If I think about robotic autonomy, I imagine a sort of layer cake:
1. Intent and Planning -> What am I doing, What will happen when I do it
2. Macro Implementation -> What are instructions down to all my parts
3. Local Implementation -> Given Instructions, let’s move my leg
4. realtime / reactivity -> OK, I’m 3 degrees off expected on this foot, need to rebalance everything.
Figure et al are demoing 1-4, but slow and without motion - just heads and hands.
Tesla’s robot day was demoing 4 only, as far as I can tell — engineers with controllers did all the macro planning and movement requesting - but there’s no way you can rebalance a robot live with a handheld controller - that seems certain to have been live from code.
BD has always existed in this kind of liminal space to me, where it’s not clear that they’re doing anything more than 4 with their demos. Extremely impressive mobility demos, to be clear. But, I’ve never seen a humanoid demo from BD that’s not in what looks a pre-scanned and planned room.
And we never hear claims about 1, 2 or 3 for humanoid robots from BD as far as I’m aware.
I think for years the low level mobility part has been so hard, BD has felt way ahead, the planning and environment responsiveness was just science fiction. Today we see that the top part of that layer can be gotten with vision models, and we’re seeing dual architectures where those vision models instruct small fast models for layers below. My guess is that this architecture is going to be relatively resilient, changing the cost structure and opportunities for robot companies.
So, I guess if I add all this up, we have a race between mobility experts and agile ML teams with robotic experience. And that adds up to some early launch products; I’d imagine they will be clumsy, and that BD will be a late launcher if it ever goes after consumer at all; it may just decide to license. Anyway I’d like to buy one, so the next few years should be interesting.
1. Intent and Planning -> What am I doing, What will happen when I do it
2. Macro Implementation -> What are instructions down to all my parts
3. Local Implementation -> Given Instructions, let’s move my leg
4. realtime / reactivity -> OK, I’m 3 degrees off expected on this foot, need to rebalance everything.
Figure et al are demoing 1-4, but slow and without motion - just heads and hands.
Tesla’s robot day was demoing 4 only, as far as I can tell — engineers with controllers did all the macro planning and movement requesting - but there’s no way you can rebalance a robot live with a handheld controller - that seems certain to have been live from code.
BD has always existed in this kind of liminal space to me, where it’s not clear that they’re doing anything more than 4 with their demos. Extremely impressive mobility demos, to be clear. But, I’ve never seen a humanoid demo from BD that’s not in what looks a pre-scanned and planned room.
And we never hear claims about 1, 2 or 3 for humanoid robots from BD as far as I’m aware.
I think for years the low level mobility part has been so hard, BD has felt way ahead, the planning and environment responsiveness was just science fiction. Today we see that the top part of that layer can be gotten with vision models, and we’re seeing dual architectures where those vision models instruct small fast models for layers below. My guess is that this architecture is going to be relatively resilient, changing the cost structure and opportunities for robot companies.
So, I guess if I add all this up, we have a race between mobility experts and agile ML teams with robotic experience. And that adds up to some early launch products; I’d imagine they will be clumsy, and that BD will be a late launcher if it ever goes after consumer at all; it may just decide to license. Anyway I’d like to buy one, so the next few years should be interesting.