It's a nice piece of machinery. Looks like they finally went all-electric and got rid of the hydraulics. The new model is much less bulky.
The task it is doing is undemanding. It's just moving things from one set of large slots to another. No need for precision placement, unstructured bin-picking, or object separation. If it could pick up engine covers from the messy pile seen atop one of the racks and slot them into the storage unit, that would be more impressive. It's cool to see this done with a humanoid, but off the shelf industrial robots could do that job. This is the same place where Rethink Robotics got stuck. They could do simple object movements in mostly-structured situations, but so can lots of other simpler approaches.
Amazon, despite substantial efforts, still doesn't have full robot picking. About two years ago, Amazon announced their "Sparrow" picking robot. But that seems to be experimental. It's not seen in videos of Amazon warehouses in 2024. Amazon is using the Agility humanoid, but only experimentally.[1]
This is how Amazon currently does picking.[2] Racks of product come to the picker on robotic platforms. The picking system projects a light square on the space in the rack from which the picker should take the product. The picker picks the item, waves it under a barcode scanner, and drops it in an outgoing bin. Repeat for 8 hours. The job requires no more than a room-temperature IQ. Machines should think. People should work.
Have you noticed how in second 13 it rotates the lower body clockwise while the upper body rotates counter-clockwise in order to optimize the movements to archive the goal?
All (bipedal/quadrupedal) robots I've seen behave either like an animal or a human, but this one is something else. Imagine the inverse kinematics required to perform such a movement, and how precise even the slight left foot's clockwise rotation is.
It does these very uncanny movements multiple times in the video, this video less about moving that thing from here to there than about how it does it. Also the recovery it had when it failed to insert one of the objects.
> Have you noticed how in second 13 it rotates the lower body clockwise while the upper body rotates counter-clockwise in order to optimize the movements to archive the goal?
Could be a fancy optimization of course, but could as well be a side effect of a decoupled planning of the locomotion part and the upper body pick and place.
Algo could be: plan a path for a lower body motion from pose (4dof) A to pose B. And given pose B plan the upper body place of the tray in the target. If no constraints are broken plan both in parallel so it looks like one smooth behavior.
There's something weird in its movement optimization, towards the middle it surprises the camera person by needlessly walking too close to it, a couple of steps away from the destination slots.
Nevertheless, it fascinating how the legs/torso/head spin in different directions. It's scary even if you know it is physically weaker than its hydraulic cousins.
On the shift from hydraulic to electric: for years Marc Raibert and others at Boston Dynamics expressed doubts that a non-hydraulic humanoid could be made to perform as well in that package.
Because of actuator density- a humanoid has a high number of actuators. Also power density, and they did a lot of work in molding the hydraulic actuators directly into the arms and legs to make them compact. The bulk is all in the backpack pumps, power pack, etc.
So this move to all-electric represents real growth in Boston Dynamics being able to make compact motors that meet their needs of performance.
I am betting this new robot is less capable of the explosive dynamics of the old one. I am not expecting dance performances and backflips from this one right away. But I am expecting them to work toward that kind of performance.
Hydraulics have a massive benefit that the actual actuators have amazing power density - and the actual bulky hydraulic pump and valves can be far away. Also, the pump can be shared with many actuators, making the whole system smaller and lighter when you have 50 actuators and only need to use 5 of them at full power at a time.
Electrics win for energy efficiency and no-leaky-fluid, and also are improving pretty quick.
> On the shift from hydraulic to electric: for years Marc Raibert and others at Boston Dynamics expressed doubts that a non-hydraulic humanoid could be made to perform as well in that package.
Then Schaft won the DARPA Robotic Challenge in 2013 with their all-electric humanoid robot with liquid-cooled motors.
They beat out Boston Dynamics' Atlas.
Then Google bought Schaft. Google got bored with robotics, couldn't find a buyer for them, and Schaft was shut down, along with several other Google robotics acquisitions.
The visualization of a 3d model of the engine cover and the points of the racking system highlighted as spheres in the visualization shows quite how much task-specific code there must be in this demo.
I don't think you'd be able to get this to move other objects or to other types of shelves without getting a programmer involved for a few days at least.
In [2], it's very surprising that the system is designed so that the worker must use steps to access bins on the transporter. Since this seems so obviously and counterintuitively bad for the workflow I assume there must be a very, very good reason for it. Would be interesting to know what it is.
That surprised me, too, because a previous version of the system used smaller bin stacks and no steps.
Amazon could have gone in the other direction. The layout could be designed so that the picker sits on a swivel stool and never stands at all. That's what they do when they test robot picking - there's one human-sized robot arm on a turntable mount doing the picking.
> The man whose whole life is spent in performing a few simple operations, of which the effects are perhaps always the same, or very nearly the same, has no occasion to exert his understanding or to exercise his invention in finding out expedients for removing difficulties which never occur. He naturally loses, therefore, the habit of such exertion, and generally becomes as stupid and ignorant as it is possible for a human creature to become.
Anyone who don't know what manufacturing automation looked like for past decades should go watch those "factory tours" video or two. Then write down what are shared characteristics of manually handled objects in it and what are deltas between those videos and this. There are just way too many fantasies around here.
At first, humanoids will not be used for these inflexible, old-style mass manufacturing factories. Humanoids can automate where flexibility is needed. E.g. a baker baking 5000 breads a day for a local community cannot afford the capital expenditure to automate every step in his pipeline with these ABB machines. He could afford to automate with one of these new humanoids though, as the same machine first does step 1, then step 2 all the way to the end of the recipe, in a kitchen which is already there.
Same idea with individual production. A toyota factory builds you one exact copy of a car in any colour you want (as long as it's black). But we might go back to a more individualised default in manufacturing, where for example tailor made clothes become affordable to most people again.
> to automate every step in his pipeline with these ABB machines.
That's still too much "inside the box" thinking.
That baker doesn't need "automate everything", in fact, no-one does.
What "that baker" needs, is to rank work on effort, cross sectioned with "costs to automate". And then automate the top item or items only.
The baker needs a machine to knead the dough. Special tools to cut the dough of twenty cookies at once with minimal waste. Trays and tools that can be cleaned in a dishwasher. A bread slicing machine.
A bakery that can be mopped in one go, rather than a fully automated humanoid mopping machine.
This is an analogy for all automation: we don't need to cover 100%, be fully autonomous, 100% flexible. We need to automate the hardest part, even if that's boring tech. Then the next hardest part. And so on.
IIUC, while this type of mass manufacturing is very efficient, it also limits the materials and what the final product can be.
Especially in food manufacturing, they need to add certain chemicals in the food ingredients so that the machines can process them and that's what separates the cheap mass produced food from artisan food.
A humanoid robot might be able to apply traditional processes and make artisan, additive free products cheap and available.
If we end up having practically unlimited energy(which we actually have, just not harnessed fully yet), we can also have small production centers on every corner that produce bespoke products using humanoid robots and traditional tools instead of having centralized and highly specialized and streamlined mass production.
Someone is drinking the 'bespoke artisan' koolaid.
Food additives are almost entirely about making products more visually appealing or shelf stable. Cellulose to keep cheese from sticking, nitrogen to keep meat from oxidizing, etc.
Your 'bespoke artisan' products are mostly improved marketing and improved staging (better lighting, less crowded stores, more personal service, etc)
what a relief it will be to be released from doing manual tasks for a living and to be able to finally engage every waking moment in an all against all struggle for power.
half of the population has IQ below average. what 'idea' are you talking about?
Actually that's getting interesting. For non-physical jobs we already got significant boost from LLMs. Robots will be another wave when they get cheap enough. For robot like on the video $20K price for mechanics looks achievable.
While humanoid robots are neat, if not uncanny, I would love to see robots with forms optimized for their work. A large octopus with legs would be ideal for this sort of parts handling job.
I think the exact point they’re trying to make is that a humanoid robot can pick up one shift exactly where a human left off. No custom robotics, no workplace changes, etc.
This is huge for the industry. Smarter Every Day visited a frisbee factory and they had automated a bunch of things. However, every automation point was extremely protected (fenced off) so that a bumbling human couldn’t walk somewhere and get a limb ripped off. If I remember correctly they joked that it was OSHA or something, which it turned out to be.
The effort required to change a process shouldn't be underestimated.
Especially considered that industrial environments are already (a) automated for lowest-hanging fruit things (e.g. moving stuff around at human height) & (b) optimized around human capabilities for the remaining things. Substituting a not-humanlike robot would require reconfiguring a lot of existing automation around it.
If you have "like a human, but costs less" that can be plugged into any existing still-human process? You can literally swap them in.
Eventually we'll get to hyperoptimized machines, but an easier sales story to say "We automate your existing human processes."
When steam and coal engines gave way to gas and electric engines in factories, it took decades before factories were reconfigured to adjust to the smaller sized engines that didn't require one major axle running through the entire factory. As a consequence the first gas engines were huge - over time they shrunk. I bet the same will happen with robotics, where humanoid will be the primary form factor at first for general tasks, then more efficient forms will emerge as processes are updated.
I think you're right, but the humanoid phase in manufacture will probably be an intermediate phase, for the reasons krunck points out. A 100% adapted to the function robot is where productivity is.
I keep hearing this but I just think it’s silly, for basically any factory, it’s trivial to change the configuration of a space so a better suited robot can operate rather than build robots that can fit into human sized spaces.
Yes there will be use cases but it just seems like a problem looking for a solution most of the time.
Optimization isn't efficient on the grand scale of things. Think of these robots more as exchangeable worker units. Lets say the warehouse isn't accepting many packages that day. So you don't need as many robots at the loading dock. But we have a lot of product that needs to be inventoried or shelved. Send a couple over and have them start doing it right away.
You actually end up running a warehouse with less robots because they can easily be repurposed for other duties.
Imagine you're working your 35-hour shift at the Fulfillment™ Center and all around you there's robotic eldritch horrors scurrying through the two-storey high shelves
Factory robots that move in general are kind of gimmicks, except for those roomba things for Amazon warehouses.
An assembly line with robotic arms has been standard for a long time now. And having many such arms working at the same time is normal. And each robotic arm will be doing one extremely narrowly defined task.
Anything involving autonomous judgment and mobility introduces uncertainty.
The legs do seem like a huge overcomplication. I can't think of many situations where they would be worth the added cost/complexity (compared to simple wheels). Sure they can walk stairs, but a place that employs freaking robots could easily make it robot accessible too, it would seem.
Did you see previous Atlas videos where this things jumps, does backflips and runs up stairs? The whole point is to make it operate in spaces designed for humans.
I liked the Tesla snake robot for charging.[1] But it upset many people.
It looks complicated but is a rather simple mechanical design. There are several motors in the base pulling on cables. The arm itself is just segments, disks, and cables. So all the complexity is in the base, and you can surround that with a metal or concrete box to prevent damage. Arm replacement wouldn't be too expensive.
Tesla is now touting wireless car charging, but that's a lot of power to transmit through air.
Bipedalism is great to evolve on uneven terrain. Here it seems to just slow the process. Also, it uses its second hand for balance instead of achieving work, a counter weight would be as effective. In a factory, where the floor is flat, a human sized self moving robot 2 or 3 wheels would be way more effective, longer arms with more joins that are not mimicking humans one could also be better. There is already a lot of automation/robots in industry and it never look like a human. Even in our houses the best approach to automation never look like a human (eg. vacuum cleaner). I think that the only part of our body that would worth copying is the hand.
Humans are precariously tall. Good for spotting threats but not for navigating terrain. Four legs are better, with a lower center of gravity. Look at mountain goats for example.
We already have highly specialized robots. They’re usually built for specific situations and are expensive and immobile. The goal here is to provide a robot that can fill as many roles as possible and the best template for that is humans around which the entire world has been built.
Or just a bunch of conveyor belts, magnets, clips etc. A "boring" but proven, optimized, sorting machine really.
Though the demo is awesome, the use case is bad. This "problem" of sorting and moving parts around, has long been "solved". It can be forever optimized, sure. But humanoid robots are definitely not that optimization.
One immediate use I can see for this robot is racking and energizing circuit breakers in electrical switchgear, which is one of the most dangerous things an electrician does. Arc flashes are very bad for humans, robots can be replaced easily.
I’m not talking about using hot sticks used for disconnecting and reconnecting live conductors, I’m talking about removing and replacing switchgear breakers for maintenance and testing from energized switchgear.
There is a risk of arc flash when you disconnect the bus fingers and also when you reconnect them to the bus, and a robot is easier to replace than a human.
The entire built environment to date is designed for humans
It’s the same reason why companies are throwing money at a self driving car which can coexist with manually driven cars, rather than building roads specifically for autonomy
The video is impressive, but how does it learn? How long does it take before it can do something new? I realize this is still very much a research project, but I'd love to understand how it works.
Happy to be corrected, but AI isn't mentioned in the title or video anywhere. What's with the expectation of learning new tasks? All this would be impressive regardless of that.
Whereas my current "robot", aka my dishwasher, has them done in 20 mins (or 90 mins eco setting).
Specialized automation is there, proven, efficient. It can be improved, sure (easier (un)loading?), but I doubt a generalized, humanoid robot is the best way towards these improvements.
Pretty sure my dishwasher takes a couple hours (maybe always on eco?). I could hand-wash them in 15 minutes though — I would just rather not.
I wouldn't really care if the robot was up all night doing dishes as long as they were done in the morning. And, you know, the robot did their work quietly.
Readit News
The task it is doing is undemanding. It's just moving things from one set of large slots to another. No need for precision placement, unstructured bin-picking, or object separation. If it could pick up engine covers from the messy pile seen atop one of the racks and slot them into the storage unit, that would be more impressive. It's cool to see this done with a humanoid, but off the shelf industrial robots could do that job. This is the same place where Rethink Robotics got stuck. They could do simple object movements in mostly-structured situations, but so can lots of other simpler approaches.
Amazon, despite substantial efforts, still doesn't have full robot picking. About two years ago, Amazon announced their "Sparrow" picking robot. But that seems to be experimental. It's not seen in videos of Amazon warehouses in 2024. Amazon is using the Agility humanoid, but only experimentally.[1]
This is how Amazon currently does picking.[2] Racks of product come to the picker on robotic platforms. The picking system projects a light square on the space in the rack from which the picker should take the product. The picker picks the item, waves it under a barcode scanner, and drops it in an outgoing bin. Repeat for 8 hours. The job requires no more than a room-temperature IQ. Machines should think. People should work.
Amazon keeps trying to automate that step.
[1] https://www.youtube.com/watch?v=q8IdbodRG14
[2] https://www.youtube.com/watch?v=wsIjagFEv84
Have you noticed how in second 13 it rotates the lower body clockwise while the upper body rotates counter-clockwise in order to optimize the movements to archive the goal?
All (bipedal/quadrupedal) robots I've seen behave either like an animal or a human, but this one is something else. Imagine the inverse kinematics required to perform such a movement, and how precise even the slight left foot's clockwise rotation is.
It does these very uncanny movements multiple times in the video, this video less about moving that thing from here to there than about how it does it. Also the recovery it had when it failed to insert one of the objects.
Could be a fancy optimization of course, but could as well be a side effect of a decoupled planning of the locomotion part and the upper body pick and place.
Algo could be: plan a path for a lower body motion from pose (4dof) A to pose B. And given pose B plan the upper body place of the tray in the target. If no constraints are broken plan both in parallel so it looks like one smooth behavior.
Obviously, I don’t know :)
Nevertheless, it fascinating how the legs/torso/head spin in different directions. It's scary even if you know it is physically weaker than its hydraulic cousins.
You mean the unnecessarily violent jerking motion it does for no great reason at 1:22 instead of gently trying again?
Because of actuator density- a humanoid has a high number of actuators. Also power density, and they did a lot of work in molding the hydraulic actuators directly into the arms and legs to make them compact. The bulk is all in the backpack pumps, power pack, etc.
So this move to all-electric represents real growth in Boston Dynamics being able to make compact motors that meet their needs of performance.
I am betting this new robot is less capable of the explosive dynamics of the old one. I am not expecting dance performances and backflips from this one right away. But I am expecting them to work toward that kind of performance.
Electrics win for energy efficiency and no-leaky-fluid, and also are improving pretty quick.
Then Schaft won the DARPA Robotic Challenge in 2013 with their all-electric humanoid robot with liquid-cooled motors. They beat out Boston Dynamics' Atlas. Then Google bought Schaft. Google got bored with robotics, couldn't find a buyer for them, and Schaft was shut down, along with several other Google robotics acquisitions.
[1] https://www.darpa.mil/program/darpa-robotics-challenge
I don't think you'd be able to get this to move other objects or to other types of shelves without getting a programmer involved for a few days at least.
Amazon could have gone in the other direction. The layout could be designed so that the picker sits on a swivel stool and never stands at all. That's what they do when they test robot picking - there's one human-sized robot arm on a turntable mount doing the picking.
Humans would get fat with that layout.
- Adam Smith
1: https://www.youtube.com/watch?v=zQeQWGqfFN0
2: https://www.youtube.com/watch?v=csifG1AM5d8
3: https://www.youtube.com/watch?v=wg8YYuLLoM0
Same idea with individual production. A toyota factory builds you one exact copy of a car in any colour you want (as long as it's black). But we might go back to a more individualised default in manufacturing, where for example tailor made clothes become affordable to most people again.
That's still too much "inside the box" thinking.
That baker doesn't need "automate everything", in fact, no-one does.
What "that baker" needs, is to rank work on effort, cross sectioned with "costs to automate". And then automate the top item or items only.
The baker needs a machine to knead the dough. Special tools to cut the dough of twenty cookies at once with minimal waste. Trays and tools that can be cleaned in a dishwasher. A bread slicing machine. A bakery that can be mopped in one go, rather than a fully automated humanoid mopping machine.
This is an analogy for all automation: we don't need to cover 100%, be fully autonomous, 100% flexible. We need to automate the hardest part, even if that's boring tech. Then the next hardest part. And so on.
Especially in food manufacturing, they need to add certain chemicals in the food ingredients so that the machines can process them and that's what separates the cheap mass produced food from artisan food.
A humanoid robot might be able to apply traditional processes and make artisan, additive free products cheap and available.
If we end up having practically unlimited energy(which we actually have, just not harnessed fully yet), we can also have small production centers on every corner that produce bespoke products using humanoid robots and traditional tools instead of having centralized and highly specialized and streamlined mass production.
Food additives are almost entirely about making products more visually appealing or shelf stable. Cellulose to keep cheese from sticking, nitrogen to keep meat from oxidizing, etc.
Your 'bespoke artisan' products are mostly improved marketing and improved staging (better lighting, less crowded stores, more personal service, etc)
Actually that's getting interesting. For non-physical jobs we already got significant boost from LLMs. Robots will be another wave when they get cheap enough. For robot like on the video $20K price for mechanics looks achievable.
This is huge for the industry. Smarter Every Day visited a frisbee factory and they had automated a bunch of things. However, every automation point was extremely protected (fenced off) so that a bumbling human couldn’t walk somewhere and get a limb ripped off. If I remember correctly they joked that it was OSHA or something, which it turned out to be.
The effort required to change a process shouldn't be underestimated.
Especially considered that industrial environments are already (a) automated for lowest-hanging fruit things (e.g. moving stuff around at human height) & (b) optimized around human capabilities for the remaining things. Substituting a not-humanlike robot would require reconfiguring a lot of existing automation around it.
If you have "like a human, but costs less" that can be plugged into any existing still-human process? You can literally swap them in.
Eventually we'll get to hyperoptimized machines, but an easier sales story to say "We automate your existing human processes."
Yes there will be use cases but it just seems like a problem looking for a solution most of the time.
You actually end up running a warehouse with less robots because they can easily be repurposed for other duties.
An assembly line with robotic arms has been standard for a long time now. And having many such arms working at the same time is normal. And each robotic arm will be doing one extremely narrowly defined task.
Anything involving autonomous judgment and mobility introduces uncertainty.
Seems at the very least it could have little quad wheel things so it doesn't have to worry about balance as much.
It looks complicated but is a rather simple mechanical design. There are several motors in the base pulling on cables. The arm itself is just segments, disks, and cables. So all the complexity is in the base, and you can surround that with a metal or concrete box to prevent damage. Arm replacement wouldn't be too expensive.
Tesla is now touting wireless car charging, but that's a lot of power to transmit through air.
[1] https://www.youtube.com/watch?v=uMM0lRfX6YI
Deleted Comment
Though the demo is awesome, the use case is bad. This "problem" of sorting and moving parts around, has long been "solved". It can be forever optimized, sure. But humanoid robots are definitely not that optimization.
Racking 480V three-phase breakers: https://youtu.be/Rytjdqj_Img
There is a risk of arc flash when you disconnect the bus fingers and also when you reconnect them to the bus, and a robot is easier to replace than a human.
Why does it need to turn its head or it's torso?
Maybe just to make it more humanoid?
How much do you benefit from 360 vision vs front facing camera.
You also get some relative directional information as well with the current design.
It’s the same reason why companies are throwing money at a self driving car which can coexist with manually driven cars, rather than building roads specifically for autonomy
Specialized automation is there, proven, efficient. It can be improved, sure (easier (un)loading?), but I doubt a generalized, humanoid robot is the best way towards these improvements.
I wouldn't really care if the robot was up all night doing dishes as long as they were done in the morning. And, you know, the robot did their work quietly.