00:01The new generation of ATLAS, it's a very exciting robot.
00:05When we started working on its design, we set out to build a robot that was designed for low-cost,
00:13ease of manufacturing, reliability, and simplicity.
00:16It really is a robot for which we've tried to reduce the complexity as much as possible.
00:22One of the cool things is that we've been able to, despite that simplification, maintain many of the things that
00:30make Boston Dynamics robots fun to work with.
00:33It's a robot that is very strong. It's a robot that is very agile and fast, and it's very easy
00:39to control.
00:40We're very successful at having fun in simulation designing behaviors that then almost always work with the real hardware.
00:49One of the goals that we've had in the last couple months since we've been working with this robot
00:54is basically being able to prove to ourselves that what we thought was possible with this robot is actually possible.
01:07So what we're seeing today is the new generation of hands for ATLAS.
01:11They're a perfect companion to where ATLAS is today.
01:14It's been designed with simplicity in mind and its trajectory to mass manufacturing.
01:20For the kinematics of the hand, we have increased the number of degrees of freedom.
01:24GR2 had 7 degrees of freedom. GR3 now has 13.
01:28The size and strength of the new hand is a little bit larger than an average human hand,
01:33but it's also, on average, a bit stronger.
01:36We went into the design process with an idea of the kinds of dextrous behaviors that we wanted to be
01:42intrinsically possible with these hands.
01:44So, for example, being able to achieve a pinch grasp that is dextrous,
01:48so like when you're grasping something small and rubbing it against your thumb,
01:52or being able to achieve a tripod grasp and you're holding something with some amount of strength,
01:58but you still want to be able to move it.
01:59Being able to grasp the handles of tools, but at the same time being able to press their trigger.
02:04So, a large variety of dextrous manipulative behaviors,
02:08we knew that were important in that roadmap of skills for ATLAS.
02:12I'm personally very excited about the fact that you can simulate it cleanly,
02:17and it enables very interesting work in reinforcement learning,
02:20but it's also close to anthropomorphic.
02:22It has a very small gap between human demonstrations and the embodiment of this hand.
02:28The reason we picked the tasks that we're looking at right now is because they're representative of the tasks that
02:33we really care about.
02:35This kind of object reorientation, both of these tasks require that.
02:39But if we're able to do these tasks, it kind of represents what we would be able to do with
02:45tools,
02:46or with pieces that we have to assemble together, which we actually do care about in the long run.
02:51So right now, reinforcement learning is used a lot for actually training the whole body,
02:55and we're trying to see how we can use those tools to actually get the robot to be able to
03:00use its arms and its hands to manipulate things.
03:03We've worked on getting these hands to be simulatable and also be transparent to force and motion.
03:12So any forces that the environment applies on these fingers get translated, and we can proprioceptively sense them.
03:19And any motion that we want to apply to the environment gets applied very transparently.
03:24Proprioception is a biological term, which means your ability to sense where your body or muscles are in space,
03:33if they're moving, how much force is applied to them.
03:37Backdrivability is not only beneficial in proprioception and being able to generate dexterous behaviors,
03:44but it's also really good in an impact.
03:46So when you hit something really hard, the joint is able to accelerate and move out of the way,
03:53whereas if it had a lot of inertia or a lot of friction, the gearbox would be damaged.
03:59So this is not only beneficial for performance, but also for ruggedness.
04:04When we want to actually simulate these actuators and this hand in a simulator,
04:10there's three things that we generally care about.
04:13The first one is kinematics.
04:15So what is the geometry of the hand?
04:18Where are the links?
04:19Where are the joints?
04:20And the second one is dynamics.
04:23Where we think about what is the friction in the system?
04:27What torques can it apply?
04:29What is the backlash in the system?
04:31And how do we simulate that?
04:32Lastly, we start worrying about contact dynamics.
04:36So how does this gripper interact with the environment?
04:39One of the things we think about when we are doing this sim to real match is not just going
04:46at it from one direction, but approaching it from both directions.
04:49We considered many, many designs, and there were many questions like, do we want the hand to have one thumb
04:56or two thumbs?
04:57Do we really want the hand to have a pinky or not?
04:59And most of these don't have straight answers, right?
05:02There's always some amount of intuition and experimentation.
05:34It boils down to a trade of different competing objectives.
05:36There's no pinky because the team determined that the additional dexterity and tasks you'd be able to accomplish is not
05:44worth the extra complexity of three additional degrees of freedom, the size, the power consumption, all of the things that
05:51come with adding additional actuators.
05:53The journey for hand is a long one, but this hand really set us in a good position for the
06:00work we have to do today.
06:01The multi-fingered hand is like a whole mini robot when you think about the number of joints and the
06:07number of motors in there.
06:08The hardest thing that we were working on this gripper, I was not expecting to do so much science in
06:14this job.
06:14I think I'm really excited for all the science that I get to do.
06:18We are actively pushing the state of the art on the Atlas team.
06:24Basically, Atlas needs to be a tool, a tool basically that is similar to humans, which means it needs to
06:30start interacting with objects and objects come in all shapes and forms.
06:35And because Atlas is a beast and we know that it's capable of lifting heavy objects, we basically wanted to
06:41jump in the deep end.
06:43The empty fridge is 50 pounds, and that is mostly what we shot with in the lab to really see
06:49what we could do.
06:50We were able to go up to 100 pounds, which was entirely unseen for the policy.
06:56It was always trained at about 50 to 70 pounds, and this is really pushing the limits of what humanoid
07:02robots can do.
07:03I don't know of any other humanoid robot that has moved a big, heavy object like a fridge.
07:09So the fridge experiment was really cool and something I've kind of always been wanting to do, which is basically
07:16instead of just grabbing things like with your hand and grasping it and kind of moving it like this, it
07:22was really using the whole body and really like leveraging like, you know, straight arms, lift up a thing and
07:28like really put your whole body into it was kind of the idea.
07:31And I've always wanted to do that with a robot, you know, a couple of things have come together recently.
07:36We have a really capable new robot that could actually do it.
07:39And we have a lot of really cool new tools on the team that we're utilizing to program it to
07:45do these kinds of behaviors.
07:53I've been here at Boston Amex for 15 years, and I've been a leader on the controls team for about
08:00that long.
08:01And we've gone through many cycles and many, you know, pivots.
08:06And I think this one where we are right now is probably the most exciting.
08:11In the six weeks, I've noticed that it's fairly easy for us to test things.
08:17There's a pipeline, there's a rigorous pipeline that's been established, and there's a lot of people working behind the scenes.
08:24The whole system in itself is a packaged system for us to test new policies on.
08:31So I could be training a policy today, and tomorrow I have a baked in policy, and then within like
08:37an hour, my policy is on the robot.
08:40I'm testing, I'm collecting data, I'm seeing what's happening, and I'm iterating and I'm getting back.
08:45So that whole pipeline is seamless.
08:47We kind of have a very close sim to real gap with just the robot.
08:53And that's been very fun to focus on and not worry about that aspect.
08:59We put together an animation, a reference of what it would look like for the robot to move the fridge.
09:05Then in simulation for millions of hours, we let the robot practice moving the fridge in the exact same way.
09:12We encouraged it to move the fridge like we told it, to not drop the fridge and all those sorts
09:17of things.
09:17And then once you get sort of the reward tuning right, the physics correct for the hand dragging along the
09:25fridge, and you run it on the hardware, it does perfectly.
09:29We have not seen the limits of what Atlas can do.
09:31We are still pushing the limits in terms of the sort of big heavy object that Atlas can pick up.
09:36We're always going to keep pushing it.
09:38Alright, backflipping. Three, two, one!
09:47Hey!
09:50We're like a champ!
09:51Put up!
09:52Put up!
09:54What can I do?!
10:25Untertitelung des ZDF, 2020
10:55Untertitelung des ZDF, 2020
11:07Untertitelung des ZDF, 2020
