Unitree H2 Destiny Awakening Skild AI Robot Parkour Gemini Robotics 1.5

That’s 19% more than the previous R1 model that was just released. And it’s demonstrating all of its agility as well as its active degrees of freedom while they’ve just released some of these new frameworks for the Unitree G1 which makes me wonder how much of this hardware is actually new. Now that we’ve seen this brand new trailer covered up sometimes with some of these clothes let’s take a closer second look at both the G1 and this new H2 side by side. So here the H2 on the left comparing the legs to the G1 on the right we could see there’s actually some new plastic molding the black piece inside of the hips of this H2 robot.

So comparatively it looks to be at least more compact hardware though it’s tough to say without knowing the internal specifications whether or not the hardware inside is really something novel or if it’s just a repackage. Then looking at the top of the robot it seems to be that most of everything is the same except for the shoulders of the H2 do have a slightly different shape but again no telling as to whether or not this is just a repackage more compact version of the G1. Tell me what you think below next we have the skilled AI brain.

This is a singular omni-bodied foundation model and it’s basically like physical AGI for humanoids look at the way that it jumps over knows just where to place its hands and its feet and this works across any robot type or task with no hardware lock-in like before and it can jump and tumble imagine this thing if it was holding a gun how it could probably protect maybe that’s later but here’s skilled AI hopping at its core it’s using a dual layer setup so first it’s a high level policy controlling the robot’s navigation and manipulation at a low frequency here you can see where it figures out how to join motions and torque and of course this is the same skilled AI that can operate across all types of form factors you can even mutilate some of these things like the robot dog and it still learns how to continue walking whereas in other tests they were using different form factors like humanoids and they were also figuring out how to make these things adapt in real time and these are never trained on other robots but they demonstrate fully emergent behaviors like here for instance when this robot breaks its knee it’s able to learn how to walk without it this is without programming another time they tried to disable the front legs of the robot and sure enough it figured out soon how to actually walk on its own this is all just with skilled’s AI generalizing figuring out here it’s standing on stilts so this is quite an impressive framework that’s pushing the next level of robotics towards really general behaviors that don’t have to be programmed step by step but here’s some more robot generalization and this is from Gemini robotics 1.5 here it’s generalizing to try to pick up this object from this woman’s hands although she’s not giving up the object so it has to figure out how to continue trying to grasp this away from her despite being unable to grab the intended object here and this is just testing to see how is this able to emergently figure out different multimodal strategies to pick out these items despite their shapes and Gemini robotics also showed off these multimodal hand strategies single finger push it’s using all different types of grabbing techniques here with just two fingers another with three next the hardest is five finger dexterity and Gemini robotics 1.5 is also generalizing to use five fingers here a power grasp while not breaking open the object here it’s able to grab with all five fingers and put it in the basket next here it manipulates a slide by not pressing too hard just pushing it over it’s actually much harder than you would think it would be for a robot system to figure this out emergently here it grasps a cylindrical object and here’s a medium grasp using not all of its fingers it looks like just three which is quite an awkward grip but it does it without a problem here next an ambidextrous grasp where one finger is basically holding the bowl against the hand and the other is kind of positioning it with both hands working there together another bi-manual slide using one to hold it another to push two hands working together real time figuring this out and here’s a bi-manual rotation with two hands again so this is generalization happening right before you multimodal strategies and a counterclockwise rotation this is all just with text prompting telling this robot to do this and it figures it out and they also showed off the Apollo robot using Gemini robotics 1.5 doing unseen object manipulation with a bunch of clutter where maybe these different objects that it’s manipulating are occluded from its cameras by the mess in front but regardless it’s able to take these different text prompts drop the yellow sponge into the wicker basket and figure it out step by step here with just one hand but you saw before it can also do it with two and then of course there is the height generalization because the robot was trained with a bunch of tabletop data from humans so it’s able to take these different text prompts pick up the shirt from the top shelf and figure out exactly what it’s working with this is a word problem blue cereal box from the middle shelf here it says pick up the orange packet from the top shelf it has to figure this out in real time map it out to motor controls and one more rightmost pops box so it actually has to read it with an OCR system next it has to pick up the blue cereal box from the top shelf this is true generalization you haven’t seen this quite this level of generalization before but it’s happening before your very eyes with Gemini robotics 1.5 but how close is this to actually being translated into a commercial system that’s going to be in the home or in the office is this going to be later this year or is it going to be sometime next year but here’s a hand that’s even more dexterous than what you just saw with Gemini robotics 1.5 this is the sharper wave hand and this is a one-to-one scale human size hand with 22 degrees of freedom throughout i’m not sure about this exploded view is that actually all the components it doesn’t look like it’s a very scalable design if so but maybe you tell me in the comments whether or not you think that this is the legitimate build here we see the active degrees of freedom bending and stretching throughout each of the fingertips these are all of the active degrees of freedom ultra dexterous just like a human hand even with inward rotation apparently they use a proprietary neural network of a bunch of different modules for their tactile feedback system and the hand has over 6 000 discrete levels of pressure detection throughout and it can sense forces between 0 to 30 newtons per fingertip and ultrafine resolution down to 0.005 newtons of sensitivity and the system uses over 1 000 tactile pixels per fingertip
[tr:trw].