If they could implement an AI on the scale of gpt-5 into this thing, then I would in a heartbeat, although just thinking about it, this could do a lot more then choirs, this could do a LOT of manual labour jobs with enough training and data, for people who don't need to spend 100k on a Tesla bot, then could spend 30k on this depending on their needs, i feel like it could even take my job as a barista by the end of the year
Just watched that video. That project is wild. Open source and 30k in COTS parts all by a part time team of 3 phd students. Made me think of something wild. It has two modes of operation. First a fully autonomous ai mode. Secondly, a human operable mode where you train it with your movements that can then be usable in mode 1. Because the robot is open source you can have people around the world build it and as they train it in different tasks you upload that data to be downloadable by the rest of the community. The rate of improvement would be insane, literally overnight.
Yes, that's a brilliant way to train them via a community across the world. Just gotta make sure some of the tasks aren't something messed up like picking locks(robbing banks)/attacking pets or people or illegal things etc We can see all the good applications and how helpful they can be, but how many criminals will use this technology too, at least robots are ultra slow right now and that was speed up at 6x speed for most of the tasks. But it's pretty wild that you can train them to do almost any task if they just need to have muscle memory. It's scary, but also cool because robot sports might be interesting to watch in the future. Robots boxing/swimming etc
What would really make this go crazy: virtual training. Use a game engine to model physics and what the AI would see. Train it on 100 years of household tasks in a few days, and then clone the AI to as many butler or chef robots as you can
I think the fact that there's possible to connect all of them wirelessly and if there are a million of them out there, then they will learn quickly anyway.
So many pieces to a full humanoid robot are coming together. This, plus the Ameca robot for facial expressions. All of the LLM stuff for conversations. The fluidity of Tesla robot mechanics and Boston dynamics.
This video brings back memories of the Chernobyl incident back in the 1980's where it would have been nice to deploy robots trained on disaster tasks so humans would not have to be deployed in such dangerous conditions. I do believe with some modifications these are capable of handling such dangerous tasks. Kudos for the video
I believe the robots they tried to use first broke down quickly from the extreme radiation, so they had to resort to human "liquidators" wearing extremely heavy (but still inadequate) lead armour...
I'm a construction guy and used to joke with my colleagues about maybe if we had enough bit extenders we could be working from home like our management. Tah-Dah! This thing's a God send!
Co-training, like you operating a drill remotely, using one of these, is absolutely going to be a big career for at least a window of time (maybe forever) while we train AIs to perform all these tasks.
@@mr-biz Good news is that the billionaires won't need the pesky roaches any more. Bezos dream of turning Earth in to a park is coming to full realization! All Hail Bezos! God, master of all! Daddy!
So it's a bit like using an AI to grade an AI, only the grading is done by humans and then the AI generalises from there. It's RLHF with the HF first. It's exciting because the AI learns to generalise after the human hands over the reins.
Soon these will cost 10,000, be less bulky, perform a wider variety of tasks, and be more consistent in successful task operation. At which point it's a no-brainer to buy, the time saved from household chores will be huge. It may be a while before it's a top tier chef, but once it can make many different decent quality meals, from store bought ingredients, that'll be huge as well.
Will this (one of your) robot assistant(s) then get to enjoy a solo trip to the store to get the ingredients it needs, do you think? Or is that a different assistant? Or does the store send their delivery robot to hand over the groceries to (one of) your assistant(s) at your door instead? Or will we rather be assisted by a fleet/swarm of rented robots, waiting to serve us on every street corner (and lake bottom) like those pay-as-you-go electric scooters?!
Well ultimately yes, any and all of those things. At the time of the robot I described above i'd still prefer going to the store, I like being able to improvise. But as things get better, yes I wouldn't mind the robot managing a list of what i usually like to have, and ordering appropriate restocks of what I like and have regularly. At which point I'd only go to a store when i want to explore options to expand my palate. On top of all this, given the extra free time i'd have in my life i'd like to manage my own garden, which would supplement the amount of food I need to get from ordering. Which i'm sure a robot capable of doing all sorts of household chores could help with as well. But things in stages, some of the capabilities you described I see as being more ahead of what described, at least when it starts being available. The way I see it is there's several development steps before get to automatic store to table at home meals/equivalent chore capabilities. Such as more compact robotics, better manipulating parts, more consistent task performance, strong safety rating of not injuring humans in the working area, task priority and ordering (thinking of getting the ingredients *ahead* of dinner sort of thing), and a few other similar hurdles. Plus price, a lot of those things are kinda possible now, but very expensive and still not to the level of quality we'd really want. My guess is a 10,000$ bot that can do full food ordering/cooking, dishes, laundry, cleaning, and other similar level tasks is more gonna be 2026, though possibly mid 2025. I think there will be stuff with about half the quality and twice the price (25000$ with much higher task quality/consistency, and wider task range) than that shown in the video, by the end of the year, and as a commercial product coming to market. So probably not released, but certainly demo'd advertised, and on it's way publicly. @@lauridskristensen9800
@@TheMajesticSeaPancake Can you imagine a bunch of fat slugs, spreading over their sofas staring into their 3D goggles or whatever, while all manual work is done by a sexy-looking robot or two? I've worked at home for year and never found housework much of a burden. Gardening on the other hand...
Well if you don't like gardening you can do all the dishes while a robot tends to your front lawn while I do the opposite. As far as what other people do on their couches, well good for them, not my problem.@@bigglyguy8429
It looks like the legs could just be an RC car or a Roomba vacuum. Don't have to buy anything special for that, it'll save you 7k just by doing that. Now 1k for the battery? You could definitely just use something that will last you for 2 hours or so, so $500 would be more than enough. Now as far as the robot arms, I do not have any ideas... But there you go, I saved you $7500!
I want to state the depth that you put into the technical aspects of the paper into this YT video. Already watched to other AI YT creators and they just touched the subject in the technicalities that you went into! Keep it this way!
George Jetson's (The Jetsons) maid is a robot named Rosie. Rosie is a Model XB-500, an older, high-tech robot that the Jetson family hires from U-Rent A Maid. This "U-Rent" business model allows an economy of scale with respect to the training cost. Working up a business plan for Venture funding now?
I just realized that Apple Vision Pro could be used as a tool for training tasks for robots. It already tracks your hands, eyes and the surrounding space. This can be huge.
Why is Google‘s name on this? This work was performed by researchers at Stanford University. Did Google pay for it or was the Aloha system developed by them?
I just want it to be verbally trainable via LLM integration. I want to be able to tell it “After flipping the shrimp, you should move it to the center of the pan so it cooks more quickly and evenly,” and I want it to remember that advice permanently. I’ll be truly impressed once it can handle zero-shot, verbally instructed tasks like “flip each shrimp using the tongs” when the tongs are hidden in some drawer somewhere. It would need to reason: “I need a set of tongs, but I see none. Where might I find one? One is likely to be in one of these drawers. I’ll search each one, but I’ll have to be quick or else the shrimp may burn. [...] I’ve found the tongs and can now continue the task.” Until it can do that successfully (and is guaranteed not to break anything or burn down my house), I’m not shelling out $32k.
Thank you for the update on this! Definitely 10 years or less. The price makes it ripe for massive optimization. Would like to see it for like $7500, maybe $10k tops?
There's definitely a lot of optimization they could do to this rig overall to improve efficiency and cost. For one bipedality could help eliminate the need for the hefty power systems, and among some of the other expensive items that would be swapped out for more anthropomorphic parts. $7000 to $10k could be very achievable for a competent robot companion.
Please god this can’t come soon enough. I spend 2-3 hours a day just on chores cleaning up after kids and laundry etc. there is so much I want to do. having a robot do all this would make me super productive. Also a robot that could garden could cut down on my grocery bill and give me healthy food.
It looks like the hardware has great potential to go substantially down in price when mass-produced. I would guess something profitable should be possible in the $4500 range, but depending on further R&D and startup costs, the price could go up.
With a bit of cosmetic appeal and a little bit of tweaking the capabilities, I believe people would easily pay 20K. You certainly could not hire a maid or butler for any extended length of time for 20K.
HUGE impact for elder care. A lot of nations are facing a looming crisis of elder care for the huge boomer generation. This would be a massive burden (financially and emotionally) on the younger generations. Robots that can handle a lot of these burdens will have a huge societal impact and save lots of money (and the kid's sanity) I should be building and customizing this robot tomorrow and fundraising on Monday.
Although this will also be a tragedy. Aged care workers often earn minimum wage, and their tasks will be "easily automated", but the human contact that decent care workers provide for many elderly is invaluable.
@@Juttutin Not a tragedy at all. A huge of amount of people will not be able to afford care workers year after year. It's these robots or the family has do all this stuff themselves. It will be a huge help for people all across the world.
@@ymroddi it probably depends a lot on where you are. I've had exposure to the budget end of for-profit elderly care, and some of those (already quite dystopian) places would cheerfully replace most of the workers with robots if it was cheaper. Also my Mum, who is very independent and living alone, would probably prefer, at some level, a mechanical rather than human cleaner (as would I perhaps, our family seems to be genetically loners). But it's not healthy. Human contact leads to statistically meaningful improvements in average health, life quality, and lifespan. Both can be true.
@@Juttutin So you are proposing family stay with the elderly to provide 24/7 support for years (decades?) - or simply have robots take away a good deal of this burden. One can have robots and human contact. It really isn't one or the other. There will no doubt be a huge need in coming decades that the care industry will never be able to satisfy with available staff - even assuming people could afford it. And they can't afford it. I many many cases it will simply be robots or no care at all.
@@ymroddi there are already many places where the human contact in elderly care is severely limited and understaffed, despite high unemployment. I am not making any proposals, just cautioning that for many, our economies are such that there will be incentives to replace, rather than augment human staff where possible, regardless of any available but underutilised workforce. Because providing the bare minimum is more profitable. I don't see (yet) the structural economic changes that will be required to prevent low-cost elderly care from sliding further towards some black mirror episode. On the other hand, I can easily see how, for the reasonably wealthy, a few robots will allow care workers to spend much more time on the intrinsically human aspects of elderly care.
I can't wait for the $32k machine to cook me a piece of shrimp with a success rate of 40% after I've demonstrated it to it 20 times. Exciting yes, "likely to be my assistant in the very near future" probably not.
The idiom "in the very near future" means "in 5-10 years" to you? ChatGPT tells me > "In the very near future" is subjective and can vary, but it generally implies a short and imminent timeframe, typically within the next few days or weeks. @@sssurreal
I forget who said it, I think Elon, but the difference between getting a robot to do what you want 90% of the time (or in this case 80%) and getting it to do it 99.999% of the time, which is the only rate acceptable in a consumer product, is night and day.
Most of these tasks were using the tele afaik. It's still cool that it's able to cook a shrimp, though. And soon it will be able to do the dishes and all the other stuff and more just by itself :D. Crazy that it can run on consumer grade hardware. Excellent video as always!
According to my wife I am incapable of being trained to rinse a dish. Suspiciously, when she saw this video, she asked me how much life insurance I had...
It might still be too expensive for your average working class person but I think if a company was to offer a payment plan for a couple hundred bucks a month for a number of years the way we finance cars, there would be a lot of people willing to have a less fancy car for a while so they can have a permanent cook and maid. Assuming it's effective at cleaning a house it has never seen as it's shown to be and it can cook a variety of meals completely on its own, that alone is worth the sticker price and a financing plan could get the monthly costs down to the price of a meal prep service. Granted, you still have to buy the groceries (or have them delivered via drone and brought in by Aloha) but still, people pay a lot more for less utility.
My first thought about this was that we need a 3rd hand .. in so many daily chores around the house, in our work, etc., such as hammering a nail while holding up the object in the 3rd hand .. or holding a large heavy carton with two hands while the 3rd opening the door handle .. Then when re-thinking it, I saw that what we really need is to improve the two hands, to provide multi-attachment capabilities, in addition to the grabber .. First and the simplest .. so many of the tools we made for our human hands have a long handle, such as a broom, a hammer, a soup ladle, etc. We could engineer a simple straight forward accommodation for this in addition and alongside the 'hand grabber', that would obtain a much better end-result .. Sure, not all the time, but most of the time .. And .. such a small price to pay for a much better result, while staying with two hands .. [which makes the actual human training session easier and more natural] ..
I believe this is already enough to run autonomusly a hotel. A mixture of this and development from a company, I think it's totally doable In the next 2 years to spend the night in one.
Tracer AGV can be replaced with cheaper and diy solution. Robot harm can be build also. In less than 2 -3 month a mod version of this project will caost 10k less. Also the realtime interference can be done with a less Power consuming HW or through a fix solution trough wifi.
Let’s hope Nvidia might gift researchers with some Jetson AGX Orins! Comparable in cost to the laptop but could support 6 cameras (1080p at 60fps). I imagine other sensors like LiDAR, infrared, et cetera could aid in recording more ground truth and enhance adaptability.
That is totally cool! It would also be really cool if they added two things to the training process where the person doing the training can speak and tell the AI what they are doing where that speech is also recorded and a camera records the video of the training process which could be used to to train the AI to be able to generalize a video of perform that task. For example when putting the pot in the cabinet the trainer can say, I am now putting the pot in the cabinet on the bottom shelf using both hands and thus one can train a chat GPT while also training the "muscle memory" where the AI can then link the two together and not only learn to perform a task but also learn how to instruct others in words how to perform it as well as build a textual representation of the robot's library of tasks it knows how to do and maybe even build an ontology model as well. BTW, as to how to do the later I have made a number of videos and uploaded them to my UA-cam channel on a way to do that which I call UniML, Universal Modeling Language. So in UniML speak, the trainer's words could be treated as Swim Lane Islands (SLI's) which the can be assimilated (gobbled up) into RDF like statements as well as can be used to train a chat bot. So then one would have a number representations, the recorded motions, the AI training data derived from them, the recorded speech and video of the training for that task, the importation of that into UniML as SLIs which it then assimilates (which could be done by its own Chat bot that has been trained to do that), and probably more. These then could be used to generate training videos and documents as well as be added as part of a world model representation of what is doable as well as how one can do it.
BTW, co creation of the Ontology model might then be thought of as an artificial consciousness (thoughts) while the AI that can perform the task as muscle memory. Thus when the robot is putting the pot on the bottom shelf it can "think" I am no putting the pot on the bottom shelf using both hands because the task was linked to the words via the ontology model. It could then take that sentence as use that as a prompt for its own chat bot to see what that might produce like be careful to make sure it is clean before doing that. The the ontology model serves not so much as a bot but as the human that ask or tells the bot things, i.e. does the prompt engineering base on its ontology models, that could have been in part created by a chat GPT. It is sort of like the left brain conversing with the right brain and vice versa.
BTW, the another reason to do all this is to facilitate sharing where one person might train their robot to put a pot on the shelf as well as some other person might to the same but with the verbal descriptions added and the ontology model creation they could share these on a network that might then combine the two into a fused AI that uses both sets of training data. Indeed that is one of the uses of ontology models is to allow different system to be interoperable in that the ontology model serves as a data broker between the two so they can know what the other system's data represents. And there are such Ontology schemes such as OWL and RDF where in UniML I try to extend that to be more graphical in nature, like going from DOS to a GUI like how Steve Jobs did with the Macintosh from DOS on a PC back in the day.
Also during training of the robot, half of the training should be by a left handed person, the other half by right-handed .. That way, the robot may excel above and beyond either of them ..
Industrial robots have been doing this for decades - repeating actions trained by a remote operator. I'd like to be enthusiastic about this, but it really just looks like 90% teleoperation, 10% repetition. If you pay attention, you'll notice that all of the demos are performed using teleoperation - in some, you can clearly see the operator in the background. You'll also notice the complete absence of adverse or unexpected circumstances from the demos. In the absence of 3rd party testing, skepticism should be the default.
One of the major appeals of this isn't necessarily the training method, but the fact that this is open source and fully accessible. Functionality can be added in and modified rather rapidly. It doesn't matter if the pilot has to do some work initially, that's the whole point. Skepticism is fine, but don't be so biased that you become hyper-skeptical.
There a re clearly no operators visible in the main demos (I just watched again to be sure). The footage at the end of the video however shows the training side of things, where people are obviously required to be operating it.
Fair enough. If it's real, then we should expect to see this replicated and improved very rapidly given it's open source. If it's vaporware, then we should expect everyone to forget about it after six months or so. I don't think it's hyper-skeptical to suspect the latter, but I'd also be pleasantly surprised to be wrong. Either way, we'll know soon enough.
bimanual manipulation is a very tricky task. HHH AI is a game changer in robot control. We'll all be put to retirement soon. To do that using classical robotics, it is extremely complex. We have built a bimanual robot and i know how hard it is. And they are very expensive. Waw, AI did replace many methods and Robotics are not an exception.
I want to buy those equipments and make one at home, for my son, who has problems with his hand and arm. I think with a certain amount of time on learning, the machine can make a massage daily. That'll be great helpful.
wow, fully open source, something is starting there! I can see a lot of opportunity in cost reduction. Also it should be possible to train remotely with a virtual headset. Train it slow then accelerate on execution. Lets build a training dataset from the crowd! Some computation can be offloaded or improved by cloud computation. But it needs safety, it is not safe to launch an autonomous robot with humans. It may break bones, it may start fire... Actuators need to have a safe force limit and human supervision.
5 years? Maybe not. Consider the pressure Mobile Aloha just put on the robotics industry as a whole to push for even better breakthroughs. Musk has to know about Mobile Aloha and I don't figure he will stand idly by while it outperforms Optimus.
"Only $32,000" Man I wish I was American if they think that something that costs so much is not expensive. On another note. The robot leaves the frying pan handle over the edge of the cooker, which everyone with kids knows is something you don't do. Also it puts the hot frying pan down on the worktop.
But I already watched this video from Wes Roth >_< Also, currently, despite the cherry picked examples, it's really not ready for use, teleoperation cool and all but not something affordable to deploy and it's few shot learning is not up to snuff yet.
Best proof of concept I've seen for Tesla's Optimus gen 2 bot in terms of use cases. They'll mass manufacture it for cheap enough eventually. Invest wisely
I don't see what exactly should cost 34k here. 4 motors? Plastic? What exactly? 2k at best. Unless you want to some gems inside of it and few layers of gold.
I'm sure someone will build this testing out cheaper parts to see how bad it would do. Great idea to keep it open source so people can iterate over the plans. I think even if you modded it to be human controlled if you were disabled in various ways it could be a great help if it was affordable.
"Mock-Up" A useful word, that you should know for the forthcoming cycle of development and hype in the AI field. It is basically a fake, but use legitimately it's intention is to give everybody an idea of what the final version should look like and how it should function Basically it generally means, producing something that appears to function the way you intend something else to function. If you're making a program, or doing an experiment, or designing the scenery for play, or creating an invention to automatically function. You will create something that appears to be the real thing so long as you look at it from just the right angle, with the right editing, and there's a lot of predictable events or human interactions. So for example, let's say you're writing a program for a calculator. You might write a much simpler program that looks like a calculator, and always prints out the number "5" when you hit '=' and then during the mock-up demonstration you type in 2+3 = and it prints out 5 even though no calculations have been done. There are legitimate uses for mock-ups. Basically you can use them to make sure that everyone on your development team knows what you're going for. You can inform investors as to your goals and give them a practical demonstration of what you want it to do. You can test aspects of an experiment to make sure that it can actually detect the result of the experiment correctly, and I know they in fact did this with the gravitational observatories, they fed it fake data to see if it could actually identify merging black holes. I've done it in biology, where I've fed in fake experimental results into a gel and into a sequence analyzer to see if it could detect the real thing when I ran the experiment. This is sometimes called a positive control.. It goes from a mock-up, to fraud when you represent it as real results obtained from natural inputs. If I say look my calculator can add two numbers, two plus three equals and a five pops out, I am misleading people as to whether or not actual addition is occurring. If I had presented my positive controls in my biology experiments as if they were the result of an experimental intervention rather than me custom designing DNA fragments to give a particular appearance, that would have also been fraud. Unfortunately fraud like that is really hard to detect unless you have independent people that have access to the same things and can verify your results as well as test for the assumed full range of applications. I do think that some of what Google did qualifies as mockup that they then mixed with semi legitimate stuff, creating what is properly called a fraud. Mockups are actually really useful tools and represent an organic part of the research and development process. They are also very easy to translate into fakery and fraud especially when there's a lot of pressure to get results fast and there's a lot of money on the line. You need people with a lot of maturity and a lot of integrity to use mock-ups properly.
He's just covering an academic research project. It's not like that team is trying to convince people to buy these research platforms with pre-orders. I get your point that it's not 100% proven and could turn out to be less capable or more limited than it looks. But they even showed the robot failing in the outro to drive home the point that it still fails sometimes right now.
These robots will implemented in basic settings like Janitorial positions, window-cleaning, and even McDonald’s!! To train the robots, you still need human interaction. This is implemented as a new position of ‘janitor’. So, to be a janitor in the future, YOU WILL NEED A COLLEGE DEGREE. THINK ABOUT THAT!!! FOOD FOR THOUGHT!!!! FUTURE, HERE WE ARE!!! 🎉❤
Ok , so it trains on certain situations. But what if the glass of wine or spillage is somewhere else? Is it smart enough to change its path and do it with the same accuracy? Do not think so... Still a cool project for sure!
In a year these will be pretty competent I imagine, the problem with robots right now is the poverty of data. We should build like a million of these and share our datasets.
Seriously...do you want one of these in your house??
It this version. No. But give them 3 years and this bitch will look like Walle and you will want one yourself.
The 5th generation one maybe but I love it!!!! This is amazing!!
Absolutely
After I wrap it in a RealDoll® skin...
If they could implement an AI on the scale of gpt-5 into this thing, then I would in a heartbeat, although just thinking about it, this could do a lot more then choirs, this could do a LOT of manual labour jobs with enough training and data, for people who don't need to spend 100k on a Tesla bot, then could spend 30k on this depending on their needs, i feel like it could even take my job as a barista by the end of the year
Just watched that video. That project is wild. Open source and 30k in COTS parts all by a part time team of 3 phd students.
Made me think of something wild. It has two modes of operation. First a fully autonomous ai mode. Secondly, a human operable mode where you train it with your movements that can then be usable in mode 1. Because the robot is open source you can have people around the world build it and as they train it in different tasks you upload that data to be downloadable by the rest of the community. The rate of improvement would be insane, literally overnight.
Yes, that's a brilliant way to train them via a community across the world.
Just gotta make sure some of the tasks aren't something messed up like picking locks(robbing banks)/attacking pets or people or illegal things etc
We can see all the good applications and how helpful they can be, but how many criminals will use this technology too, at least robots are ultra slow right now and that was speed up at 6x speed for most of the tasks. But it's pretty wild that you can train them to do almost any task if they just need to have muscle memory.
It's scary, but also cool because robot sports might be interesting to watch in the future. Robots boxing/swimming etc
hopefully they build the site to do this, someone will clearly
A wireless, "growth spurt", every night while it is idle.
@@phen-themoogle7651autonomous security deposit box lockpicker and burgaler
@@kevinnugent6530Jesus it’s over 😂
What would really make this go crazy: virtual training. Use a game engine to model physics and what the AI would see. Train it on 100 years of household tasks in a few days, and then clone the AI to as many butler or chef robots as you can
I think the fact that there's possible to connect all of them wirelessly and if there are a million of them out there, then they will learn quickly anyway.
Nice
Nvidia is doing that
Did my headhunter at Tesla already called you? If not.. keep your phone nearby!
now we just need Gordon Ramsay to make a training video game.. or someone to take all his videos of him cooking and make them POV.
Hmmm...considering the previous Google AI "demo"....
Exceptionally well done... those PhD students just accelerated progress by a huge amount of time.
This is next level!
So...we didn't learn anything from the whole Google Gemini stunt I see.
So many pieces to a full humanoid robot are coming together. This, plus the Ameca robot for facial expressions. All of the LLM stuff for conversations. The fluidity of Tesla robot mechanics and Boston dynamics.
Please make one as it's open source
Also Chinese sweat shops can now be replaced, finally!
By chinese people remotely controlling the robot?
This video brings back memories of the Chernobyl incident back in the 1980's where it would have been nice to deploy robots trained on disaster tasks so humans would not have to be deployed in such dangerous conditions. I do believe with some modifications these are capable of handling such dangerous tasks. Kudos for the video
I believe the robots they tried to use first broke down quickly from the extreme radiation, so they had to resort to human "liquidators" wearing extremely heavy (but still inadequate) lead armour...
In Fukushima, Japan, the problem was not lack of robots but that they were breaking very quickly due to high radiation
And we are just in the first week of the year!
I'm a construction guy and used to joke with my colleagues about maybe if we had enough bit extenders we could be working from home like our management. Tah-Dah! This thing's a God send!
Co-training, like you operating a drill remotely, using one of these, is absolutely going to be a big career for at least a window of time (maybe forever) while we train AIs to perform all these tasks.
At least part of our careers anyways.
The only thing that these robots will be unable to replace is builder's crack when they take your construction jobs.
@@mr-biz Good news is that the billionaires won't need the pesky roaches any more. Bezos dream of turning Earth in to a park is coming to full realization! All Hail Bezos! God, master of all! Daddy!
@@mr-biz You under-estimate the technology....
Most of these videos show the robot being teleoperated and not acting autonomously
So it's a bit like using an AI to grade an AI, only the grading is done by humans and then the AI generalises from there. It's RLHF with the HF first. It's exciting because the AI learns to generalise after the human hands over the reins.
Great observation
This is incredible! I'd love to have one of them
This is the most important channel in YT right now
The machine that will kill us at failure.
Compleatly agree this is mind blowing. I believe this project can spur the creation of new real-life practical robots everywhere.
5:20 My brain had some tricky perception visual glitch.
I think most of those videos of complex tasks are those guided during training
9:11 that logic is NICE!! flipped the lid so it would fit better!
Cyberdyne Systems T-0001 has this tech already working for general use. Can’t wait for the next models they may release in the near future!
0:52 it even shakes the fuckin pillow in 😂😂😂 Idk why that was so funny 💀
manufacturing and my dishes are extremely excited!
Great research video Matthew
I think all time lines are gonna move up every year, 5 years before these are almost everywhere
Soon these will cost 10,000, be less bulky, perform a wider variety of tasks, and be more consistent in successful task operation. At which point it's a no-brainer to buy, the time saved from household chores will be huge. It may be a while before it's a top tier chef, but once it can make many different decent quality meals, from store bought ingredients, that'll be huge as well.
Will this (one of your) robot assistant(s) then get to enjoy a solo trip to the store to get the ingredients it needs, do you think? Or is that a different assistant? Or does the store send their delivery robot to hand over the groceries to (one of) your assistant(s) at your door instead? Or will we rather be assisted by a fleet/swarm of rented robots, waiting to serve us on every street corner (and lake bottom) like those pay-as-you-go electric scooters?!
Well ultimately yes, any and all of those things. At the time of the robot I described above i'd still prefer going to the store, I like being able to improvise. But as things get better, yes I wouldn't mind the robot managing a list of what i usually like to have, and ordering appropriate restocks of what I like and have regularly. At which point I'd only go to a store when i want to explore options to expand my palate. On top of all this, given the extra free time i'd have in my life i'd like to manage my own garden, which would supplement the amount of food I need to get from ordering. Which i'm sure a robot capable of doing all sorts of household chores could help with as well. But things in stages, some of the capabilities you described I see as being more ahead of what described, at least when it starts being available. The way I see it is there's several development steps before get to automatic store to table at home meals/equivalent chore capabilities. Such as more compact robotics, better manipulating parts, more consistent task performance, strong safety rating of not injuring humans in the working area, task priority and ordering (thinking of getting the ingredients *ahead* of dinner sort of thing), and a few other similar hurdles. Plus price, a lot of those things are kinda possible now, but very expensive and still not to the level of quality we'd really want. My guess is a 10,000$ bot that can do full food ordering/cooking, dishes, laundry, cleaning, and other similar level tasks is more gonna be 2026, though possibly mid 2025. I think there will be stuff with about half the quality and twice the price (25000$ with much higher task quality/consistency, and wider task range) than that shown in the video, by the end of the year, and as a commercial product coming to market. So probably not released, but certainly demo'd advertised, and on it's way publicly. @@lauridskristensen9800
@@TheMajesticSeaPancake Can you imagine a bunch of fat slugs, spreading over their sofas staring into their 3D goggles or whatever, while all manual work is done by a sexy-looking robot or two? I've worked at home for year and never found housework much of a burden. Gardening on the other hand...
Well if you don't like gardening you can do all the dishes while a robot tends to your front lawn while I do the opposite. As far as what other people do on their couches, well good for them, not my problem.@@bigglyguy8429
It looks like the legs could just be an RC car or a Roomba vacuum. Don't have to buy anything special for that, it'll save you 7k just by doing that. Now 1k for the battery? You could definitely just use something that will last you for 2 hours or so, so $500 would be more than enough. Now as far as the robot arms, I do not have any ideas... But there you go, I saved you $7500!
It looks like a 3D printer then went rogue and printed itself into a robot.
I want to state the depth that you put into the technical aspects of the paper into this YT video. Already watched to other AI YT creators and they just touched the subject in the technicalities that you went into!
Keep it this way!
This is it buddy. The cost will come down radically over the next year and they'll be everywhere. The robots are here.
George Jetson's (The Jetsons) maid is a robot named Rosie. Rosie is a Model XB-500, an older, high-tech robot that the Jetson family hires from U-Rent A Maid.
This "U-Rent" business model allows an economy of scale with respect to the training cost. Working up a business plan for Venture funding now?
I just realized that Apple Vision Pro could be used as a tool for training tasks for robots. It already tracks your hands, eyes and the surrounding space. This can be huge.
this is crazy good. 30k AI OPEN source robot can do dish, cook and laundry ??? OMG.... it makes coffee for you
I think it’s ready to go out into the field with some soldiers and learn from them. Skynet is close.
Why is Google‘s name on this? This work was performed by researchers at Stanford University. Did Google pay for it or was the Aloha system developed by them?
I just want it to be verbally trainable via LLM integration. I want to be able to tell it “After flipping the shrimp, you should move it to the center of the pan so it cooks more quickly and evenly,” and I want it to remember that advice permanently.
I’ll be truly impressed once it can handle zero-shot, verbally instructed tasks like “flip each shrimp using the tongs” when the tongs are hidden in some drawer somewhere. It would need to reason: “I need a set of tongs, but I see none. Where might I find one? One is likely to be in one of these drawers. I’ll search each one, but I’ll have to be quick or else the shrimp may burn. [...] I’ve found the tongs and can now continue the task.”
Until it can do that successfully (and is guaranteed not to break anything or burn down my house), I’m not shelling out $32k.
Too bad we can't download their trained AI model. You have to train by yourself! 😢
Thank you for the update on this! Definitely 10 years or less. The price makes it ripe for massive optimization. Would like to see it for like $7500, maybe $10k tops?
There's definitely a lot of optimization they could do to this rig overall to improve efficiency and cost. For one bipedality could help eliminate the need for the hefty power systems, and among some of the other expensive items that would be swapped out for more anthropomorphic parts. $7000 to $10k could be very achievable for a competent robot companion.
I'd say 2 years to reach businesses for 10k and 3 to reach consumers for 5k.
@@VperVendetta1992 2nd hand sexbots for 500...
Please god this can’t come soon enough. I spend 2-3 hours a day just on chores cleaning up after kids and laundry etc. there is so much I want to do. having a robot do all this would make me super productive. Also a robot that could garden could cut down on my grocery bill and give me healthy food.
I can totally see myself doing the dishes with my robot, hi-fiving each other haha.
It looks like the hardware has great potential to go substantially down in price when mass-produced. I would guess something profitable should be possible in the $4500 range, but depending on further R&D and startup costs, the price could go up.
With a bit of cosmetic appeal and a little bit of tweaking the capabilities, I believe people would easily pay 20K. You certainly could not hire a maid or butler for any extended length of time for 20K.
Not gonna lie, I giggled when you said $32k after you said it was affordable. :D It pulled me back to reality. :(
HUGE impact for elder care.
A lot of nations are facing a looming crisis of elder care for the huge boomer generation.
This would be a massive burden (financially and emotionally) on the younger generations.
Robots that can handle a lot of these burdens will have a huge societal impact and save lots of money (and the kid's sanity)
I should be building and customizing this robot tomorrow and fundraising on Monday.
Although this will also be a tragedy. Aged care workers often earn minimum wage, and their tasks will be "easily automated", but the human contact that decent care workers provide for many elderly is invaluable.
@@Juttutin Not a tragedy at all. A huge of amount of people will not be able to afford care workers year after year.
It's these robots or the family has do all this stuff themselves.
It will be a huge help for people all across the world.
@@ymroddi it probably depends a lot on where you are. I've had exposure to the budget end of for-profit elderly care, and some of those (already quite dystopian) places would cheerfully replace most of the workers with robots if it was cheaper.
Also my Mum, who is very independent and living alone, would probably prefer, at some level, a mechanical rather than human cleaner (as would I perhaps, our family seems to be genetically loners).
But it's not healthy. Human contact leads to statistically meaningful improvements in average health, life quality, and lifespan.
Both can be true.
@@Juttutin So you are proposing family stay with the elderly to provide 24/7 support for years (decades?) - or simply have robots take away a good deal of this burden. One can have robots and human contact. It really isn't one or the other. There will no doubt be a huge need in coming decades that the care industry will never be able to satisfy with available staff - even assuming people could afford it. And they can't afford it. I many many cases it will simply be robots or no care at all.
@@ymroddi there are already many places where the human contact in elderly care is severely limited and understaffed, despite high unemployment. I am not making any proposals, just cautioning that for many, our economies are such that there will be incentives to replace, rather than augment human staff where possible, regardless of any available but underutilised workforce. Because providing the bare minimum is more profitable.
I don't see (yet) the structural economic changes that will be required to prevent low-cost elderly care from sliding further towards some black mirror episode.
On the other hand, I can easily see how, for the reasonably wealthy, a few robots will allow care workers to spend much more time on the intrinsically human aspects of elderly care.
I can't wait for the $32k machine to cook me a piece of shrimp with a success rate of 40% after I've demonstrated it to it 20 times. Exciting yes, "likely to be my assistant in the very near future" probably not.
Yeah, a bit of skepticism would be nice in these hype videos. 😂
5-10 years is the near future
The idiom "in the very near future" means "in 5-10 years" to you? ChatGPT tells me
> "In the very near future" is subjective and can vary, but it generally implies a short and imminent timeframe, typically within the next few days or weeks.
@@sssurreal
I forget who said it, I think Elon, but the difference between getting a robot to do what you want 90% of the time (or in this case 80%) and getting it to do it 99.999% of the time, which is the only rate acceptable in a consumer product, is night and day.
My success rate is 90 to 97%. But the robot need cams to check success of corse. Plan, do, check, correct. Like I do. Coming next.
Most of these tasks were using the tele afaik. It's still cool that it's able to cook a shrimp, though. And soon it will be able to do the dishes and all the other stuff and more just by itself :D. Crazy that it can run on consumer grade hardware. Excellent video as always!
Another interesting Video, Thank You. Oh BTW I Like the "Lazy Weekend look"
thanks for highlighting the significance of this new innovation - great video!
yup! nice 2 see! pretty cool! thanks 4 the vid Matthew! the bloopers were helpful!
According to my wife I am incapable of being trained to rinse a dish. Suspiciously, when she saw this video, she asked me how much life insurance I had...
I just want to be alive to see Gordon Ramsey yells to a bunch of cooking robots
how many "CHANGES EVERYTHING" would be there , even change would cry
Did that thing just make a cup of coffee? That's AGI
It might still be too expensive for your average working class person but I think if a company was to offer a payment plan for a couple hundred bucks a month for a number of years the way we finance cars, there would be a lot of people willing to have a less fancy car for a while so they can have a permanent cook and maid. Assuming it's effective at cleaning a house it has never seen as it's shown to be and it can cook a variety of meals completely on its own, that alone is worth the sticker price and a financing plan could get the monthly costs down to the price of a meal prep service. Granted, you still have to buy the groceries (or have them delivered via drone and brought in by Aloha) but still, people pay a lot more for less utility.
It is already better than me at cooking certain dishes!
ok, so take the shrimp out of the saucepan, this robot will still continue to "cook it" in the pan, when it is not there anymore ;)
That thing better wash its "hands" after cleaning the toilet before cooking my shrimp! 😜
My first thought about this was that we need a 3rd hand .. in so many daily chores around the house, in our work, etc., such as hammering a nail while holding up the object in the 3rd hand .. or holding a large heavy carton with two hands while the 3rd opening the door handle .. Then when re-thinking it, I saw that what we really need is to improve the two hands, to provide multi-attachment capabilities, in addition to the grabber ..
First and the simplest .. so many of the tools we made for our human hands have a long handle, such as a broom, a hammer, a soup ladle, etc. We could engineer a simple straight forward accommodation for this in addition and alongside the 'hand grabber', that would obtain a much better end-result .. Sure, not all the time, but most of the time .. And .. such a small price to pay for a much better result, while staying with two hands .. [which makes the actual human training session easier and more natural] ..
What really brings it all together is the personification of the machine. You all saw how beautiful it's googly eyes were right?
The sexbots are gonna be a hit...
I believe this is already enough to run autonomusly a hotel. A mixture of this and development from a company, I think it's totally doable In the next 2 years to spend the night in one.
I've seen a video of a place in Japan where the gimmick is that the staff is largely robots. Right around the corner
Its abilities to create shrimps out of thin air are certainly impressive!
Tracer AGV can be replaced with cheaper and diy solution. Robot harm can be build also. In less than 2 -3 month a mod version of this project will caost 10k less. Also the realtime interference can be done with a less Power consuming HW or through a fix solution trough wifi.
Let’s hope Nvidia might gift researchers with some Jetson AGX Orins! Comparable in cost to the laptop but could support 6 cameras (1080p at 60fps). I imagine other sensors like LiDAR, infrared, et cetera could aid in recording more ground truth and enhance adaptability.
0:50 WOW 😊👍
That is totally cool!
It would also be really cool if they added two things to the training process where the person doing the training can speak and tell the AI what they are doing where that speech is also recorded and a camera records the video of the training process which could be used to to train the AI to be able to generalize a video of perform that task.
For example when putting the pot in the cabinet the trainer can say, I am now putting the pot in the cabinet on the bottom shelf using both hands and thus one can train a chat GPT while also training the "muscle memory" where the AI can then link the two together and not only learn to perform a task but also learn how to instruct others in words how to perform it as well as build a textual representation of the robot's library of tasks it knows how to do and maybe even build an ontology model as well.
BTW, as to how to do the later I have made a number of videos and uploaded them to my UA-cam channel on a way to do that which I call UniML, Universal Modeling Language.
So in UniML speak, the trainer's words could be treated as Swim Lane Islands (SLI's) which the can be assimilated (gobbled up) into RDF like statements as well as can be used to train a chat bot.
So then one would have a number representations, the recorded motions, the AI training data derived from them, the recorded speech and video of the training for that task, the importation of that into UniML as SLIs which it then assimilates (which could be done by its own Chat bot that has been trained to do that), and probably more.
These then could be used to generate training videos and documents as well as be added as part of a world model representation of what is doable as well as how one can do it.
BTW, co creation of the Ontology model might then be thought of as an artificial consciousness (thoughts) while the AI that can perform the task as muscle memory.
Thus when the robot is putting the pot on the bottom shelf it can "think" I am no putting the pot on the bottom shelf using both hands because the task was linked to the words via the ontology model.
It could then take that sentence as use that as a prompt for its own chat bot to see what that might produce like be careful to make sure it is clean before doing that.
The the ontology model serves not so much as a bot but as the human that ask or tells the bot things, i.e. does the prompt engineering base on its ontology models, that could have been in part created by a chat GPT.
It is sort of like the left brain conversing with the right brain and vice versa.
BTW, the another reason to do all this is to facilitate sharing where one person might train their robot to put a pot on the shelf as well as some other person might to the same but with the verbal descriptions added and the ontology model creation they could share these on a network that might then combine the two into a fused AI that uses both sets of training data.
Indeed that is one of the uses of ontology models is to allow different system to be interoperable in that the ontology model serves as a data broker between the two so they can know what the other system's data represents.
And there are such Ontology schemes such as OWL and RDF where in UniML I try to extend that to be more graphical in nature, like going from DOS to a GUI like how Steve Jobs did with the Macintosh from DOS on a PC back in the day.
Also during training of the robot, half of the training should be by a left handed person, the other half by right-handed .. That way, the robot may excel above and beyond either of them ..
Remote control magic
Using a mobile desk for the bottom is brilliant.
This is how the Daleks started ... just needs a plunger attachment. 😀
Would love for you to go through technical papers and AI research more, to have something else than doing new model reviews etc.
Absolutely engaging and convincing proof of concept that can be built on a weekend. Amazing. The future is so promising.
Industrial robots have been doing this for decades - repeating actions trained by a remote operator. I'd like to be enthusiastic about this, but it really just looks like 90% teleoperation, 10% repetition. If you pay attention, you'll notice that all of the demos are performed using teleoperation - in some, you can clearly see the operator in the background. You'll also notice the complete absence of adverse or unexpected circumstances from the demos. In the absence of 3rd party testing, skepticism should be the default.
One of the major appeals of this isn't necessarily the training method, but the fact that this is open source and fully accessible. Functionality can be added in and modified rather rapidly.
It doesn't matter if the pilot has to do some work initially, that's the whole point. Skepticism is fine, but don't be so biased that you become hyper-skeptical.
There a re clearly no operators visible in the main demos (I just watched again to be sure). The footage at the end of the video however shows the training side of things, where people are obviously required to be operating it.
Fair enough. If it's real, then we should expect to see this replicated and improved very rapidly given it's open source. If it's vaporware, then we should expect everyone to forget about it after six months or so. I don't think it's hyper-skeptical to suspect the latter, but I'd also be pleasantly surprised to be wrong. Either way, we'll know soon enough.
Called it. Two months later and there's been zero activity on this project's Github pages. Just another staged demo of a vaporware project.
bimanual manipulation is a very tricky task. HHH AI is a game changer in robot control. We'll all be put to retirement soon. To do that using classical robotics, it is extremely complex. We have built a bimanual robot and i know how hard it is. And they are very expensive. Waw, AI did replace many methods and Robotics are not an exception.
I want to buy those equipments and make one at home, for my son, who has problems with his hand and arm. I think with a certain amount of time on learning, the machine can make a massage daily. That'll be great helpful.
wow, fully open source, something is starting there! I can see a lot of opportunity in cost reduction. Also it should be possible to train remotely with a virtual headset. Train it slow then accelerate on execution. Lets build a training dataset from the crowd! Some computation can be offloaded or improved by cloud computation. But it needs safety, it is not safe to launch an autonomous robot with humans. It may break bones, it may start fire... Actuators need to have a safe force limit and human supervision.
5 years? Maybe not. Consider the pressure Mobile Aloha just put on the robotics industry as a whole to push for even better breakthroughs. Musk has to know about Mobile Aloha and I don't figure he will stand idly by while it outperforms Optimus.
...jesus christ its like a toddler trying to cook the shrimp...im good
Yet more sped up speed cut videos of something robotics experts first managed 15 years ago.
Don't really want a Macbook Grope rolling around. I definitely want a Spot robot dog though
OMG! this is what we had been waiting since the 80s!!! thanks god
"Only $32,000" Man I wish I was American if they think that something that costs so much is not expensive.
On another note. The robot leaves the frying pan handle over the edge of the cooker, which everyone with kids knows is something you don't do. Also it puts the hot frying pan down on the worktop.
Imagine giving over your entire existence of being human to a robot so you can sit on your couch all day and play video games. The world looks bleak.
So when will I get a robot servant to do all household chores upon my instruction in natural language?
But I already watched this video from Wes Roth >_<
Also, currently, despite the cherry picked examples, it's really not ready for use, teleoperation cool and all but not something affordable to deploy and it's few shot learning is not up to snuff yet.
They had me at doing your dishes.
Best proof of concept I've seen for Tesla's Optimus gen 2 bot in terms of use cases. They'll mass manufacture it for cheap enough eventually. Invest wisely
Imagine having a personal robot constantly chastizing you about toxic habits and imaginary privilege.
Those are all hardcoded routines though, right?
I don't see what exactly should cost 34k here. 4 motors? Plastic? What exactly? 2k at best. Unless you want to some gems inside of it and few layers of gold.
Left arm: 7k, right arm: 7k, battery: 500, laptop: (it has to be a good one) guess
@@HUEHUEUHEPony full car costs like 2k
The base is $7000, too.
But the base could be built DIY for less than 1/10th of that.
@@BillTrowbridge but it is scary. This setup can execute any crime.
I'm sure someone will build this testing out cheaper parts to see how bad it would do. Great idea to keep it open source so people can iterate over the plans. I think even if you modded it to be human controlled if you were disabled in various ways it could be a great help if it was affordable.
"Mock-Up"
A useful word, that you should know for the forthcoming cycle of development and hype in the AI field. It is basically a fake, but use legitimately it's intention is to give everybody an idea of what the final version should look like and how it should function
Basically it generally means, producing something that appears to function the way you intend something else to function. If you're making a program, or doing an experiment, or designing the scenery for play, or creating an invention to automatically function. You will create something that appears to be the real thing so long as you look at it from just the right angle, with the right editing, and there's a lot of predictable events or human interactions.
So for example, let's say you're writing a program for a calculator. You might write a much simpler program that looks like a calculator, and always prints out the number "5" when you hit '=' and then during the mock-up demonstration you type in 2+3 = and it prints out 5 even though no calculations have been done.
There are legitimate uses for mock-ups. Basically you can use them to make sure that everyone on your development team knows what you're going for. You can inform investors as to your goals and give them a practical demonstration of what you want it to do. You can test aspects of an experiment to make sure that it can actually detect the result of the experiment correctly, and I know they in fact did this with the gravitational observatories, they fed it fake data to see if it could actually identify merging black holes. I've done it in biology, where I've fed in fake experimental results into a gel and into a sequence analyzer to see if it could detect the real thing when I ran the experiment. This is sometimes called a positive control..
It goes from a mock-up, to fraud when you represent it as real results obtained from natural inputs. If I say look my calculator can add two numbers, two plus three equals and a five pops out, I am misleading people as to whether or not actual addition is occurring. If I had presented my positive controls in my biology experiments as if they were the result of an experimental intervention rather than me custom designing DNA fragments to give a particular appearance, that would have also been fraud.
Unfortunately fraud like that is really hard to detect unless you have independent people that have access to the same things and can verify your results as well as test for the assumed full range of applications.
I do think that some of what Google did qualifies as mockup that they then mixed with semi legitimate stuff, creating what is properly called a fraud.
Mockups are actually really useful tools and represent an organic part of the research and development process. They are also very easy to translate into fakery and fraud especially when there's a lot of pressure to get results fast and there's a lot of money on the line. You need people with a lot of maturity and a lot of integrity to use mock-ups properly.
He's just covering an academic research project. It's not like that team is trying to convince people to buy these research platforms with pre-orders.
I get your point that it's not 100% proven and could turn out to be less capable or more limited than it looks. But they even showed the robot failing in the outro to drive home the point that it still fails sometimes right now.
THis is awesome especially for folks that need assisted living. Something like this robot gives the a degree of the dignity of independence
This is amazing work! How can you perceive depth with a 2D camera??
Robot is cooking the shrimp at too high a temp
These robots will implemented in basic settings like Janitorial positions, window-cleaning, and even McDonald’s!! To train the robots, you still need human interaction. This is implemented as a new position of ‘janitor’. So, to be a janitor in the future, YOU WILL NEED A COLLEGE DEGREE. THINK ABOUT THAT!!! FOOD FOR THOUGHT!!!! FUTURE, HERE WE ARE!!! 🎉❤
Ok , so it trains on certain situations. But what if the glass of wine or spillage is somewhere else? Is it smart enough to change its path and do it with the same accuracy? Do not think so... Still a cool project for sure!
Reminds me of BBT. An arm is all you need, right?
I would absolutely buy one to cleanup after me and serve me in my home.
What makes this Google's breakthrough? It's a Stanford project, right? Did I miss something?
Exactly the paper is not even mentioning Google...
In a year these will be pretty competent I imagine, the problem with robots right now is the poverty of data. We should build like a million of these and share our datasets.
This looks way better than Optimus 😂