Slashdot Mirror

http://www.plyojump.com/

Favorites:
BigDog http://www.bdi.com/content/sec.php?section=BigDog
RHex http://www.rhex.org/
ZMP, Inc. http://www.zmp.co.jp/e_home.html
RiSe http://www.youtube.com/watch?v=JzfP0Ig7eVQ

NaturalMotion isn't quite as dynamic as it may appear. It's not doing a full physical simulation most of the time. There's still considerable use of motion capture data and kinematics. Blended approaches like that are the norm in animation; full control of a humanoid character with real physics is still not that great. There's continual progress, though. Today, we have the MIPS and many of the algorithms. When I was first working on this, we had 20 MIPS, and things were rather slow. We could get the theory to work, but the hardware wasn't there yet.

Much of this was done a decade ago, by Motion Factory, which was a Stanford spinoff. That was purely kinematic, but produced reasonably good movement. Eventually, Softimage/Avid bought that company, but didn't do much with it. Boston Dynamics has also done work in that area.

Kuffner (above) is right, of course. Dynamically stabilized walking has been around for years. It's not easy to do, but it's been done. Raibert first did it in the 1980s. See his book, "Legged Robots that Balance".

Most of the self-balancing walkers, as Kuffner points out, use a ZMP-based approach. This works for walking, although it's not quite enough for effective running.

Many of the dynamically balanced robots can rebalance after a shove. BDI's Big Dog can. So can some Japanese hobbyist robots.

If you're not up to date on how far along Japanese hobbyist robotics has progressed, see these videos of this month's humanoid robot soccer match. These robots are mostly manually controlled, but have computers managing some functions. Many have rate gyros to assist with balance. Gradually, the computers and sensors are taking over more of the control. The hobby robotics manufacturers in Japan now have about 70% of the functionality of Asimo at 2% of the price. There are hobbyist robots with WiFi links and cameras on board. A few more improvements and you'll be able to do all the Asimo stuff with a $1500 robot. But it will only be about 60cm high.

What you're looking at is a one-legged hopper. Locomotion researchers find it useful to work on one-legged hoppers because the system is simple enough to be analyzed analytically.

The new feature of this leg seems to be that it has three active joints with sufficient power behind them for jumping. Most legged robots, such as the BDI Big Dog, only have two active joints in the leg, although some have a weak or passive ankle. This is enough to position the leg to any point in the working envelope, and it's not obvious what a third joint buys you. ASIMO has an "ankle", but it's used only to align the foot with the ground, not for active running; ASIMO runs flat-footed, not on the ball of the foot. This Toyota machine seems to have both an ankle and a toe joint.

This is a big win, as I described back in 1995 in my "Why Legs have Three Joints" paper. With three joints involved in running and jumping, you gain control over the force vector for ground contact, which allows slip control. Also, the hip joint (which is usually the most powerful) can be used more effectively; the lower joints position the leg so that the hip muscles can do most of the work. For humans, this is subtle, because the ankle-toe distance is small. It's much clearer for horses, where the hind leg has three sections of roughly equal length.

I hate to send a slashdotting to such a nice group of guys, but the Big Dog made by Boston Dynamics might be what you are looking for.
http://www.bdi.com/content/sec.php?section=BigDog

While it doesn't have 2 legged operation it does operate in the "falling forward" way of walking. Check out that video where the guy gives it a swift kick to the side and it just sways and regains it's balance. Looks uncannily lifelike.

I mean, none of those seem nearly as cool as the BigDog quadriped, or the ball balancing robot, or my personal favorite, R.O.B.

Robots are awesome though, my roomba just cruised around my desk.

Most of the little humanoid toy-sized robots are a joke in the sensor department, but this one has the gyros, accelerometers, and force sensing to, maybe, get it right. Importantly, it has three axes of force sensing on the legs. The Aibo and BDI's Little Dog do not, which limits them to semi-static gaits. If you have that force sensing, you can do slip control and potentially run up hills.

I have a long-standing interest (and some results and patents) in the legged running area, and I'm glad to see the hardware catching up. Simulation is nice, but limiting.

And if you can't outrun them you can just kick them and they'll fall over. Oh wait...

Raibert did some great work in the Leg Lab's early days. Raibert's big insight was that balance is more important than gait, and he did work with one-legged machines with springy actuators to force the issue. In his day, the Leg Lab had one, two, and four-legged running machines. But he left MIT to do a startup, which seems to have ended his dynamics work. BDI does mostly kinematic models.

The next professor to head the Leg Lab was Gill Pratt, who was more of an actuator guy. He didn't accomplish too much, and is now at some lesser school. Under Pratt, the Leg Lab backed down from running machines to walking machines.

There was somebody after Pratt, but apparently the Leg Lab is now defunct. It's sad. They made so much progress under Raibert.

It's possible to go beyond walking and running on the flat. Legs are really for traction control. All the MIT work assumes that the "feet" don't slip. That doesn't work on real hills or slippery surfaces.

There's two phases to dealing with slip. First, you need to limit joint torques to below where the feet start to slip. Once you do this, you can climb some hills. (Video, 8MB .mov file). That work is ten years old, and still, nobody else seems to be handling leg slip at all.

The next step is to use the three joints of a leg to adjust the vector at which the normal force is applied to keep the ground contact inside the friction cone. Then you can climb more serious hills. Once you get this figured out, much of how humans move when dealing with terrain becomes clear. Leaning forward and bending the knees more when going uphill is all about slip control. Think about it.

Working on this diverted me off into physics engines, because everything that was available ten years ago sucked. So I did a physics engine that worked, which turned into a business. There are still very few physics engines good enough for legged locomotion work. Most physics engines, especially the Baraff-type impulse/constraint ones, don't do friction well. Since legged locomotion is all about managing foot-ground friction, you need a simulator that gets friction right. (Hint: if a simulator can't do a driving game without special-casing the wheel/ground contact, it won't work for legged work.)

All this is patented, of course.

Brooks did some great insect-level AI. It's purely reactive, with no world model. That was good work, and a reasonable reaction to the logic-based planning crowd then running MIT AI. But then Brooks started going around saying that the reactive model was far more powerful, and started making human-level AI noises. From this came Cog.

When Brooks came by Stanford to talk about his plans for Cog, I asked him "Why don't you go for mouse-level AI", something that didn't seem totally out of reach. He said "Because I don't want to go down in history as the man who created the world's greatest robot mouse". As one of the grad students on the Cog project said, "It just sits there. That's what it does. That's all it does". And, years later, it still doesn't do very much.

The model-free approach is just too dumb. With no world model, you can't get beyond insect-level AI. That approach works mostly for creatures in environments where inertia doesn't matter much. For insects, banging a feeler into something is fine. For large animals, you get bruises or worse. As creatures get bigger, faster, and stronger, they need models with some predictive power, so they can avoid mistakes before damaging themselves. I tell people in academia that you need to be "less formal than Latoumbe (who formerly headed Stanford's robotics operation) and more formal than Brooks". The game development community has absorbed this lesson, but it's only starting to get through to the robotics community.

Raibert's work on legged locomotion was very impressive. I'm very familiar with that work; I've done some improvements on it. Raibert had one great insight - balance is more important than gait. People have been studying locomotion for a century, and almost all the studies center on gait. Raibert realized that balance was more important, and built a one-legged hopper to force the issue.

But, in fact, the way Raibert does locomotion is very simple. There are two controllers, both simple hand-tuned PID loops, and a state machine that swiches between them. This can handle simple locomotion on the flat, and some preprogrammed moves like flips, but it doesn't generalize. I'd expected the adaptive control people to pick up from where Raibert left off, but so far, nobody has really done that.

My insight there was that slip control is more important than balance. On the flat, traction control isn't a big deal, but on hills or rough terrain, traction control dominates balance control. That's what legs are really for. If you add automatic traction control to Raibert's approach, legged running on hills becomes possible. Otherwise, you slip out climbing hills.

Raibert himself left MIT and did a startup company, Boston Dynamics. But they ended up selling products to DoD which are game-like kinematic simulations. They don't seem to stress dynamics work any more.

The MIT Leg Lab was taken over by Gill Pratt, who was more of an actuator and controls guy. He didn't accomplish much. The next head of the Leg Lab was some guy who was into prosthetics. The Leg Lab now seems to be defunct. Their web site hasn't been updated since 1999.

Have a look at their Engineering page. You'll see images of both Rugged RHex (featured on Slashdot earlier) as well as Sony's bipedal Qrio robot. Marc Raibert has assembled an impressive team of people to work on these very cool projects.

Motion capture is not the final solution. Breaking down realistic human motion into mathematical equations is. Just ask the folks at Boston Dynamics. I saw a demonstration of one of their models (programmed by a graduate from my lab) and it was superb.

Gil Pratt took over the Leg Lab, and focused more on actuator design. Raibert's machines worked, but needed hydraulic, electrical, and pneumatic umbilicals. Better machine design has produced more compact robots.

The idea of springy joints has been around for a while. It's common to model muscles as springs and dampers for which the spring constant, neutral point, and damping factor are adjustable. It's well-known that in mammal running, most of the energy of each stride is stored as spring energy in muscles. (As I recall, about 80% of the energy is recycled for the next stride, so this is a big win.) There's been work at Stanford on flexible manipulators, although that's more related to arms. McGill has a small, high-efficiency hopping machine.

Unless you use pneumatic actuators, off the shelf components aren't well-matched for this approach motion control. That doesn't mean it can't be done, but you spend a lot of time on component development. That's what the Leg Lab has been focusing on under Pratt, and that's why the little dinosaur model was tough to build.

Rod Brooks from MIT also tried a robotic startup, IS Robotics, which produced a $100K robotic insect. Didn't sell. It's really hard to sell mobile robots; I've known several people with failed startups.

I work on this sort of thing for games and animation.