EXCLUSIVE: Brett Adcock on Figure’s 8-Hour Autonomous Shift

Ajouté : 2026-05-15

[00:00:00]Hang on. Rex in the house.

[00:00:03]>> Uh oh.

[00:00:07]>> We have red [ __ ] in the house, ladies and gentlemen.

[00:00:10]>> All right. All right.

[00:00:12]>> Great to see you, Brett.

[00:00:14]>> Well, drop us bring us back in so you can order down in the bottom corner.

[00:00:21]>> Yes, absolutely. In a minute.

[00:00:24]>> What have you guys been doing?

[00:00:26]>> Well, we've been watching Pink Drive.

[00:00:27]Awesome.

[00:00:29]You're watching the robot.

[00:00:30]>> Yep. Absolutely. Good.

[00:00:34]>> And we're all gonna be So So yeah, we we've been watching it, Brett. Uh you as well. I mean, you see it all the time.

[00:00:40]So I imagine you've got more other other things to worry about than watching the bot for eight hours. Like you got a real job.

[00:00:48]>> We run like we run like we have uh many uh we run a bot just like in this use case in many many different places in the office like this.

[00:00:57]>> Yeah. Right.

[00:00:58]>> Yeah.

[00:00:59]>> Pretty crazy.

[00:00:59]>> We've been doing this for like for months. Like this is just uh um Yeah. We like we like it's I think it's it's in a good spot. I mean I think like the the Okay, so we're like the crazy thing for us is it's we're at a point where the the system is like really stable now. And uh meaning um one is like this is running Helix 2. It's basically running a full body uh like neural network from pixels all the way to actions. uh everything from like subtle footsteps to all the joints like commanded from helix. Uh we have to reason through pixels. So we have to be able to like see the world and understand what to do at like a very fast frequency. Uh we need to be able to push the packages down the conveyor system basically around like 3 seconds a package which at the human speed KPIs.

[00:01:48]We have to do all this while being at like 90% successful with barcodes being face down on the conveyor system. they get scanned and then ultimately the there's another label shot on top on the other side for the for the delivery truck and um and then we need we need like not only uh there but if we want to run like 247 we have to have robots communicating with each other um as one robot is low on state of charge in the conveyor it message a robot behind it to come in and take its place. So the while like while that's happening the robot's coming behind it waiting robot's backing off the conveyor and we just basically the robot just jumps in. So within 30 seconds, we're back to moving packages again with like basically like very minimal downtime. The robot then goes back and charges. Takes only about an hour to full charge. And then anytime the robot has any problems, they head to maintenance. And then we message the fleet to bring another robot in. So we have like a the state space is actually pretty like pretty large.

[00:02:40]>> And the robots are doing this all autonomously uh self-nworked with each other, messaging each other, and moving like basically moving robots around. I mean, it's relatively straightforward. It's like a simple pretty straightforward use case, but it's like there's a lot of different things that could go wrong uh with the robot and with other robots there that uh have to kind of be like, you know, autonomously, um reason through. So, um so anyway, like yeah, this is like it's awesome.

[00:03:06]It's it's great to see um and we're we're excited to see if we can pull forward on eight hours.

[00:03:11]>> So, so Brett, we noticed that there are three bots there instead of two.

[00:03:14]Obviously, for charging, you would only need two. is the third one there is kind of a fail safe in case one of them has a hardware failure or something like that.

[00:03:22]>> Yeah, we have um we have and then then you know worst case scenario if like if the the bots like have issues with hardware or even software and they need to like roll out. We could just we have hundreds of robots in the office. They can they'll just call robots in the fleet and they'll just come in and dock uh they can dock anywhere. And so, you know, in some way we could basically just run for months and months and be like subbing out different robots, robots having problems uh and you know, we'd be basically uh be able to put any robot into the use case at the office.

[00:03:54]>> So, in this demo you're running now, do you collect Nice to meet you by the way, are you collecting data and training in at the same time or is this purely a demo or kind of like >> By the way, u before you respond um Gustav Anderson is Dr. Gustav Anderson.

[00:04:09]He's a he's a hand surgeon in Sweden.

[00:04:11]So, um >> uh great guy all over the horizon to do a lot of uh >> pick apart your hand audience.

[00:04:19]>> Yeah. Yeah. Yeah. Yeah.

[00:04:20]>> Um >> over to you.

[00:04:22]>> Well, you guys want to see the You guys want to see the robot? We go. You guys want to check it out?

[00:04:26]>> Awesome. That would be awesome. Yeah.

[00:04:29]>> Answer questions.

[00:04:31]>> All right. Hold on.

[00:04:33]>> Ultimate Snory Cam.

[00:04:37]Awesome.

[00:04:39]>> This is like Christmas.

[00:04:40]>> I know. Seriously.

[00:04:45]>> There we go.

[00:04:46]>> Oh gosh. So, there he is. Oh, we got we got a different point of view than everybody else does. Yay.

[00:04:50]>> That's right. That's right.

[00:04:52]>> Wow. That's fantastic.

[00:04:55]>> So, Brett, was he was he primarily RL trained or was he primarily um Yeah, there you go.

[00:05:06]If he if he say hi Brett, will the bot respond?

[00:05:11]>> I don't know if I wonder if he cannot hear us because he's Yeah, it might be too loud.

[00:05:15]>> All right. What do you think?

[00:05:18]>> That's pretty cool.

[00:05:19]>> Amazing. Yeah, that's really cool.

[00:05:24]>> And of course, you've got the others.

[00:05:25]Yeah. Right.

[00:05:27]>> Hey Brad, it's Phil Trouy here. Um, there was a question actually, Dr. Noah, I'll just ask.

[00:05:33]How how was this bot trained to do this task? Can can you kind of tell us just even in general like what kind of techniques he used?

[00:05:41]>> Yeah. Um so we trained this end to end with Helix 2. Uh there's some like there's some good information actually on our site. Uh okay we've done on on both this use case and other stuff.

[00:05:50]Helix 2 is the is a basically a full body uh like basically neural network policy we've designed in house. Uh it's a vision language action policy and um this uh we we've trained it um the data set is basically like um uh we basically we've we've we've now um collected this data uh done training and then this is all test time inference happening here.

[00:06:14]Um so uh the the robot is now reasoning through a basically a whole body controller. So it's um we basically from from just like vision and state The robot can command things like subtle footsteps like the robot can take like you know lunge and do footsteps all coming from like basically like basically camera conditioning vision conditioning.

[00:06:35]>> Um so uh so just through like looking around the robot's been able then to command all joints uh simultaneously on how to move the body fingers pelvis footsteps everything uh fully end to end uh all that uh AI work is happening on board um on on in actually in the robot torso. the brains are in the uh torso chest right >> of the robot and uh does all that locally on board with no there's like no network needed to do any AI inference >> yeah but how how is how is the training data captured was it like uh somebody wearing a a GoPro headset and and doing these motions because somehow the the the the robot has to learn these motions and and I'm just wondering if that was through u through uh video or from um teleop or combination.

[00:07:23]>> Yeah.

[00:07:25]>> Yeah. I don't want to specifically comment on exactly what we've done uh as like a recipe, but like you know, we've we've collected everything from like tele operation data all the way through uh you know, human video data to like online online video work and uh we've we've we've been training and we've been rolling out policies on all all of those areas. Uh ultimately we need to be able to >> you know we we kind of see the world like a human and we kind of move the around the world from like like a human.

[00:07:51]So ultimately the the the best and most uh like like the most ubiquitous and largest data set in the world is human data. So ultimately um in the limit like figure needs to learn from humans uh as much as possible. I mean we're like we're a humanoid, right? We have this unfair advantage. or like we have five fingers, wrists, pelvis like uh you know cameras on the head that like see an RGB space you know passively through the world like we should be we we need to be like uh we need to be trained on human data scale >> and so actually speaking of that can I ask you because these figures have the cameras in the hands is that correct these ones do or not?

[00:08:26]>> So are are you able to use that? Are you able to sort of go into cheat mode with that and use the cameras in the hands?

[00:08:32]It doesn't look like it right now but maybe they are. Um, I don't know if I actually I don't know if this this policy h is using cameras in the hands or not in this exact one, but we do use like cameras in the hands quite a lot for all of our stuff. I wouldn't be surprised if the cameras in the hands are being used for I just actually don't know top of my head, but like uh >> we we do have a camera in every figure three hand as of now that's like that's operational. Um, and we have seen a lot of benefits of of doing that. I like you know there's a lot of areas where you're like uh that like certain head cameras really can't see into could be like bins and boxes and then um and then you know for like uh a lot of cases we're like grabbing items in front of the head we're like partially oluded from uh and then uh so having like a we've seen a lot of benefits from having camera in the hands here.

[00:09:16]>> Cool. Cool.

[00:09:17]>> Yeah, that's that's what we're wondering is is the camera with the hands could scan to see where the barcode is.

[00:09:23]>> So if it's to run behind it, it it would be able to do that. Okay. We just didn't know if we was doing this case.

[00:09:27]>> In this use case, we don't need to scan the barcode. We just need to know where the barcode's at. Like the >> Yeah. to see it. To see it.

[00:09:32]>> Yeah. Just for like for the for the for the users watching, the use case is the uh find find the package uh figure out where the barcode's at. Put the barcode down and then put on the conveyor. The conveyor scans the barcode from underneath. And then um a few seconds later, there's a pneumatic label machine that basically prints out a new label really fast and pneumatically throws it on top of the of the package. uh the other side and it's used for like the the delivery trucks and elsewhere. So um and so so you know some people ask like why not you know put the put the barcodes up or like uh why not why not be able to scan them like our job in this case is not to do this. It's just it's basically a small package like sorting line here.

[00:10:13]>> Okay. Okay where it is as well. It looks at the box and it doesn't see the label but it knows that the label is on the back so it flips it correctly. or does it has to see the label to kind of infer it or can it >> Yeah, I was watching a little bit today.

[00:10:27]There's like a lot of um there's a lot of good times where like label like it'll be like a say like a cube um a square box and the label will be on the bottom and if it if it determines like a uh has never seen a label on either of the four sides or on top, it'll determine it's in the bottom and just not even look and place it on and it's usually fairly accurate.

[00:10:43]>> Yeah, I spotted that. That that was that really stood out. I spotted that. Yeah.

[00:10:47]Yeah.

[00:10:47]>> Yeah. So, like the cube boxes are funny.

[00:10:49]Like it holds it like a Rubik's cube and I can move it around.

[00:10:54]>> You didn't have that last year, right? I think the the boxes were more rectangular last year. I don't remember any cubes last year.

[00:11:00]>> Cubes as well.

[00:11:01]>> So, it looks like it's it's it's a new one that you may have thrown in because again, it seems to be stuttering a little bit on that because it's not quite sure where the label is. Where it's a flat box, it's pretty obvious.

[00:11:11]It's a package. It's much quicker.

[00:11:13]>> These are all like these are all just data problems. like the the faster speed is has been like there's some architecture decisions that we've made on the helix side to definitely make it faster.

[00:11:21]>> So you have like that side of things are getting better and then you have like data like like better data better high quality data more diverse data flowing into Helix for for training like that's all you we've done some abdations on the website you can go look at where we've um more data is increasing success rate and also decreasing speed. Uh so I think if you went back to the figure two work we did um on some logistics posts last year, we had some uh stats on the site where we showed um both speed and accuracy improving with more data. Uh so like like largely the box or anytime you see any issues, these are all data issues at this point. If you want this use case to run faster and higher success rate and um you know every once in a while like it'll it'll have a you know the barcode on the wrong side.

[00:12:03]These are all just data problems. Um, >> did you did did you upgrade the uh Nvidia processor between Figure 2 and Figure 3?

[00:12:11]>> I don't think so. No.

[00:12:13]>> Okay. And are are you looking forward to to to an upgrade to a processor upgrade presumably in the future?

[00:12:20]>> Well, I we talked about we talked about this earlier. We were wondering if you're if you're running up into the limits of of your current inference processor. That's what we're wondering.

[00:12:30]>> Um, yeah, we like we uh Yeah, listen.

[00:12:32]Would we love to have like more memory and more flops on device? Like the answer is like we would [ __ ] love that. Like uh we are um yeah we are like largely memory constrained uh and uh on device for for inference today. Um we do have a planned upgrade coming uh on the robots for future generation to help with this. Um but as you can see like having ondevice uh like models that are running is like super critical. um you know there's a bunch of advantages.

[00:13:03]One's speed and you know we have like such fast dynamics that are happening from the control side uh that you know like so having ondevice models is important there and obviously like you know being able to operate in an area where you don't have a network or maybe even like a you know like a high latency network is like uh is extremely beneficial to have on on ondevice models. So I think you know in the limit you're getting better more you know more memory more flops you're getting bigger models and the robots are getting both smarter and faster uh with ondevice uh ondevice stuff.

[00:13:35]>> So do you see do you see a future where like you know in in the future where you have a hundred different t you know completely unrelated tasks that you that your bot can do. Do you see a future where you'll be downloading where the bot will download specific models for a specific task like an app store for for figure kind of thing?

[00:13:55]>> No, I think the I don't I think the it's like um the one beauty of having a humanoid robot is you can basically benefit from transfer learning. You can have a robot that can have like one single pool data, one single ideally one single neural network that is pulling in all this like very diverse uh different types of use case work across like a say say a humanoid hardware and you're basically learning from this as a as a pool like like you do in pre-training for like text LLMs and uh this is what you don't get from like having multiple different robot form factors is you don't get the benefit of being able to uh like like uh you don't get the benefit of transfer learning from other things. This is this is a this is a humanoid one. Like one of the biggest benefits to humanoid is like you can basically have like one AI brain. You can have data flowing in from all these different use cases and it's bettering the entire system. And I I'll tell you a story. We were doing um we were doing some work on figure 2 about a year and a half ago where we were like running uh uh Helix our first version uh of you know of Helix Helix one and we were I think putting some items into a fridge and the policy was generating like um like 60% like 55 60% success rate of putting things in the fridge like just you know different condiments like uh like cheese other things like in the fridge like in the fridge drawer and it was just with fridge data it was just trained on Helix on I think it was like 55 or 60%. I remember us like 3xing the amount of data that was like nothing to do with a fridge. It was like opening cabinets and drawers and like things on a tabletop and training like a you know training basically the model again with like more diverse data that was like fridge plus other things. And that same model I watched basically do in a limited eval like the next like the next day like basically do 90% in the fridge work. So we went from like 60% success rate with fridge only data and then we trained that same model again with more data like that's not fridge plus non-f fridge not even increasing the fridge data and the numbers went up and that's when we were like holy cow like the this there's a real uh benefit to having like extremely diverse like data and like it makes sense because like the more we handle like bags in other areas will like will also help us like handle the cheese bag better as we move to the fridge. Um, and so like this is a huge benefit. This means like as you pull more diverse data in across different use cases in the robot, the robot's getting smarter and better. We we saw this on figure three, we've seen it on figure two. This is a massive phenomena.

[00:16:21]It's like it's just like a miracle for humanoids where you basically have a single platform that can go out and basically do almost anything a human can. And you're getting better collectively across that fleet as more data pulls in to the robot. So instead of having like these specialized models you're pulling out like um like an app store uh you're basically pulling in all this data into one centralized model uh into training and then you're training a big model uh offline model and then you're basically like like otaa and you're like over the air updating all the weights uh that are on on every single robot. So basically the like unfair advantage is like a humanoid that gets good at one thing say logistics every robot in the fleet will get better at logistics the same same level of performance >> uh which is unlike humans humans you know I'm watching my kids they all have to like learn independently and they they want to learn independently they don't want to help um and they're all like you know at their own pace that's not the same for humanoids there'll be a collective learning through the whole system centralized through a central brain that uh basically gets shared across the whole network of the fleet So I have a couple of questions um and I I want to go around the panel and uh because Brett I'm sure your time is limited so I want to give everybody an opportunity especially Gustav for the hand because he's our good hand doctor and Devong welcome Humaloid Hub um think of a question that you want to ask um Brett but I have two for you first of all Helix AI is phenomenal um the is it the two S2 layer that you guys introduced recently that brings in the more granular role. How do you see that AI stack evolving uh into a more kind of intuitive uh feature-based evolution? A and B, you've you've teased us with a lot of not a lot, but just a few bits and pieces of information about figure 04.

[00:18:07]Um between generations, what do you see evolving faster, the hardware or the AI stack? And what's the SL chips? And what's the sort of incremental jump that you're expecting to get in features and capabilities in generations?

[00:18:25]And then I have another one about the China supply chains.

[00:18:29]>> Um I think uh we we just you know yesterday we uh we just we just finished our critical design review for figure four. So it's our last major design gate for uh like as we like enter design lock and ship parts out.

[00:18:44]Um and um I I will say this I can't you know I don't want to talk about any I don't want to like you know this isn't a product launch for figure four but I'll tell you like you know we have a you know we have a spot in our lab you come to and you can see like every version say figure one I mean Scott you've seen this figure two figure three and you can see this huge capability jump from like figure one to figure two. It's like figure one's like a you know like a you know dorm room prototype it looks like compared to figure two. Figure 2 is like nice, little big, uh a little heavy, and then like figure three got like slimmer, like less mass, better sensors, better hands. Um, and you know, you saw this like step up every single time between models. Um, figure four will be the largest step up we've ever made between versions in history. Um, and I wouldn't have believed this a year ago, like because when we ship figure when you ship figure when you ship when you're when new for the like last design gate for figure three, you think like you have everything you'd want in there and you're like so happy about it. Yeah, you're over the moon. And then, you know, in the last year, as we've gone, you know, as we've like, you know, designed Figure 3, we've gotten out, we've learned, um, and we've also progressed on Helix, we've taken a step back and we've been able to think about if we had to redo this from scratch again, what it would look like. I think figure four will be the first like I it'd be like an iPhone one moment for the space.

[00:19:58]>> It is so dramatically different on both on every aspect. Um, it's it's unrecognizable as a robot like from like the previous versions we have. Um, and I think it's the first version of like a a like a true like architecture that will scale into the millions of robots and um we've learned a lot. I think like I think if I had to put if I if I had to compress every my knowledge into like a single couple sentences, we've designed uh figure 4 for helix helix 3 and helix 3 was designed for for data and the whole robot was designed around data which is our like the largest limiting factor by far for figure to get into like uh robots at mass scale to the world. Um, so you know, I I don't know if that helps, but like I think um, you know, figure 3 is probably I think arguably one of the best maybe the best humanoid in the world, I think, you know, by a long shot. I think Figure 4 is just like completely out of this world. I can't um, you know, we Yeah, we don't we don't like we don't show anybody anymore. We don't like show basically anything. It's in a a secret room here and only a few folks in the whole company really understand it well.

[00:21:11]But like my god, it is it is it crazy.

[00:21:13]And I'm like I'm excited. There's a lot of work left to get to get it ready ready to go, but man is it is it going to be uh the the you know, sorry I'm a little longwinded in my answer. I always kind of thought humanoids would like level out to saturate like iPhones did at one point like iPhone 13 or 15, you know, like nobody really cares anymore.

[00:21:28]You can't really tell. Um, we are so early with humanoids. We're like I think you know we were in flip phones and Nokias and you know I think maybe figure four would be the first version like iPhone one level moment where the architecture is just perfect and yeah things will get better as we go to iPhone 2 and three and four and you'll see those leaps but like we made a big leap forward here.

[00:21:49]>> We're laughing because we were talking about black we were talking iPhones.

[00:21:53]>> Yeah, we were just talking about that.

[00:21:55]>> We were just talking. So Brett, um first of all, I have to um just an observation, but I I've watched you over the years from afar, uh and I've admired how you've handled every question that seeks to compare you to the competition and um Tesla optimists. Um and it's I have a lot of admiration for the way that you focused on what's right in front of you um and not really spoken or talked down about the competition. So a lot of respect for that and your attitude and I think that trickles down to your entire team. Um having said that every humanoid bot company in America and every non-Chinese humanoid bot company is deeply reliant on Chinese supply chains. I'm curious how you're thinking about maybe the whole process of onshoring and insulating yourself from geopolitical risks that may impact your supply chains in the future and how much are you thinking about vertically bringing everything inhouse?

[00:23:05]Um so we so I think maybe maybe for folks in the audience maybe listening like so I figure we design almost everything in the whole robot from scratch. Um we design the the motors the the rotor stator the we design the gearboxes we design we design gear gear boxes here.

[00:23:23]We make them here. We make gearboxes here on site too which is crazy. Um uh and um we we make we make sensors uh we design sensors um uh like we designed you know h there's over 100 PCBs in the robot we do all that work here internally on the electrical engineering side um cameras are designed we design the cameras here from figure three. So like all the hands like just like everything is just like a tremendous amount of complexity all the way down the stack. Um we don't like we don't actually manufacture every single part obviously but we design almost everything. Uh and then some of that we actually manufacture individual parts and then we >> Does that frustrate you that you have to go to China for these parts or other parts?

[00:24:04]>> I think I need to know the exact numbers but I I my forecast is maybe by next quarter I don't think we'll have anything coming out of China on the supply chain side.

[00:24:13]>> Wow. Really? We've moved most of our supply chain like in ter like outside of China the last like year uh for tariffs, geopolitical risks, everything else. So we don't we have very little uh manufacturing uh on individual individual parts coming out of China these days like very very low like uh low enough where like there's no risk really anymore from us from that perspective. Um >> wow.

[00:24:37]>> I think every major major company in the world has been doing that last year. So I don't think it's uh I I think every there's not a really a major group I don't think with like hu like huge exposure there uh at this point. So um I mean there may be some but like anyway we we've made a a concerted effort to to move to derisk that significantly last like year.

[00:24:54]>> All right. I'm going to give Devong and uh Gustav a chance to ask a question or two.

[00:25:01]>> You want to come first?

[00:25:03]>> Uh sure. Hey thanks and congrats Brett.

[00:25:07]Uh the the demo looks super impressive and the the speed improve uh improvements have been noticeable with this one. Uh my question was um the the AI models that uh humanoids use are very data hungry and I'm assuming from different data modalities you are working on gathering uh useful data. Um so when you look at one year out do you think uh you you would still be data scars for some of the important data or will you would you be compute uh constrainted um it's a good question I mean there's a lot of constraints right we need to make a lot of robots there's a constraint there we need a lot of compute there's a constraint there we need to both train and inference compute um and then we need a lot of data. Um this is I mean a hell of a place to be in where we have these constraints now. We're not asking like will humanoids work and can they do useful work and can they run for five or eight hours or longer like these are great things to be uh you know to be thinking through. Um I think like you know I'll answer the question is like certainly right now like our biggest constraint is data.

[00:26:25]Like the robot if we had the right data we could snap our fingers I think general robotics would be solved today.

[00:26:32]I think we'd be able to do most things a human can in the world. Um and then you're probably like energy compute and manufacturing constrained by like like by a long shot. Um maybe in the orders of like data constrained then compute constrained then manufacturing constrained. Maybe that's like the right way to think think about this a bit. Um and then long long term again you're you're computer constrained again because you're like there's so many units out the market that you're you're running in front with. Um, so I think like right now what we need to solve like what we care about I figure is solving for a general purpose machine that has like common sense raising like a human that if we if we want to like if we want to solve if if we want to just like show you what we're doing now and just like ship that thing and get thousands and get tens of thousands or hundred thousands of robots out like we can do this like we we know the playbook we just got to go execute here. Um but if you want to solve like for like a you know the experience you feel with like a human in a suit or meaning other humans that that we're data constrained by like orders of magnitude like we just need so much of the right data out there um this is like it's like not it's like um that that needs to be solved like before we like talk about how we're computed uh internally here we're not compute constrained we have like a great relationship with Nvidia um and they've supplied us we just launched a new cluster of B200s that we announced a few weeks back that just went This month uh we're training uh into pre-training some of the largest models we've ever done. I think last week we trained like we started pre-training like one of the largest models we've ever trained by like several factors. Um these are large models now that take like they take time to train on like very large clusters uh very large cluster for us. Um so you know right now we're not like internally at figure trying we're you know like we're not we're not hoping that we would snap our fingers and get more compute.

[00:28:22]Obviously be great to have more compute, but like we have all the comput we need to train the models that we're that we're designing for Helix now, which is good. And then over time, yeah, having access to more compute would be great.

[00:28:31]Uh we we're planning on that. And we have the right partner and partners in place to do this. Uh we think today. So I think we're we're okay. We got like we got to go knock this thing down, man. We got to go like solve general robotics.

[00:28:42]We got to build like why we're we're all on this call hoping will happen, which is like we build like a human in a body suit that can like I can talk to and just do anything.

[00:28:50]>> Yeah. Speaking of which, u I know Gustav has a couple of technical questions about the hands for you. Um as as as much as you're willing to share, of course, but hey, when can we get one for ourselves?

[00:29:02]Um yeah, I mean, we've been like I mean, there's different models, right?

[00:29:09]You can take there's other groups that will just like sell your robots they day one, they don't, you know, maybe they work fine, maybe they don't. Like that's we're not really in here to do this.

[00:29:16]We're in here to ship like an Apple style quality product. like we want to ship a product that works and we don't want to ship something that isn't going to work to you guys' house.

[00:29:24]>> We don't want to ship this in the commercial customers that don't work.

[00:29:26]This stuff needs to really work well.

[00:29:27]So, we're like we we don't want to ship crap >> or slot and so um we're and we're not going to be impetuous about this decision. So, we're going to we're going to try to you know, Figure is going to try to ship an unbelievable product.

[00:29:39]>> And this is such a hard space to be in.

[00:29:42]Uh >> I I have to ask you because we've talked about that demo of a figure in your house with your kids. Uh we've we've talked about it so many times here on over the Rise of Scott and yourself and I was that was super impressive. What were your feelings when what were what were you feeling when the pot was right there next to your kids and obviously you felt that it was safe enough to do that? I mean you wouldn't ever take a risk, right? But I mean did it give you a kind of a view of of the future, the potential?

[00:30:12]>> Oh yeah, of course. Uh I mean I'm listen guys, I'm around here every day. Like I'm seven days a week. I'm here every night till midnight. I see these bots all like like I I see my bots, you know, like more than I see my family. I'm like at work a lot, you know. So, these are just like I um uh I don't know. I'm around them all the time. I I can't even imagine the last time we had a bad failure with a robot. I don't like, you know, I don't I haven't been around a robot at an office when I'm around all like all of them all the time where we had this. It's uh used to happen every day or almost every hour uh years ago and now we're in a place where the robots live.

[00:30:46]Um, it's good to do testing. Like we it was in my house that was running autonomous policy doing I think it was in this case uh putting like like laundry in the wash machine. Um, and uh it was doing great. Like super proud, but like you know the bar's here, we're here. Uh, we need to design something that can like do everyday things all day every day. Like you know stuff we're doing now, but like do it in your home, in anybody's home 247. Like that's like that's where we want to head. So we just like we feel like we're far away from the, you know, we're far away from that today. Like we're like we're not there yet. So we're like we're dying to get there and that's a data problem you know we just talked about. So like we're it's good to like it feels great. We're like super pumped to have robots there doing that and feel like a you know like um feel very fortunate to work on the problem but like we want to solve this problem. We are going to feel like uh it's going to be embarrassing if we don't you know we're we're ready. We got the cash. We got a great team. We're working working our asses off. Like we got to go solve for a general purpose machine. And that that that's that's where that's where we're pushing 80 100 hour weeks here. uh is just like how do we go solve that problem?

[00:31:46]>> Yeah. I mean it's clearly evident because I mean the fact that you were willing to pick up the gauntlet and you were confident enough the fact that you're willing to just laid out their warts and all is just mad respect for that and so appreciative of that because I mean the entire audience and the entire community around the world is is learning so much about this and we also need to sensitize people um and not only just make them aware but just get them you know mentally ready for this new era. So, I mean it's it's it's I I cannot just tell you how how crazy good and you know awesome this is. So, I'll hand it over to Gustav now who I'm sure has a couple of technical questions being the hand surgeon.

[00:32:28]>> Well, I don't know if they're technical but I mean so your body is doing useful work. Why would you need to redesign the hand? Aren't isn't this good enough? I mean >> Yeah. Um we we have done more hands than robots here. Um you know we started as like a t we I I started like thinking that the the right hand was like a tendon based hand. Uh very very similar to um to how humans are today where most of your actuators are actually in your forearm and they're being you know um mechanically driven from like tendons.

[00:33:09]Uh we designed and shipped that for figure one. It's in our it's in our uh design studio. Um and I I was just so proud of this. The hand was like so incredible. Very high degree of freedom.

[00:33:19]We customized all the actuators for the wrist and all the fingers and the forearm. We like custom designed all the all like basically the whole tendon system. Uh I personally was there when we brought it up. Uh and I spent about a month with it and I ultimately killed the program.

[00:33:34]And this was in 2022 2023.

[00:33:38]And um and since then we've now like worked on five or six new generations of hands that some of you seen, some of you haven't maybe seen. Um tendance is for sure the wrong way to go with hands.

[00:33:51]It's u it's a local maximum and it uh it it is not the right engineering design long term for humanoids.

[00:34:01]>> And I learned that firsthand and as like you know was an engineering mistake I've made. I mean, we learned a lot, so I don't know if I'm a mistake, but like uh um and since then, we've now designed kind of better and better hands last like four years. We have a new generation hand that we teased about a month ago that has like basically equivalent level of human of human range of motion and uh like KPIs as a human.

[00:34:26]And it is just it's bananas. I don't know. There's not a single hand in the world that's even close. Not even close.

[00:34:32]Like in terms of like the the reliability of that system, the the kinematics like um the the KPIs in your own speed and torque. Um we have like like some of the greatest sensors in inside the robot for for like for tactile. Uh the rob the robot hand is just absolutely insane. Thermals are insane in the hand. Uh yeah, and we we'll talk more about this obviously as we as we go on. Um I think to achieve general purposeness in the robot the a human hand a human hand at par a robot hand at parody with a human hand is is uh is critical >> and it seems like that's where most people are struggling at the moment as well and I mean you manage a lot with just having the pure flexion extension but in the demo you kind of the >> I mean let me give you like >> do you really need it you think We we think you do. Like I think if you you know when I when I grew up my mom my mom would always like fold socks like this inside of each other and um our current hand on figure three can't do that.

[00:35:41]>> It doesn't have the degrees of freedom.

[00:35:43]Um our new generation hand can >> that we'll put out with figure 4. And if our goal is to if we go if we go back to like talking about the earlier conversation around like what is the constraint if constraints data and we want to learn from human uh data at scale and we have human data that can do that stuff and we're not able to do that in the robot. This is data that will uh a we're not able to learn from it and b it will pollute our data set.

[00:36:11]>> Yeah.

[00:36:13]So, I would argue that in order to get to humanlike intelligence, having a hand at par, a robot hand at parody with human hands is fundamental.

[00:36:24]>> Yeah, I think she Yeah, I totally agree.

[00:36:26]I mean, it's it's what I'm looking most forward to is having the the humanoid bot with a hand that is equivalent to ours. And I think we're getting there, but there, okay, you may might not want to tell, but how do you solve for the like packing everything into the hand and the heat? Is it all did you have to kind of make up come up with all yourself or are there products out there that can kind of do that? Today, >> we we designed this whole hand and end ourselves.

[00:36:56]Um, so like I just left the hand review for figure four and it was like maybe a hundred slides on thermals, electrical sensing, kinematics, uh, actuation like uh, structure uh, all of it. And it was just um, many many many people working over a long period of time doing some of the best engineering work I've ever seen in my whole career.

[00:37:22]So uh this is like cumulatively like years of work that we did just for this hand that is just too hard to explain like in you know how do we sell the thermals here or there it's just it's just uh you know we worked on every millimeter every gram uh we worked towards like very hard requirements in terms of KPIs for this hand and we like we nailed them all. So like uh it just like the robot hand itself now is you know the robot hands for us have more actuators now than the rest of the body. It's just it's uh it's the most complex thing we have on the robot and um you know it's like it's hard to explain how we design a humanoid robot here in like few minutes right it just uh I you know I left a 10-hour meeting yesterday just talking about this in detail >> and you Yeah. So we only >> we do this. It's like uh we spent like dozens of engineers for four years have been working on this problem and it's like uh culminated to like you know this like very it looks like it looks like um it's like watchmaking man. These these parts are small super compact. There's no space. It's very it gets hot. It's um you know there's a lot of motors there.

[00:38:32]It's just this it's like a lot of electronics wiring. Things are bending trying to come loose. uh stuff trying to get in there like dirt and debris and water and so it's just it's a very very significant system engineering problem that we had to go through.

[00:38:47]>> Brent, do you think you'll ever reach a point of matching or catching up with evolution or do you need to or can you even you know go past evolution? I mean you don't need to be limited by the human form factor in terms of functionality. Right. I'm a big believer of like the of not of of uh there's there's a there's a there's a there's a school of thought out there is like we'll just go beyond the human form and get to beyond like go beyond human performance. And I'm just like a um I I just don't believe that personally.

[00:39:21]We've um like a simple thing is like you could think of the world as like eight billion humans got dropped on here today and then we work to go build a world around it so we can work with it. We get shelter and we get doors and we have tools that we can use. Like we like built the world for our bi. You can think of this like we didn't get to choose the our body composition, but we did get to choose what we built the world uh world of. And we went out and used the the way we look biologically to go build a world that we can interact with. We live in a human operating system in the physical world. And we built that human operating system so we can interact with it every single day.

[00:39:56]There is no greater like general purpose machine than a human just because of that very like ve like very reason.

[00:40:02]We've designed it just for us. And you know the you take like some you know little bit more like uh aggressive view at like if if I was too short or too tall or if I had you know one arm and not two like you just you're you're not able to do as much things as a normal human can. And uh it's because we like a normal human, you know, 5 foot whatever has designed the world around ourselves so we can interact with it every single day. So there is no getting better. Like you know what I mean? Like it's just like it is we've we've designed it specifically for a 10-fingered human that's roughly 5'2, 5 foot five that can like do everything in the world. That's the way the world was built. It was optimized perfectly for that across a little bit different form factors here and there, but mostly an average human.

[00:40:46]And if you're too short, you're two feet tall, you can't reach up on the cupboard and you can't reach up on the cabinets.

[00:40:50]And if you're too too tall, like if you're 20 feet tall, it obviously won't be able to work well in the world either. So it's like we've optimized the whole world just for humans. It's perfectly built for a humanoid, like a mechanical human. And it's a perfect key to a keyhole you could think of, >> right?

[00:41:05]>> And that's why a humanoid will just build like something magical is because you can build one hardware with one AI model that can take all the data in like a sponge.

[00:41:14]>> Yeah. can get better as more data comes in that sponge.

[00:41:17]>> Yeah.

[00:41:17]>> And the hardware doesn't need to change for that. And that's why human >> Roden, can I jump in on this because actually this was this was actually related to my question. I actually wanted to talk to you about the bedroom um making thing that you guys released a couple of days ago where the two bots are working together because that is um actually Scott and I did a video on that a couple of days ago and and I thought the the fascinating part about this was it's clear that with the head nods and the the visual communication, not just like Wi-Fi telepathic communication, that you're kind of building out a world for intelligences, not for the robot.

[00:41:50]It's like the robots have to interact with humans. They have to interact with each other. They may have to interact with other forms of intelligence like large language models or something.

[00:41:58]System two thinking, the stuff that you're doing, you know, kind of outside the box. So, I I'm curious what your thought about that is because I find it really Yeah, there we go. I find it really fascinating. The head nods in particular, I thought were really really remarkable. And then um this is related to this in a slight way, but I was curious. You've got perfect lighting in this scene. Do you guys mess around with the lighting some to try to get imperfect lighting so that these guys have to work in more natural environments where you might have slashing lights and things like that?

[00:42:25]So, two two things about the bedroom.

[00:42:29]Um I think one thing about a humanoid is like you're basically humanoids are using intelligence to basically like create output >> and like economic value and um you know in this case like it's doing you know it's like cleaning up a room making a bed in the case of you know the live stream today it's like you know moving packages around. Um, but you can think of the humanoid as like just basically like a uh you like basically like I mean it's basically a robot is just turning intelligence into useful output.

[00:43:04]And um I mean this is something I want in my my my house every single day. Uh yeah, if like if lighting does change significantly, we we've seen policy degragation for sure. Um this is where like data it's just a you know there's maybe um maybe maybe there's a way you can train through it. Um there's other like you know different types of techniques you can use to you know uh change the training sets to have like different type lighting conditions and doing some alterations there. Uh it's also just a just another like answer for like we just need more we just need more data like in the system of like places of like varying lighting conditions across the world and all this. Um so uh yeah I think the headnot was cool like um in this case they needed to like tension the bed at a certain time so one's not pulling and the robot's not ready. Uh, and they use a headnot as a gesture of communication here that's happening, which is like really cool.

[00:43:54]Um, as a way to kind of communicate with each other. Um, we we think, >> but do they need to? I mean, I it kind of for me it looks as if the head nod is more for humans. I mean, as a signal uh rather than because you they don't really need to give a a physical kind of a a nod or any indication because they're they're connected to each other, right? It's like a hive mind.

[00:44:18]Well, in this case, like there's no um like explicit messaging between the robots. They like they're coordinating their actions fully visually with the head nods. So, they're not that's how they're communicating. Um there's other ways to do this like um in you know um that's how the robots are performing here. Um >> but yeah, I don't know. I think this is like um it's cool to see robots to be able to do things like this with a with an AI policy and it's cool to be able to show like multiple robots being able to engage with each other. And we're doing this on live stream right now. We'll have like robots, you know, like taking turns, you know, throughout the shifts, um, and communicating with each other, uh, and messaging each other, which is very cool.

[00:44:56]>> Now, now I assume it's also so it can communicate with a human because I know you want to sell me two. I I understand actually you want to sell me three bots.

[00:45:03]Um, but I can only afford one. So, it' be interesting to actually see a human on the other side to see how they're able to because I assume you that's just enough. So, it will it'll read my head now.

[00:45:13]>> Yeah. Yeah. We're spend a lot of time on it right now. almost like how to do like effective speech to speech with humans.

[00:45:19]Um, you know, actually the figure robots today are using the HARK voice model that we designed. Um, >> cool.

[00:45:26]>> To to to do this, which is really cool.

[00:45:27]Um, so if you go come in here and talk to the robots here, they're using the HARK voice model. And um, >> yeah. And then here you go. There's the robot. Man, that thing is >> first one is back.

[00:45:39]>> So I've got two questions about this customer. Uh, one you said like 90% is the rate. So that's what you have to do and that's something I assume the customer set.

[00:45:48]>> Yeah, we basically need to do like roughly 90% success rate of the of barcode scans. We track this internally.

[00:45:53]It's not on the live stream of like you know out of the 5700 59 or whatever packages it is now like what percentage of those got scanned. Um well so we have it internally it's like 90 something percent. We have to do 90% success rate there and we have to do roughly three seconds a package.

[00:46:08]>> Um is the rough the rough uh you know KPIs.

[00:46:12]>> Yeah. And and the other thing is there's no rails on the conveyor. Is that because that's the customer setup?

[00:46:18]>> This is the customer setup.

[00:46:19]>> Oh, okay. Because we're wondering why isn't the why don't the customers put rails on there just to keep packages from falling off?

[00:46:25]>> This is like this is we we like from we are trying to reproduce the exact use cases we see at real real sites >> and um so like one for one it doesn't help us designing systems here that don't look like the customer sites and things like this. So uh all of our stuff we do we try to get it close as possible and you know then you know this point it's just like how do we get the robot to a point where it's it's running nominally without any failures >> right and and the final comment about uh Vulcan when when it goes off to repairs >> um why don't you call it I've got two names that you have to use one of them I I think is that you can either call it the bots >> or you can call it bash Bash where Bash is the bot ailment specialty hospital.

[00:47:15]>> My god. Um like Morris who led the project named it uh >> no Vulcan Vulcan is fine but where it goes I mean the actual location should be off to the bots.

[00:47:26]>> Yeah, bots might be good. We'll I'll I'll uh I'll put it in the idea jar over here.

[00:47:31]>> Okay. Okay. That and Bash those those are the two.

[00:47:34]>> Okay. Were you worried about the data transfer when you if you're designing a new hand now and you have all this data with the older version of the hand and I guess they go for the whole bot but focusing on hand do you think it would be a big problem to kind of just transfer it? Do you have to get a whole new data set to practice with or do you think it would be quite easy to just go on with a hired off hand?

[00:47:59]I actually think it'll be much easier because we're trying to learn from human data at scale and like our current hand is not doesn't have the same kinematic mapping as a human and it's there's um the way the fingers been and everything and the way the linkages are set up in the fingers now are uh are not ideal for this. So I actually think um from what I've you know from the work we've done internally almost on every metric the hand's gonna be better >> like uh like in terms of like I think the hand even though it'll be like crazy uh humanlike like like I've seen it take a beating internally we've built them and um they're extreme the reliability is just crazy and um so almost every and then from like AI learning perspective all of it I think it'll be I I think it'll be just um one of the best decisions we've ever done. But like you know going to this from like if four years ago if we tried this I don't think we'd be able to be successful. Like we've you know we put our top motor specialists on it. We took our top sensor person on it. Like it's had all it's had like the top people in the organization come through there over the last like year to kind of get where we're at now. And it's just like you know it's like a building a turboan engine. It's like a complex system and it's hard. And uh there'll be a lot more hard stuff now in the future is getting this thing to work. But I'm actually I actually think we'll be okay on the data side. And then um we do plan to scale figure four to unprecedented levels versus even figure three. So I think from that perspective it will have a mountain more of data um both like you know like on robot side and even uh and even better I think it'll be able to do things better coming from Helix and from training.

[00:49:40]That's okay.

[00:49:42]>> We'll bring them up. We'll bring up a bunch of them. We'll go try them and we'll we'll know. But like uh I'll know this year, but I I feel pretty confident it's going to be great.

[00:49:49]>> Hey Brad, are are are you going to be adding uh speech output to figure anytime soon? And and then my real question that I've always want to ask all roboticists is when are we going to when do you think you're going to get to the point of, you know, taking a brand new figure uh to a uh let's say a u factory workstation that it hasn't seen before. No human has shown anything about this new factory workstation. and the human describes to the robot what needs to be done and maybe even demo something you know quickly like you would a human like how you would train a human when do you think we'll get to that nirvana >> um oh yeah uh so I have to leave after this question uh I just have a meeting here I get back to you but like the I think a few things one is we're um we're using so I started a new AI lab called HARK last summer uh we have a team of about 70 we're working on like new kind of AI kind of ignition generation AI models um and some hardware and uh one of the areas we're spending a lot of time in is like like real time speech to speech. So the robots every robot at figure now has the hardc voice model on it and you're able to stop any robot and talk to it here at the office uh which is which is which is great. Um and then um uh I guess on your second question this is like the holy grail of robotics man like we want to be able to take a robot anywhere and do whatever we want with it. This is like the number one goal I figure is how do we solve for a general purpose machine and solve general robotics and um I the soonest we'll be able to do this demonstrate something like this would be like this year and I don't know we'll see if we'll hit it. I don't know >> that is crazy.

[00:51:36]>> Maybe maybe not. Uh but we whether you know we we believe internally we have a we have a plan that if we execute on that plan we can do something like that. Um now it could be that the plan is crap and we do this and it doesn't work. Um or it could be we do this plan and it works and it just works at such small scale and we'll we'll know more as we you know we just launched the pre-training run for our next generation Helix 3 model to test this last week on the new cluster of B200s we just like of black wells that we just series >> that we just set up um here internally that launched this month.

[00:52:14]>> So I think um I I mean we'll we'll we'll evaluate that model in the coming weeks. Um I don't think at that point it'll be able to do things like this but it is the model architecture that we think should be able to do things like this. Uh so um you know listen hopefully over the next coming years for sure we hope that you'd be able to do this and yeah it's just it's just foggy right because nobody's ever gotten to a point be able to do this. So I think we're we were like we have a plan now that we think if we execute well on we'll have like the best chance possible to get this done. If you look at the plan and get pitched internally, you're like, I'm like, damn, this is like a this this plan. I I think this should be able to solve this.

[00:52:55]>> Okay.

[00:52:55]>> And um but it needs a ton of data. It needs a ton of compute and we need to see patience to do the right ablations in the right model policy evaluations and stuff. So um but this is our number one goal. If you came in here and say Brett, what are we doing? What's I'm like this? We're trying to solve general robotics and you know doing things like you said is like at the top of the list there to be able to do and it could be in a factory it could be in your home could be able to do that you know the bed tidying but like in an unseen home like a random Airbnb we just we we check we book that morning and we drop the robot into and we ask it to go do something and it does it like these are kind of these are the kind of valuation sets we want to be able to solve. Um you know we're trying to solve as fast as possible. Hopefully in the next you know one to two to three years we can be able to do this. Um, if if we can solve it this year, we're going to we're working we're working as hard as possible to try to solve it as fast as possible. But as soon as as soon as we solve it, I will tell you >> even if it happen Yeah. Even if it happens next year, we'll our minds will be blown.

[00:53:49]>> Yeah. I think if it happens within this decade, I think this will be like greatest the world's ever seen.

[00:53:54]>> Okay. So, I I I please if you can just spare a couple of more minutes because I have some uh just two questions which I'm sure you'll love. first being long-term horizon off world deployment of bots for labor on the moon on Mars.

[00:54:10]Um are you thinking about it? Is that part of your plan? Um how big is a challenge uh a ch engineering challenge is it to adopt these uh adapt these bots to off-world deployments and the swag.

[00:54:25]>> Absolutely love the swag.

[00:54:27]>> We want the figure three merch. Yeah, we want the figure that's whatever that is. And if you have time for one last question, one quick last question from Dang, please. That would be wonderful. So, space deployment, please. First, >> okay. Um, so there is a pre-training versus fine-tuning uh trade-off in robotics AI right now. Uh, if you train pre-training with too much data or too less data, you don't get generalization.

[00:54:59]But then you spent uh too much time on fine-tuning. So what does that balance look like right now? And is there a goal within the company to get fine-tuning uh as less as possible?

[00:55:13]>> Um on the space side is before we move into the pre-training question. Uh yeah, we would love to be in space. Uh it's a part of our master plan from when we started. Um and we're having some of those conversations now. Um I would like our robots will most certainly be in space uh and hopefully as soon as possible. Um on the second question on training and I have to leave I'm sorry I just have a 2:30 is um uh we are spending all of our time on pre-training and the way we get to like real generalization from the robot is in is in pre-training. Um that's the stage that we learn this and um like the semantics of the world and this this understanding. Um so um yeah I would say like just just uh hang tight man. We're we're working on it and uh like every day I see new experiments.

[00:56:10]Some work, some don't. And we're working through it. We want what you want. We want like a a world we live in where a human or robot can just do everything like we we like Figure wants to build iRoot the movie.

[00:56:22]>> Yeah.

[00:56:22]>> In real life.

[00:56:24]>> We saw it in the demo.

[00:56:25]>> Yeah. Robots are robots are walking around everywhere. They're inside.

[00:56:29]They're outside. They're at home.

[00:56:30]They're in the workforce. They're in the billions. Like we want to do that. In order to do that in 2026, figure needs to solve generalization.

[00:56:38]>> Stuff that Phil talks about stuff we talking about here. That's the number one goal by a mile. It's not manufacturing. It's not Um, it's not bomb cost, it's not supply chain, it's not it's not any of these things. It's generalization. That's what everybody should be figuring out as question is like uh how do we solve this and figure I think will be the first to get there and we're working really hard right now.

[00:57:00]>> But um listen guys, I got to run. I got I got a meeting. Thanks for having me.

[00:57:03]>> Thank you.

[00:57:03]>> And uh it's always Thanks.

[00:57:06]>> Thank you. Thank you for your time and congratulations. It's been awesome.

[00:57:11]>> Hope to have you back. See you guys.

[00:57:12]Bye. Bye.