3 ms·

This analogy does a surprisingly good job of highlighting what is wrong with robotics right now as well. Robots are hard to control at a low level, and typical

This analogy does a surprisingly good job of highlighting what is wrong with robotics right now as well. Robots are hard to control at a low level, and typically you don’t want to. But sometimes you have to. We don’t have very good ways of specifying tasks or behaviors at different levels of granularity and having the robot figure out the right affordances to that tasking. Huh. I may have just figured out my next research proposal.

xg15 7y ago
Interestingly, we do exactly that when controlling our own bodies: You're typically able to delegate the exact mechanics of walking, holding things, speaking, etc to your subconscious "muscle memory" (what ever that is exactly) - however you'll take "direct control" when performing an unfamiliar movement (e.g. dancing or doing Yoga for the first time). Interestingly, the "direct control" is also gradual: You're able to position your hand very precisely without meeting to know where your arm is going. The system is even flexible enough that it can eventually learn unfamiliar movements and allow you to "delegate" them as well. (E.g. an experienced dancer or Yoga practitioner)
s_kilk 7y ago
Well, we do have a body of knowledge on those problems, as a sister-comment points out, Biology, and Cybernetics which abstracted the lessons from biology and generalized them to all machine-like systems. It's just that this body of knowledge is not really compatible with the kind of stuff that the hyper-rationalist cult of lesswrong are into, and so tends to be ignored. In fact, the author here is re-inventing Cybernetics from whole cloth (especially that of Ashby and Beer, in particular, Ashby's Law of Requisite Variety and Beer's Viable System Model), but without any engagement with the mass of existing knowledge. [EDIT: I've just realized that the original author is being re-posted on lesswrong in this case, and isn't necessarily part of that milieu. I'd still encourage them (if they're reading) to check out Ross Ashby and Stafford Beer, because their intuitions align very strongly with the kind of research done by those Cyberneticians!]
LeifCarrotson 7y ago
What do you mean "Robots are hard to control at a low level"? That's the only level where they're easy! At the lowest granularity, they have servomotors, encoders/resolvers, and kinematics that convert motor rotation to a motion in 3D space. Robot manufacturers have made the code to make a joint or linear move to a predefined X/Y/Z/i/j/k location in space using those tools really quite simple. Inputs and outputs to control grippers and other end-of-arm tools are also straightforward. As long as you keep the process low-level, they're incredibly reliable and easy to program: a clever maintenance technician can pick up a pendant for the first time and have a well-scoped task programmed in a week. A competent sales engineer can clamp a small robot to a conference room table and demonstrate picking up and putting together your product in a one-hour sales demo. It's when you expand the scope when it gets really complicated: ask it to pick up a variety of objects, located by vision algorithms in arbitrary orientations and locations, that are moving in real time, in a changing environment, is simply a hard problem - and I think it always will be. Robots really only know how to move their tool center point from one location to another, and asking them to "do what I mean" and "pick up one of these and put it on one of those without making me worry about low-level stuff" is a level of simplification that has never really worked in any other domain.
- GlenTheMachine 7y ago
  It’s hard in two senses: one, it’s hard to translate low-level commanding into meaningful work in all but the most controlled of environments. That’s why robots that are economically viable today are nearly all on factory floors: that’s basically the only setting where you can control the environment well enough for them to be useful. And two, for “advanced” robots (quadrupeds/humanoids) it’s very hard to get them to even balance, much less walk. Yes, there are robots that do this today — surprisingly well, in fact. But there is a tremendous amount going on algorithmically under the hood to make it happen. Figuring out how to make them work took decades of research in controls. This is largely a restatement of what you said above, so I suppose it depends on what you mean by “low level”. My intent above was low level in horse terms, which includes standing and walking. Things humans think of as atomic behaviors.
gumby 7y ago
You might be interested in the Three Layer Architecture approach, e.g. http://www.flownet.com/gat/papers/tla.pdf http://www.flownet.com/gat/papers/tla.pdf

xg15 7y ago

Interestingly, we do exactly that when controlling our own bodies: You're typically able to delegate the exact mechanics of walking, holding things, speaking, etc to your subconscious "muscle memory" (what ever that is exactly) - however you'll take "direct control" when performing an unfamiliar movement (e.g. dancing or doing Yoga for the first time). Interestingly, the "direct control" is also gradual: You're able to position your hand very precisely without meeting to know where your arm is going. The system is even flexible enough that it can eventually learn unfamiliar movements and allow you to "delegate" them as well. (E.g. an experienced dancer or Yoga practitioner)

s_kilk 7y ago

Well, we do have a body of knowledge on those problems, as a sister-comment points out, Biology, and Cybernetics which abstracted the lessons from biology and generalized them to all machine-like systems. It's just that this body of knowledge is not really compatible with the kind of stuff that the hyper-rationalist cult of lesswrong are into, and so tends to be ignored. In fact, the author here is re-inventing Cybernetics from whole cloth (especially that of Ashby and Beer, in particular, Ashby's Law of Requisite Variety and Beer's Viable System Model), but without any engagement with the mass of existing knowledge. [EDIT: I've just realized that the original author is being re-posted on lesswrong in this case, and isn't necessarily part of that milieu. I'd still encourage them (if they're reading) to check out Ross Ashby and Stafford Beer, because their intuitions align very strongly with the kind of research done by those Cyberneticians!]

LeifCarrotson 7y ago

What do you mean "Robots are hard to control at a low level"? That's the only level where they're easy! At the lowest granularity, they have servomotors, encoders/resolvers, and kinematics that convert motor rotation to a motion in 3D space. Robot manufacturers have made the code to make a joint or linear move to a predefined X/Y/Z/i/j/k location in space using those tools really quite simple. Inputs and outputs to control grippers and other end-of-arm tools are also straightforward. As long as you keep the process low-level, they're incredibly reliable and easy to program: a clever maintenance technician can pick up a pendant for the first time and have a well-scoped task programmed in a week. A competent sales engineer can clamp a small robot to a conference room table and demonstrate picking up and putting together your product in a one-hour sales demo. It's when you expand the scope when it gets really complicated: ask it to pick up a variety of objects, located by vision algorithms in arbitrary orientations and locations, that are moving in real time, in a changing environment, is simply a hard problem - and I think it always will be. Robots really only know how to move their tool center point from one location to another, and asking them to "do what I mean" and "pick up one of these and put it on one of those without making me worry about low-level stuff" is a level of simplification that has never really worked in any other domain.

GlenTheMachine 7y ago
It’s hard in two senses: one, it’s hard to translate low-level commanding into meaningful work in all but the most controlled of environments. That’s why robots that are economically viable today are nearly all on factory floors: that’s basically the only setting where you can control the environment well enough for them to be useful. And two, for “advanced” robots (quadrupeds/humanoids) it’s very hard to get them to even balance, much less walk. Yes, there are robots that do this today — surprisingly well, in fact. But there is a tremendous amount going on algorithmically under the hood to make it happen. Figuring out how to make them work took decades of research in controls. This is largely a restatement of what you said above, so I suppose it depends on what you mean by “low level”. My intent above was low level in horse terms, which includes standing and walking. Things humans think of as atomic behaviors.

gumby 7y ago

You might be interested in the Three Layer Architecture approach, e.g. http://www.flownet.com/gat/papers/tla.pdf http://www.flownet.com/gat/papers/tla.pdf