The world’s most advanced humanoid robot, Boston Dynamics’ Atlas, is back, and it’s moving some medium-weight car parts. While the robot has mastered a lot of bi-pedal tricks like walking, running, jumping, and even backflips, it’s still in the early days of picking stuff up. When we last saw the robot, it had sprouted a set of rudimentary hand clamps and was using those to carry heavy objects like a toolbox, barbells, and a plank of wood. The new focus seems to be on “kinetically challenging” work—these things are heavy enough to mess with the robot’s balance, so picking them up, carrying them, and putting them down requires all sorts of additional calculations and planning so the robot doesn’t fall over.
In the latest video, we’re on to what looks like “phase 2” of picking stuff up—being more precise about it. The old clamp hands had a single pivot at the palm and seemed to just apply the maximum grip strength to anything the robot picked up. The most delicate thing Atlas picked up in the last video was a wooden plank, and it was absolutely destroying the wood. Atlas’ new hands look a lot more gentle than The Clamps, with each sporting a set of three fingers with two joints. All the fingers share one big pivot point at the palm of the hand, and there’s a knuckle joint halfway up the finger. The fingers are all very long and have 360° of motion, so they can flex in both directions, which is probably effective but very creepy. Put two fingers on one side of an item and the “thumb” on the other, and Atlas can wrap its hands around objects instead of just crushing them.
Atlas is picking up a set of car struts—an object with extremely complicated topography that weighs around 30 pounds—so there’s a lot to calculate. Atlas does a heavy two-handed lift of a strut from a vertical position on a pallet, walks the strut over to a shelf, and carefully slides it into place. This is all in Boston Dynamics’ lab, but it’s close to repetitive factory or shipping work. Everything here seems designed to give the robot a manipulation challenge. The complicated shape of the strut means there are a million ways you could grip it incorrectly. The strut box has tall metal poles around it, so the robot needs to not bang the strut into the obstacle. The shelf is a tight fit, so the strut has to be placed on the edge of the shelf and slid into place, all while making sure the strut’s many protrusions won’t crash into the shelf.
One limitation here is that at least some of the smarts in the video are pre-calculated—at one point we see what looks like Atlas’ vision processing, and it has a perfect 3D scan of the car strut ready to go. So this is either attempt-number 5,000 and it has already seen the strut from all angles, or Atlas was pre-programmed with topographical data for this exact model car strut. Either way, for all the lifts in the video, Atlas is saved from trying to figure out the shape of the object in real-time. Atlas has a Lidar sensor on its face and can generate a point cloud of what it’s looking at, so it just needs to line up the pre-baked model with the point cloud, and it has perfect knowledge of the strut topography. A harder level of difficulty would be picking up an object Atlas has never seen before, but you’ve got to break down the challenges into smaller parts and start somewhere.
When Atlas picks up a strut, it has to walk around a pallet, and as always, the robot shines when it comes to bipedal movement. The simpler way to move around the pallet would be a set of straight-line walking paths with pivots in between. Atlas’ path-planning is way more complicated, though, and involves more advanced side-step moves, leaning into turns, and just dynamically stumbling around the pallet any way it can. This version of Atlas moves less like a robot and more like a drunk person, which is a big compliment. At one point it even stumbles and recovers, drawing an excited reaction from onlookers in the background.