These are all highly valid points and the going won't be easy.
The good news is that general purpose reasoning and path planning should scale quite easily in simulation. It is the precise manipulation and close contact where you have to take into account friction and the physics of the surfaces where the computation can be heavy and inaccurate.
But for the latter likely need to delegate to a physics-aware model anway that only worries about grip and not the big picture.
The other issue is that the hardware is expensive and wears out. So, won't be quick, but some of the work on general AI since 2020 or so (and availability of compute) will help a lot.
Something I wish I had found a space to include is that evaluating hardware is its own can of worms which is not done particularly well at the moment. That’s the thesis of the Amelia Michael piece I linked to in the next—it’s worth a read
These are all highly valid points and the going won't be easy.
The good news is that general purpose reasoning and path planning should scale quite easily in simulation. It is the precise manipulation and close contact where you have to take into account friction and the physics of the surfaces where the computation can be heavy and inaccurate.
But for the latter likely need to delegate to a physics-aware model anway that only worries about grip and not the big picture.
The other issue is that the hardware is expensive and wears out. So, won't be quick, but some of the work on general AI since 2020 or so (and availability of compute) will help a lot.
Something I wish I had found a space to include is that evaluating hardware is its own can of worms which is not done particularly well at the moment. That’s the thesis of the Amelia Michael piece I linked to in the next—it’s worth a read