27 ms·

I'm surprised that a chapter titled "Intelligence and embodiment" (ch.1) would not include even a reference to Rod Brooks (MIT Mobile Robot Lab). Brooks' inven

I'm surprised that a chapter titled "Intelligence and embodiment" (ch.1) would not include even a reference to Rod Brooks (MIT Mobile Robot Lab). Brooks' invention of "subsumption architecture" (augmented finite state machines) and layered control systems in which higher level behaviors suppress low-level behaviors has been demonstrated to lead to robust intelligent behavior without centralized control, internal mapping, and emergent behavior. In short, a biologically-inspired solution which had tremendous influence in the field. To my mind, such an omission is a bit like writing a book on computer graphics and not including a reference to Watt, Foley or vanDam. Not saying it's wrong, just very surprising.

Symmetry 11y ago
An introductory textbook probably isn't the place to be introducing heterodox views. None of the commercial robotics companies I'm familiar with, including Rethink Robotics where Rodney Brooks works now, use subsumption architectures.
- gugagore 11y ago
  I agree with you. But Roomba.
  - Symmetry 11y ago
    Right, there is the Roomba. But don't the newer Roombas do mapping?
    - gugagore 11y ago
      I think that is true, but as you say that's the most recent model.
  - Animats 11y ago
    The Roomba algorithm is roughly: - 1. Go in an increasing spiral until you hit something. - 2. Wall follow for a while. - 3. Turn away from the wall and go a few meters. - 4. Go to step 1.
Animats 11y ago
Brooks at MIT, and Latoumbe at Stanford, headed the two opposed camps during the AI Winter. Brooks wanted purely reactive robots, and did some nice insect robots. Latoumbe was into very rigid planning systems, where you approach manipulation as path planning using maze solving algorithms in a N-dimensional space with obstacles. Both turned out to be dead ends. Purely reactive systems don't get beyond insect level, and high-dimensional planning requires total information about the environment. Then the statistical machine learning people took over and started to get real results, especially in sensor data reduction.
- eikenberry 11y ago
  Neither was a dead end. Each made real contributions to the field and have real applications but, not surprisingly, turned out not to be the silver bullet that could do all things. The statistical crowd was always there and have been good at the same set of things, it just finally matured as a field and was able to produce usable results. They all go through the same motions as anything in this industry. They show promise, everyone gets all worked up about the new thing, it produces some results and gets integrated where it makes sense, gets boring and people start looking for the next thing.
robotresearcher 11y ago
Subsumption was also fast dead end. It didn't scale up. In practice the main ideas were absorbed into the now orthodox two or three layer architecture. IIRC, Arkin was early to show the hybrid reactive+planner approach that we mostly use now.

Symmetry 11y ago

An introductory textbook probably isn't the place to be introducing heterodox views. None of the commercial robotics companies I'm familiar with, including Rethink Robotics where Rodney Brooks works now, use subsumption architectures.

gugagore 11y ago
I agree with you. But Roomba.
- Symmetry 11y ago
  Right, there is the Roomba. But don't the newer Roombas do mapping?
  - gugagore 11y ago
    I think that is true, but as you say that's the most recent model.
- Animats 11y ago
  The Roomba algorithm is roughly: - 1. Go in an increasing spiral until you hit something. - 2. Wall follow for a while. - 3. Turn away from the wall and go a few meters. - 4. Go to step 1.

Animats 11y ago

Brooks at MIT, and Latoumbe at Stanford, headed the two opposed camps during the AI Winter. Brooks wanted purely reactive robots, and did some nice insect robots. Latoumbe was into very rigid planning systems, where you approach manipulation as path planning using maze solving algorithms in a N-dimensional space with obstacles. Both turned out to be dead ends. Purely reactive systems don't get beyond insect level, and high-dimensional planning requires total information about the environment. Then the statistical machine learning people took over and started to get real results, especially in sensor data reduction.

eikenberry 11y ago
Neither was a dead end. Each made real contributions to the field and have real applications but, not surprisingly, turned out not to be the silver bullet that could do all things. The statistical crowd was always there and have been good at the same set of things, it just finally matured as a field and was able to produce usable results. They all go through the same motions as anything in this industry. They show promise, everyone gets all worked up about the new thing, it produces some results and gets integrated where it makes sense, gets boring and people start looking for the next thing.

robotresearcher 11y ago

Subsumption was also fast dead end. It didn't scale up. In practice the main ideas were absorbed into the now orthodox two or three layer architecture. IIRC, Arkin was early to show the hybrid reactive+planner approach that we mostly use now.