Comment by rao-v
5 days ago
I’ve been playing a bit with toy robots and vision models and I can’t shake the sense that somebody has got to be close to simplifying the control game with high fps cameras. Why are there so many systems to control motors and ground positioning … shouldn’t it all be cameras on the arm figuring out if the motors have positioned the arm correctly and updating to fix?
This is obviously naive and computationally wasteful but I really want to see vision first rebuilds of most control and feedback systems.
Biological analogies are always a bit dodgy, but biology hints that this doesn't work very well. Vision is important, particularly for path planning, but the immediate feedback of proprioception and touch is more important for stability. Think of how hard it is to walk normally when your leg has fallen asleep --- your visual cues remain unchanged, but you lose the immediate feedback loop from your leg. You can do it, but your gait is very unnatural and inefficient.