, we were thinking of, like, some extensions of this in a more foundation model setting, and we realized that correspondence is probably one of the things that had to go first. [00:29:19] S3: Like, to remove it, not to improve it. [00:29:21] S2: Yeah. Yeah. I totally agree. Yeah. [00:29:23] S3: I think [00:29:24] S2: ideally you break the correspondence completely by basically learning invariance in your training. So yeah, I think the next step here would be basically, can you basically not have the system rely on correspondence at all? And, still change points, but like break the strict requirement that like, you know, the first point corresponds to like, you know, something semantic. So instead your transformer is learning something more geometric and spatial as opposed to having to keep this like specific ordered set of points consistent.