Humanoid Robot Shows Off New Advanced Hand

The transition from hydraulics to electric Atlas gave us an interesting jumping point. We knew that right about the same time that Boston Dynamics and Atlas wanted to focus more on manipulation, not just mobility and locomotion, and that was just a natural point to explore grippers with higher dexterity. We've known from the beginning that the gripper journey is a long one. Grippers are one of the most intricate components of a humanoid robot. They pack tons of functionality. We have to put in a lot of actuation and sensing into a very small space, so it's a very hard design problem. We've taken that long-term perspective and tried to maximize learnings through that path. GR1 is a gripper that is the first gripper of what we call the GR line, where we basically were looking to build something that had the most fundamental and, in a way, minimalistic capabilities of a human hand. GR1 allowed us to learn a ton of what it means to mount a gripper to a humanoid robot that works most of the time well, but sometimes it falls, and inevitably once in a while it's going to fall on top of the gripper. So we wanted to figure out what is a good way to build a gripper that can resist that kind of failure mode, which is very important to our development path, being able to build grippers that are rugged and reliable. This is our second generation gripper. It has seven degrees of freedom. We have seven different actuators within it, two for each of the three fingers, and then one for this articulated thumb joint. We have tactile sensing on the fingertips, and we have cameras here in the palm. The whole module is self-contained with all the actuation within it. It's very easy to put on and off of the robot, and it was an iteration from our first generation gripper, which had a similar three fingers, but there was no thumbs. They're all on the same line. Maybe the most important distinction from GR2 to GR1 is the addition of a thumb, an opposable thumb, that really expands the type of grasps that it allows us to do. It really increases the flexibility of which objects we can grasp with it, and really, over the last year or so, we've realized that we can grasp almost anything that we throw at it. We landed on three because we believe that that's the fewest amount of fingers that can achieve very complex manipulation tasks. So the fact that you do have three fingers brings a lot more to the table over just two, because the fact that there's some distance between them can help in maintaining a stable grasp, especially when you're manipulating larger or heavier objects that want to kind of rotate out of your hand. We do have the opportunity with this thumb to do some two-finger pinch grasps, which can be very helpful for delicate manipulation of very tiny objects, but that third finger coming into play when you are grasping something a little heavier or a little larger can be really helpful in maintaining a stable grasp. We've debated a lot, and we do wonder if there's an opportunity to add more fingers and whether they bring much to the table. We decided not to add more fingers because adding more fingers, it just adds more complexity that if it's not necessary, it always ends up meaning lower reliability, higher cost, just in general, lower speed of development. We shouldn't take that as a dogma, though, right? Three fingers was the right answer for the first phases of this program. We'll see what is the answer moving forward. Tactile sensing is force feedback. You can think of it as your sense of touch that you have in your own fingers. The tactile sensors live here on the fingertips in this elastomer here that's at the top that provides two functions. One is to give a high friction surface, and then also as it deforms, we have sensors within those fingertips that translate that into a force feedback. By having that sense of touch on your fingertips, you can apply very gentle forces to something. If you're perhaps manipulating a fragile object, you might want to handle it gingerly. You can also detect if you've dropped something. So if you didn't mean to, but it's fallen out of your hand, those tactile sensors can give you that feedback to understand what's going on. The goal is to apply as little force as possible, but maintaining a stable grasp. That's very similar to when you pick something up. You don't want to crush the object, but you want to make sure that it stays in your hand and it doesn't slip out of your hand. Too light a grasp and something might slip away, and too heavy a grasp and you might actually damage it. So inward, we can rotate up to 90 degrees, very similar to what your hand can do, but this has some extra human capability in that all these fingers can bend backwards completely. We can do some very clever things where you would actually be grasping on the back side of the gripper. So we actually have a left and right version of this gripper, similar to how our own hands, we have a left and right hand that are mirrored, but not the same. So this thumb always comes around on this one side. Atlas will usually be planning for the most optimal route to get the job done that it's trying to do. So if standing in a certain position and reaching in with the left hand will provide a more stable grasp or allow you to avoid objects in the environment, that's usually the way that it will choose. It doesn't tend to have a dominant left or right hand like you and I do. We're setting the next milestone in the direction of dexterity. It's becoming clear that tasks like bin picking or tool use or in general handling small objects are very prevalent in manufacturing. Interestingly, all of these things kind of are pushing us, not necessarily by design, but sort of naturally, organically, into more anthropomorphic designs of the gripper. Finding the gripper with the right sweet spot of dexterity, actuation and sensing is one of the most interesting journeys of the next few years and we're very excited to see where it takes us.