I guess the main takeaway is that they are working on manipulation of small objects. Something Boston Dynamic Atlas has not been seen doing and we see very little of from bipedal robots. They manage to move pretty fast but not crush the egg showing haptics, so if the robot accidentally would step on a cat tail or something it might be able to stop before the cat screams. Imo very impressive. And also this helps a lot when interacting with soft objects.
They show a basic feature extraction/detection algorithm used to learn the environment ie SLAM.
Note that most keypoints are in smaller inner square of the images. Maybe they just crop the image at this point. The images seems distorted likely optimized to operate on shorter distances than cars.
Here we can see the camera setup of the robot:
So basically 3 forward facing cameras, 2 sideways facing cameras and 2 downward facing cameras. No rear view, I would have thought that could be handy but I guess the robot will just turn the head/torso to make sure it will not hit anything if it's moving in that space. Seems like an Elon solution, the best part is no part. Anyway an interesting solution, God should take notice when he design humans 2.0.
Later we see some instruction:
Not unlike
this work... You can see that the human moves his legs when he is manipulating the objects, but the robot's legs remain fixed. So the robot has translated the movement to only use the upper body. We know that this method works, even just in small scale it quickly learns to master many tasks such as tying shoe laces. Tesla can scale this up and take it to the next level. The industry will need some time to process that this will happen and a lot faster than what most people think.
This image is a gold mine of information, we will see hours of youtube videos dissecting it:
Those kids are living the dream, getting to play with some very expensive toys! They show that the robot finally managed to grab the soft object by itself and celebrate so I assume this was one of the first times it actually worked. Then they need to teach the system to do it repeatedly, I would guess the success rate is an S curve that their data engine is built to solve.
Later they show a setup where the robot is more successful so I guess now the robot has become better at manipulating objects and is training itself:
Note the egg in the background. Clearly they thought eggs are very useful to test the system. Maybe an actual development goal is to be able to cook some eggs. Do I see a smiley face on the egg?! Maybe the egg is like a crash test dummy they are trying to protect, if they can avoid harming an egg then humans should be safe!
It seems like the robot is doing some slightly wasteful moves. I think for now they rely on imitating human moves from its movement library rather than optimizing for energy/time. Basically [how would a human return from this position to the start position], not [NN guided monte-carlo tree search to find the optimal path from start to end]. They will get there eventually, but for now it's imitation.
Finally we see the robots marching and there are more of them. I think that's the message they want to send. There robots are coming and there will be more and more of them: