Last year, we unveiled the new Atlas—faster, stronger, more compact, and less messy. We’re designing the world’s most dynamic humanoid robot to do anything and everything, but we get there one step at a time.
Our first task is part sequencing, a common logistics task in automotive manufacturing. Discover why we started with sequencing, how we are solving hard problems, and how we’re delivering a humanoid robot with real value.
Learn More: https://bit.ly/3DqQJh5
Our first task is part sequencing, a common logistics task in automotive manufacturing. Discover why we started with sequencing, how we are solving hard problems, and how we’re delivering a humanoid robot with real value.
Learn More: https://bit.ly/3DqQJh5
Category
🗞
NewsTranscript
00:00The core reason people are so interested in humanoid robots today is the promise of a
00:23truly flexible robotic solution that can switch between tasks and applications in a way that
00:28we just haven't seen before in robotics. The world was built for humans and the huge promise
00:34of humanoids is that they will be able to enter into those human built environments
00:39and immediately add value. The previous generation of atlas was able to do a lot of really athletic
00:43things from running around jumping flipping and so on and so that made it a really interesting
00:48research platform for us to explore the limits of whole body control but there are some downsides
00:54to hydraulics as well right they're very complex they're quite expensive they're difficult to
00:59maintain and they're messy our old lab used to have a thin coat of oil on it at all times
01:05and we got to the point in our humanoid development where we realized that using the
01:09battery technology that we developed on the previous generation of atlas and by building our
01:14own electronic actuators that we could build a humanoid that was strictly faster strictly stronger
01:21and more compact and less messy just manipulating objects is something that people
01:26spend very little time consciously thinking about we've had an entire lifetime of interaction with
01:32the world to train our control system so that we know how to predict what's going to happen
01:38and we know how to send commands to different parts of our body in order to achieve manipulation tasks
01:44robots by and large do not have that wealth of experience and part of the exciting work that
01:49is happening in ai and that we're doing ourselves as well is trying to use the rich data sources
01:55that we are generating now from all of our robots whether they're deployed out in the real world
02:00whether they're in our labs or whether it's being generated in simulation in order to train
02:05models that have a similar ability to generalize and a similar ability to error correct when trying
02:11to do fundamentally new manipulation tasks and i think that's one of the places where i would expect
02:15ai to significantly change how we perceive the capabilities of robots over the next couple years
02:25sequencing is a logistics task in the automotive manufacturing world that was created in order to
02:30allow manufacturers to produce many different customizations and variations of vehicles
02:36sequencing has commercial value but also from a research perspective from a building blocks
02:43perspective has a lot of the complexities that we think are necessary to be able to develop general
02:49purpose manipulation robots the sequencing process is taking thousands and thousands of different
02:55variants of parts and putting them into the specific order in which cars are going to be
02:59produced making sure that the cars can be produced more quickly and more accurately than they were
03:04before for this task we've given atlas a basic description of the end state that we want so we
03:12have this output dolly that needs to be filled with engine covers in a particular order and we've
03:17also given it a description of where it can look to find those engine covers once it has this it
03:23basically goes in order of the output dolly so it picks say the first slot in the output what kind
03:28of engine cover do i need and then consults where can i find that and then it goes and tries to pick
03:35it up from there and if it can find it it proceeds it has that kind of basic description and then it
03:40has a description of the kind of whole architecture of the task that we've provided and then it puts
03:46those two together to to do the sequencing atlas has cameras mostly in the head and with those
03:54cameras it perceives the fixtures or containers that it has to interact with not just the one in
04:00particular where it has to extract an object or insert an object but all containers around it so
04:05that it can up its footsteps to avoid collisions for example and then at the same time it also
04:10sees objects it has an idea of the object that it's going to manipulate and with that mental
04:16model it sort of uses directly rgb cameras to estimate the pose of the object and then equally
04:24track it over time sometimes it knows that it's successfully put in place so it knows that it's
04:29failed if it starts pushing and senses that it's not being able to move the cover the way it wants
04:34to and then it'll back up and try again it'll get the cover in and then let go and at that point it's
04:41it thinks that it has successfully put it in place it'll go ahead and give it a little extra push
04:45just to make sure it's fully seated the 360 rotations are funny they're there for a very
04:52particular reason which is efficiency right so atlas can turn around faster can turn its head
04:58it wants to look something behind its head and it can just turn the head atlas is faster at
05:04moving back and forth because it has that 360 motion boston dynamics has a very long history in
05:12robotics and in particular legged robots and in bipeds alone we've been working on them for
05:17probably over 15 years but we've really designed it to be a very high performance machine and so
05:22we're interested in again being able to explore the very frontier of what's physically possible
05:26in humanoids and everything all the infrastructure they've built around doing that i think is a real
05:31accelerant that shows itself through the performance of atlas already i'm biased for me
05:38it's the pinnacle of humanoids today it's capable of all of the agility that hd atlas had but
05:45probably even more than hd atlas
06:01you