Self-Introspecting Robot Learns to Walk

StCredZero writes "There's something about these things that seems eerily alive! The Starfish Robot reminds me of the Grid Bugs from Tron. But it's very real, and apparently capable of self introspection. In fact, instead of being explicitly coded, it teaches itself how to walk, and it can even learn how to compensate for damage."

Comments

I'm not one to complain about newsworthiness

[doombringerltx]

but come on! "Update 24-Nov-2006:"

Self-reflection, literally!

[RyanFenton]

This is a very well-done video. I really like how it shows the virtual model to illustrate how the system 'sees' itself. Self-reflection of a sort is usually present in most complex programmed systems in one form or another - usually in terms of disjointed status variables and variables for their hard-coded implications. This is neat because the implications can be a little more dynamic.

I hope this becomes a more general library that can be used to help self-reflection of this sort become a more separate part of physical designs. Even if the implications of the physical model aren't dynamic, a standard way of quickly seeing how your model 'sees' itself would help debugging and development in many future projects.

The only problem if it becomes more prevalent would be same one that quantum mechanics holds - people think that 'observer effects' has to involve consciousness, in the same way they'd think that a program's self-reflection would mean that it 'thinks' the same way they do. Neither is true - they're all mechanical terms wrapped in common language. Anything that can record an effect on the world (a falling rock's scratches in another stone would work) is a quantum observer - consciousness has nothing to do with the 'collapsing wave function'. The same here - a bit of self-reflection on the part of a program doesn't mean it's eerie self-corrections are capable of the complexities of our mind. If anything, such mechanical results would imply that our own minds act simpler in some ways than we may think, and that consciousness doesn't necessarily have to be as inscrutable and special as we might want.

Ryan Fenton

Re:Self-reflection, literally!

[Original Replica]

"You could conceive of natural selection never producing consciousness "

Oddly enough, you could not conceive of anything without consciousness. Understanding is a mental, not physical process. You could however conceive of consciousness without the physical world. Indeed every culture has been doing so for all of recorded history in the form of spirit worlds, afterlife, etc.

Occam's razor can be much abused depending on how you frame your observation. "I think therefore I am." is much more straight forward than "I am incredibly complex and elaborate, therefore I think." Let's set Occam's Razor aside for this discussion, it doesn't seem to be the right tool for the job here.

If you allow yourself to view the conscious world as more fundamental than the physical world, then the observed consistency/connectedness of all physical phenomena would require some sort of governing over-consciousness that is responsible for the physical world. That of course would be a form of creationism, much reviled here on /.

Creepy

[Spy der Mann]

That thing almost looks alive. After seeing it, it reminded me of the nurses in Brookhaven Hospital trying to move. Eew.

Re:Creepy

[Warbothong]

At the start it looks creepy when it's moving around looking a little like a spider. Then it gets damaged and looks genuinely scary, in terms of "WHY WON'T IT DIE?!". At the end it just looks like its makers enjoy pulling the wings off flies (although I did laugh when it flipped itself upside down). It's be interesting to see whether this modelling system could be made to learn from its experiments and failures as well as creating initial similations to work from. What I mean is, its internal simulation lets it determine effective ways to move around like brains can do, but brains are also able to factor unknown environmental effects in as well. For example someone might slip on ice and fall over, since they hadn't realised how slippery the ground would be in their internal modelling. A person would get up and start walking differently (making sure each foot was firmly planted before putting weight on it, etc.) since they would factor this into their internal simulation. It would also be interesting to see whether this could get a type of "pain" added. This isn't to be sadistic or anything, as the machine wouldn't be made to "feel pain", I just mean that another factor could be added into the simulation, like "this leg is barely attached anymore, better not put too much stress on it" or "this area is important and delicate [eg. batteries, sensors, etc.] and thus shouldn't have too much weight put on it". I say this since in the video the robot has one leg damaged and then tries to walk by slamming its full weight down on the "stump", which does not seem a particularly amazing survival ability (for instance, natural selection seems to favour limping, which puts less stresses on damaged body parts and more on healthy ones which should be able to cope).

Re:Damage

[hasbeard]

Well, it seems to me that combat really isn't a good time for too much introspection. I mean with all the bullets flying and all.

Two questions...

[LynnwoodRooster]

1. Did they give it a navel?

2. Can it contemplate it?

Re:Poor thing...

[KDR_11k]

Supposedly a mine-clearing bot (lots of legs designed to be blown off by mines, the bot just walks around and triggers them) that was literally on its last leg was pulled out of the testing (it would have crawled onto a final mine and be destroyed in the process) because the supervising officer felt sorry for it. People are capable of feeling empathy for the dumbest animals, why wouldn't they for a robot?

Skipping the blogodreck, here's the real info

[Animats]

First, get past the blogodreck to the actual work. (Slashdot editors missed a blog troll again.) Also, this work is several years old. The papers are from 2004 to 2006.

The original article says that the robot has "tilt and angle sensors in all its joints", but that's wrong. It only has one central tilt sensor. That's significant, because if it did have tilt sensors at each joint, system identification would be easier. The algorithm is doing better than one might expect.

This thing is doing what controls people call "automatic system identification". You have some set of sensor inputs and some set of control outputs, and the control system has to figure out how they relate. It does this by adjusting the outputs and watching what happens. There are various statistical techniques for doing this. Calling this "introspection" isn't really correct.

After system identification, the model is inverted, or solved for the inputs in terms of the outputs. The inverted model can then be used as a controller. Given desired outputs, the inputs needed to achieve them can be computed.

The novel result here is that a reasonably decent system identification for a nonlinear system is being performed with a small number of physical tries. That's an improvement over previous methods, which tended to "learn" very slowly. I'd looked at approaches like this for legged locomotion in the past, but the available system identification algorithms weren't good enough. This looks promising.

Good robotics work, crap Slashdot article.

So would a beowulf cluster of these....

[SnoopJeDi]

...be called a herd?

The next step...

[Jafafa Hots]

...is to make it capable of autonomous self- introspection.