We’re introducing an efficient, on-device robotics model with general-purpose dexterity and fast task adaptation.
Acknowledgements
We gratefully acknowledge contributions, advice, and support from Abbas Abdolmaleki, Saminda Abeyruwan, Joshua Ainslie, Jean-Baptiste Alayrac, Montserrat Gonzalez Arenas, Travis Armstrong, Maria Attarian, Ashwin Balakrishna, Yanan Bao, Clara Barbu, Catarina Barros, Robert Baruch, Nathan Batchelor, Maria Bauza, Lucas Beyer, Michael Bloesch, Michiel Blokzijl, Steven Bohez, Konstantinos Bousmalis, Demetra Brady, Philemon Brakel, Anthony Brohan, Thomas Buschmann, Arunkumar Byravan, Kendra Byrne, Serkan Cabi, Ken Caluwaerts, Federico Casarini, Christine Chan, Oscar Chang, Jose Enrique Chen, Xi Chen, Huizhong Chen, Hao-Tien Lewis Chiang, Krzysztof Choromanski, Adrian Collister, Kieran Connell, David D'Ambrosio, Sudeep Dasari, Todor Davchev, Coline Devin, Norman Di Palo, Tianli Ding, Adil Dostmohamed, Anca Dragan, Yilun Du, Debidatta Dwibedi, Michael Elabd, Tom Erez, Claudio Fantacci, Cody Fong, Erik Frey, Chuyuan Fu, Frankie Garcia, Ashley Gibb, Marissa Giustina, Keerthana Gopalakrishnan, Laura Graesser, Simon Green, Oliver Groth, Roland Hafner, Leonard Hasenclever, Sam Haves, Nicolas Heess, Brandon Hernaez, Tim Hertweck, Alexander Herzog, R. Alex Hofer, Sandy H Huang, Jan Humplik , Atil Iscen, Mithun George Jacob, Deepali Jain, Sally Jesmonth, Ryan Julian, Dmitry Kalashnikov, M. Emre Karagozler, Stefani Karp, Chase Kew, Jerad Kirkland, Sean Kirmani, Yuheng Kuang, Thomas Lampe, Antoine Laurens, Isabel Leal, Alex X. Lee, Tsang-Wei Edward Lee, Jennie Lees, Jacky Liang, Yixin Lin, Li-Heng Lin, Caden Lu, Sharath Maddineni, Anirudha Majumdar, Kevis-Kokitsi Maninis, Siobhan Mcloughlin, Assaf Hurwitz Michaely, Joss Moore, Robert Moreno, Thomas Mulc, Michael Neunert, Francesco Nori, Dave Orr, Carolina Parada, Emilio Parisotto, Peter Pastor, André Susano Pinto, Acorn Pooley, Grace Popple, Thomas Power, Alessio Quaglino, Haroon Qureshi, Kanishka Rao, Dushyant Rao, Krista Reymann, Martin Riedmiller, Francesco Romano, Keran Rong, Dorsa Sadigh, Stefano Saliceti, Daniel Salz, Pannag Sanketi, Mili Sanwalka, Kevin Sayed, Pierre Sermanet, Dhruv Shah, Mohit Sharma, Kathryn Shea, Mohit Shridhar, Charles Shu, Vikas Sindhwani, Sumeet Singh, Radu Soricut, Andreas Steiner, Rachel Sterneck, Ian Storz, Razvan Surdulescu, Ben Swanson, Mitri Syriani, Jie Tan, Yuval Tassa, Alan Thompson, Dhruva Tirumala, Jonathan Tompson, Karen Truong, Jake Varley, Siddharth Verma, Grace Vesom, Giulia Vezzani, Oriol Vinyals, Ayzaan Wahid, Zhicheng Wang, Stefan Welker, Paul Wohlhart, Chengda Wu, Markus Wulfmeier, Fei Xia, Ted Xiao, Annie Xie, Jinyu Xie, Peng Xu, Sichun Xu, Ying Xu, Zhuo Xu, Yuxiang Yang, Rui Yao, Sergey Yaroshenko, Matt Young, Wenhao Yu, Wentao Yuan, Martina Zambelli, Xiaohua Zhai, Jingwei Zhang, Tingnan Zhang, Allan Zhou, Yuxiang Zhou, Guangyao (Stannis) Zhou, Howard Zhou.
We also thank the operations and support staff that performed data collection and robot evaluations for this project.
Read the original article
Comments
These are going to be war machines, make absolutely no mistake about it. On-device autonomy is the perfect foil to escape centralized authority and accountability. There’s no human behind the drone to charge for war crimes. It’s what they’ve always dreamed of.
Who’s going to stop them? Who’s going to say no? The military contracts are too big to say no to, and they might not have a choice.
The elimination of toil will mean the elimination of humans all together. That’s where we’re headed. There will be no profitable life left for you, and you will be liquidated by “AI-Powered Automation for Every Decision”[0]. Every. Decision. It’s so transparent. The optimists in this thread are baffling.
MIT spinoff Google-owned Boston Dynamics pledged not to militarize their robots. Which is very hard to believe given they're backed by DARPA, the DoD/Military investment arm.
By arcticfox 2025-06-253:16 This pledge would last five seconds in an actual conflict, if it makes it even that far.
By jagger27 2025-06-2419:42 Militarize is just bad marketing. Call them cleaning machines and put them to work on dirty things.
By paxys 2025-06-2419:48 Was owned by Google. Then Softbank. Now Hyundai.
By bamboozled 2025-06-2422:13 How would these things be competitive with drones on the battlefield? They probably cost the equivalent of 1000 autonomous drones and 100x the time and materials to make, way more power would be required to make them work too.
Terminator is a good movie but in reality, a cheap autonomous drone would mess one of those up pretty good.
I've seen some of the footage from Ukraine, drones are deadly, efficient, they are terrifying on the battlefield. Even though those robots will get crazy maneuverable, it's going to be pretty hard to out run an exploding drone.
Maybe the Terminators will have shotguns, but I could imagine 5 drones per terminator being a pretty easy to achieve considering they will be built by other autonomous robots.
By JumpCrisscross 2025-06-2420:051 reply > These are going to be war machines, make absolutely no mistake about it
Of course they will. Practically everything useful has a military application. I'm not sure why this is considered a hot take.
The difference between this machine and the ones that came before is that there won’t have to be a human in the loop to execute mass murder.
By m00x 2025-06-256:44 There's a clear task being given to the robot. If anything this will save lives. There are plenty of soldiers that love to kill for the hell of it, at least this will be easy to track down to who gave the order.
By JumpCrisscross 2025-06-2423:07 > there won’t have to be a human in the loop to execute mass murder
This looks like an increasingly theoretical concern. (And probably always has been. Wars were far more brutal when folks fought face to face than they are today.)
By m00x 2025-06-256:42 Good!
I’m optimistic about humanoid robotics, but I’m curious about the reliability issue. Biological limbs and hands are quite miraculous when you consider that they are able to constantly interact with the world, which entails some natural wear and tear, but then constantly heal themselves.
Industrial robots at least are very reliable, MTBF is often upwards of 100,000 hours[0]. Industrial robots are optimized to be as reliable as possible because the longer they last and less often they need to be fixed, the more profitable they are. In fact, German and Japanese companies came to dominate the industrial robotics market because they focused on reliability. They developed rotary electric actuators that were more reliable. Cincinnati Millicron(US) was out competed in the industrial robot market because although their hydraulic robots were strong, they were less reliable.
I am personally a bit skeptical of anthropormophic hands achieving similarly high reliability. There's just too many small parts that need to withstand high forces.
[0]https://robotsdoneright.com/Articles/what-are-the-different-...
By ragebol 2025-06-256:05 If you E-stop an industrial robot, it stops immediately, all OK. If a humanoid were to freeze like that, it would fall over and hurt you and your stuff on the way down, when it'll damage itself.
Mechanical reliability is not the main concern IMO
By marinmania 2025-06-2417:391 reply It does either get very exciting or very spooky thinking of the possibilities in the near future.
I had always assumed that such a robot would be very specific (like a cleaning robot) but it does seem like by the time they are ready they will be very generalizable.
I know they would require quite a few sensors and motors, but compared to self-driving cars their liability would be less and they would use far less material.
The exciting part comes when two robots are able to do repairs on each other.
By marinmania 2025-06-2419:231 reply I think this is the spooky part. I feel dumb saying it, but is there a point where they are able to coordinate and build a factory to build chips/more of themselves? Or other things entirely?
By bamboozled 2025-06-2421:44 Of course there is
By pryelluw 2025-06-2419:00 2 bots 1 bolt ?
But this still has a massive cost. Replacing or repairing an actuator isn't cheap, in material and in time of unavailability.
To maybe get a little carried away with the sci-fi for a minute, why does the Actuator need to cost anything?
When the tree of costs that make up a product are traced, surely all the leaf nodes are human labour? As in, to make the actuator, I had to pay someone to assemble it and I had to buy the parts. Each part had a materials cost and a labour cost. So it goes for the factory that made the fasteners, the foundry that made the steel, the mine that extracted the ore.
Shudder to think of how to regulate resource extraction in a future where AI humanoid robots are strip mining and logging for free.
> When the tree of costs that make up a product are traced, surely all the leaf nodes are human labour?
What about energy, real estate and taxes?
Even at the extreme end of automation, if you want iron ore, you need to buy a mine from somebody, pay taxes on it, and power the machines to extract the minerals and transport them elsewhere for processing.
The same logic applies to energy I think. We don't have to pay money to a wind turbine, or to a coal mine. We only pay money to humans to build the power plants and the grid.
If I were writing a sci-fi novel about this I don't know how I'd handle something real estate (or mineral rights or water rights). You already need permission from the government to extract resources.
As for taxes, why does the government even want the money? What are they going to do with it?
By david-gpu 2025-06-2615:08 Energy, ultimately, requires real estate --and thus property taxes-- even at the logical extreme of automation.
> As for taxes, why does the government even want the money? What are they going to do with it?
There are websites that break down how e.g. different national/federal budgets are divvied up in the real world. Alternatively, I suggest a good book on macroeconomics; I am partial to Steve Keen's "Debunking Economics", but there are many others.
By UltraSane 2025-06-2419:28 Consumable components could be automatically replaced by other robots.
By bamboozled 2025-06-2423:15 I'm interested how differences with robots work overtime, there are a lot of machines in this world that have been patched or "jimmied up" to continue working, let's say a mining robot, it would probably get quite heavily contaminated with dust, wear would occur in different places, rock falls might bend parts.
So even though another robot could probably do the "jimmy up". it seems like overtime, the robots will "drift" into all being a bit different.
Even commercial airlines seem to go through fairly unique repairs from things like collisions with objects, tail strikes etc.
Maybe it's just easier to recycle robots?
I think those problems can be solved with further research in material science, no? Combined that with very responsive but low torque servos, I think this is a solvable problem.
It's a simple matter of the number of motors you have. [1]
Assume every motor has a 1% failure rate per year.
A boring wheeled roomba has 3 motors. That's a 2.9% failure rate per year, and 8.6% failures over 3 years.
Assume a humanoid robot has 43 motors. That gives you a 35% failure rate per year, and 73% over 3 years. That ain't good.
And not only is the humanoid robot less reliable, it's also 14.3x the price - because it's got 14.3x as many motors in it.
[1] And bearings and encoders and gearboxes and control boards and stuff... but they're largely proportional to the number of motors.
The 1%/year failure rate appears to just be made up. There are plenty of electric motors that dont have anywhere near that failure rate (at least during the expected service life, failure rates certainly will probably hit 1%/year or higher eventually).
For example, do the motors in hard drives fail anywhere close to 1% a year in the first ~5 years? Backblaze data gives a total drive failure rate around 1% and I imagine most of those are not due to failure of motors.
Yes, obviously that 1% figure is a simplification. Of course not all motors are created equal, and neither are all operating conditions!
But the neat thing about my argument is it holds true regardless of the underlying failure rate!
So long as your per-motor annual failure rate is >0, 43x it will be bigger than 3x it.
By mewpmewp2 2025-06-275:56 Uuids have failure possibilities, yet they are used very successfully. It is all about the failure rate.
By mrheosuper 2025-06-258:48 your calculation is true, but the absolute number is needed here.
43x of 1% failure rate is tragic, but 43x of 0.1% is acceptable in my book.
To an extent, yes.
For example, an industrial robot arm with 6 motors achieves much higher reliability than a consumer roomba with 3 motors. They do this with more metal parts, more precision machining, much more generous design tolerances, and suchlike. Which they can afford by charging 100x as much per unit.
By bamboozled 2025-06-2423:173 reply Also factory robots arms are probably operating in highly sterile, dry environments? How would working in a muddy / dusty / wet environment change this?
By Robelius 2025-06-254:09 When designing hardware, you usually define what the expected operating environments are. Some typical environmental considerations are the min/max temperature, debris ingress, shock & vibration. If you know your product is going to operate in an area where material is likely to enter the product, then you can either try to keep that material out (sealing the product up), or make sure that dust entering the product won't cause failures (i.e. electrical shorts won't happen on a board by covering exposed areas with glue or making sure a mechanism can crush/clear particles). It's not necessarily more complexity in the product to navigate these constrains, but it is another thing to consider in the design.
For example, if you're making a phone that is going to be sold around the world, then you're going to worry about arctic/equator temps (will some of your components melt or ICs fail), salty sea air (will the product begin to corrode for people living by a beach), or fast moving elevators (will the speakers pop from a sudden change in pressure).
You can check out this manufacturers robot arms as some examples of existing products. They list some data sheets for their robot arms, including some arms that are IPxx rated. I don't think looking at robot arms is a 1to1 comparison for what you could expect from a humanoid robot since the considerations in the design process are going to be different.
website is kuka dot com/en-at/products/robotics-systems/industrial-robots/kr-agilus
By michaelt 2025-06-256:37 Some are, some aren't.
For example, MIG welding robots tend to life a hard life. And if you look at photos of industrial painting robots, you'll find they're often fitted with plastic smocks.
If you look up photos online you'll only get marketing images from robot makers, where everything is shiny and brand new - I can assure you, it's not like that after they've been operating for a decade or two :)
Search for CNC videos, those machines work in oily, soapy, dusty and full of metal shavings environments and do fine.
By bamboozled 2025-06-254:40 It's still a fairly controlled environment with splash guards, liquid based dusts suppression and or dust collection, even then my friend has a factory and the CNC is a reliable machine but things screw up.
If the dust collection was disabled, the workshop and the machine would be caked in debris.
It doesn't move, it doesn't fall over or have anything falling on top of it either (like a robot could).
By chrsw 2025-06-2517:20 Yes but like almost everything it comes down to cost. Most consumer devices are extremely cost constrained. Industrial robots can justify higher costs that come with higher reliability.
By elcritch 2025-06-255:23 With more motors and joints also comes some degree of redundancy however. Having multiple fingers means one finger dying won't be as big of an impedement. It'd require feedback and the ability for the motion planner / AI to account for it.
Plus they'll likely be modular and able to be replaced.
IMHO, the bigger design issue for humanistic is lowering the need for mechanical precision which requires lots more metals and instead using adaptive feedback and sensors to obtain accuracy similar to how humans and animals do it. AIs should be really good at that, eventually. I think the compute will need to be about 10x what it is now though.
By Toritori12 2025-06-2415:441 reply Does Anyone know how easy is to join the "trusted tester program" and if they offer modules that you can easily plug-in to run the sdk?
By technotony 2025-06-250:11 There's a sign up button at the bottom of the article...
