Dexter Walks (first dynamically balancing biped robot)

http://www.paulgraham.com/anybots.html

by perler • 19 years ago

55 20 19 years ago

paulgraham.com

chris• 19 years ago

I had the opportunity to see Dexter live in action, it was one of the coolest things I've ever seen!

brezina• 19 years ago

Congrats Tevor - I suspect the military has spent 1000X the money with 1/ 1000X the progress.

jwecker• 19 years ago

Brilliant. Well done. It's very cool seeing him go up onto his tip-toes to keep balance. I also think it's cool that they built a whole other robot just to try to pick a fight with Dexter.

BrentCastle2• 19 years ago

Would anyone care to comment on what the downfalls of the commercial gyros were? Is it precision or time-lag? I know I can't get specifics, but I'm curious in general what is different in this application that commercial gyros were not capable. Any info would scratch my itch! Thanks!

BrentCastle2• 19 years ago

Sorry. Accidentally refreshed and reposted. A moderator can delete this comment!

mynameishere• 19 years ago

Reminds me of an old person who had a stroke and has to shamblingly re-learn how to walk.

Alex39172• 19 years ago

It's interesting to look at the differences between the way Dexter's doing it and the way we do it. Dexter appears to be lifting his foot straight up off the ground and then falling forward to produce the forward motion.

The way humans walk is exactly the opposite. The toes are at a different angle from the foot (when walking), so that our feet effectively look something like ___/ , with the underscores being the foot and the / being the toes. We then transition from the ___ part of the foot to the / part of the foot and push forward, and land flat footed with the other foot. For whatever speed you are going, there is a proper ratio between landing on the ___ part of the foot and pushing off the / part of the foot, and that's where balance comes from. That is, the faster the acceleration of the fall, the faster you transition from ___ to / .

So in humans, balance comes from the way one step connects to the next step, whereas Dexter appears to be treating each step as a discrete unit, where it has to attain balance at the end of one step before going on to the next step.

If you actually have to balance yourself after each individual step instead of using the next step for balance, the problem seems in some ways to be much harder.

Also, the balance problem is greatly compounded by the small steps Dexter takes. Compare this to doing lunges. When you do a lunge, you basically take a long stride forward and then sink into it, all while keeping your torso perfectly upright. This gives you a much bigger platform for stability. As terrain gets steeper, the way humans walk becomes more like a lunge for this reason.

So my guess is that taking one tiny step and then balancing is the hardest part of this problem, because you have the least tools at your disposal. For example, you can't use the next step to balance you. However, once you can do the last step of the sequence, all the preceding steps working backward from the last one get progressively easier.

I don't know anything about robotics, but as an athlete and sports physiology geek I find the problem interesting.

jadams3• 19 years ago

> [...] Dexter appears to be treating each step as a discrete unit, where it has to attain balance at the end of one step before going on to the next step

I doubt that. I don't think that would qualify as dynamic balance. Remember, it's only been a couple of weeks since Dexter's first step -- he's walking quite well.

This is one of those situations where the approach bounds the eventual utility. There are a lot of non-dynamically balanced robots that would just blow the minds of researchers from 20 years ago, but they are still limited.

By solving the dynamic balance problem, even though to the untrained eye it may currently look less impressive than some other systems, there's no more ceiling to what it can do, dynamically (control-wise, not accounting for power density). All of the behaviours you mentioned can be implemented on top of this substrate.

Alex39172• 19 years ago

That makes sense. Thanks for the insight.

jadams3• 19 years ago

Gah! Why do I keep double-posting? Someone please delete this.

ecuzzillo3• 19 years ago

It seems to me that things like this have been done in academia for a while; because I've only seen a video (rather than, say, some code or a paper), I'm unsure exactly what Dexter is doing, and whether it's different from the stuff done with bipeds at CMU, MIT, or in Japan. I wonder if there is any more information available.

pg• 19 years ago

Really? Do you have a reference?

ecuzzillo3• 19 years ago

I'm in a class on the subject at CMU, so I should be able to find out. What you said about Asimo is not quite true; for instance, one guy here at CMU wrote some code that took Asimo and constantly replanned in real time, so that he could respond to various changes in the environment dynamically.

Here, for instance: http://www.cs.cmu.edu/~Joel/footstep/videos.html#ASIMO

It would help to know more about exactly what Dexter is doing, so I can find some similar stuff. There is a large amount of literature on the subject.

Also, you note that Asimo's floor has to be hard and flat-- generally true, but also true for most bipeds, which seems to include Dexter. Can Dexter really deal with irregular terrain? How irregular can it be? An interesting question is how to get a biped to walk on *really* irregular terrain, like boulder-size.

jadams3• 19 years ago

Sony claims that Qrio can respond to outside pushing, as here:

http://www.plyojump.com/movies/qrio/sdr-4x_recoverfromfall.mpg

altay• 19 years ago

Incredible!! Congrats, Trevor!

Now, if you'll excuse me, I need to go prepare my bunker for the coming War with the Machines.

jamiequint3• 19 years ago

wow, that's the most insane negative feedback system I've ever seen

jamiequint3• 19 years ago

um, why the mod-down... http://en.wikipedia.org/wiki/Negative_feedback

ecuzzillo3• 19 years ago

The negative feedback wasn't the problem; it was more the ogmz lozl amazing.

Dexter Walks (first dynamically balancing biped robot)

Related Stories