This is huge. There's been so little progress on manipulation over the years. The 1960s Stanford AI lab videos don't look much worse than current work.
Part of the trick here seems to be using Gaussian splatting as an intermediate form for vision data. That's a new idea. Gaussian splatting is a brute-force process that doesn't try to generate a full 3D model reconstruction or object recognition. It can be applied to any visual scene where at least two viewpoints are available.
This just might get robotic manipulation unstuck. Good target for VC funding.
This looks like a nice application to get inductive bias into the model. But I think right now there is no solution to get fine grained motor skills besides tele operating and doing behavior cloning. And even then it is far from perfect..
Someday, maybe 100 years from now, someone is going to figure out a minimal codeset to "bootstrap" machine-learning for a wide range of devices so they basically teach themselves slowly but steadily and then each other
and at first it's going to be absolute amazing, followed by absolutely terrifying
The robots teaching themselves soccer (football) still sticks in my mind - at first they were falling over but after many iterations they had learned to pass the ball sideways and backwards to yet move it forwards to the goal
I'm surprised that you think this it going to be 100 years from now. I think 20 minutes from now would be a better estimate. Or 20 minutes ago; I don't know what people are doing alone in their nasty little labs and notebooks.
We are the apex predator. Once there is another entity capable of general purpose learning that is no longer the case. It's an undeniable existential threat.
It’s interesting that China has already thrown a Robot Olympics. We’re laughing about it, but I suspect it will get better and better every year. When China decides winning Gold in everything is a national imperative, watch out.
This is huge. There's been so little progress on manipulation over the years. The 1960s Stanford AI lab videos don't look much worse than current work.
Part of the trick here seems to be using Gaussian splatting as an intermediate form for vision data. That's a new idea. Gaussian splatting is a brute-force process that doesn't try to generate a full 3D model reconstruction or object recognition. It can be applied to any visual scene where at least two viewpoints are available.
This just might get robotic manipulation unstuck. Good target for VC funding.
Wonderful. Feed them combat videos for a secure future.
Shouldn't they be able to learn to use tools better than us? https://deepmind.google/discover/blog/genie-3-a-new-frontier...
Depends how you define that, but the biggest issue is fine motor control. Millions of years of evolution has made us a highly optimized machine.
This looks like a nice application to get inductive bias into the model. But I think right now there is no solution to get fine grained motor skills besides tele operating and doing behavior cloning. And even then it is far from perfect..
Someday, maybe 100 years from now, someone is going to figure out a minimal codeset to "bootstrap" machine-learning for a wide range of devices so they basically teach themselves slowly but steadily and then each other
and at first it's going to be absolute amazing, followed by absolutely terrifying
The robots teaching themselves soccer (football) still sticks in my mind - at first they were falling over but after many iterations they had learned to pass the ball sideways and backwards to yet move it forwards to the goal
I'm surprised that you think this it going to be 100 years from now. I think 20 minutes from now would be a better estimate. Or 20 minutes ago; I don't know what people are doing alone in their nasty little labs and notebooks.
Why is it terrifying to you?
We are the apex predator. Once there is another entity capable of general purpose learning that is no longer the case. It's an undeniable existential threat.
Watch the Animatrix
(no really, this is not a joke answer, it has some great philosophy among the entertainment)
Ever seen Terminator?
It’s interesting that China has already thrown a Robot Olympics. We’re laughing about it, but I suspect it will get better and better every year. When China decides winning Gold in everything is a national imperative, watch out.
But can a robot learn to drink 3 beers before lunch?
"Can you?"
In my corner of the world the accepted form is a 100ml bottle of vodka, so hopefully it's flexible in this regard.
[flagged]