Slashdot

This is very nice work. It's good to see Raibert doing robotic locomotion again, and finally, with a big enough budget.

Back in the 1980s and early 1990s, Raibert headed the MIT Leg Lab, which produced the first legged robots with real balance control. Raibert started with one-legged hopping machines, to force the balance issue. His big insight was that balance is more important than gait. In 1992, he left MIT and did a startup, Boston Dynamics, and went off into simulation. Most of the simulations weren't dynamic, just kinematic. Now he's back to robotics, and dynamics, again.

I've worked on control of robot running on rough terrain [youtube.com]. So I understand the problems. Watching the Big Dog video, I have a reasonably good idea of how it works. This is quite impressive. DARPA got its $40 million worth.

First, it has slip control, like automotive ABS, for its feet. The first insight on the hard cases for locomotion is that balance is more important than gait. The second is that slip control is more important than balance. The key to slip control is keeping the transverse forces at foot-ground contact below the point where the feet break loose. ("Inside the static friction cone", for those familiar with the terminology.) Watch it move on ice. The feet do not slip at all unless there's real trouble, as when someone kicks the thing. The transverse forces are being held below the break-loose point. Given the load on the foot, the actuator forces (hydraulic cylinders on Big Dog) must be coordinated to keep the transverse force below the ground coefficient of friction times the longitudinal load. Finding the ground coefficient of friction can be either trial and error (if it slips, reduce the value) or they may have actual slip sensing in the foot, like humans and animals. Humans, incidentally, tend to maintain a contact force about 20% above the break-loose point, as a safety margin.

Big Dog's reaction to a slip is to immediately raise the foot and go for a new foot placement. That's an emergency behavior, though; it's the prevention of slip that makes it work. Watch the robot's reaction when it slips on ice, and, once you know what to look for, you'll see how it does it. The first priority is to recover traction. As soon as a foot slips, it's lifted and placed in a new position. The second priority is to recover balance. As the robot starts to roll to the right, it executes a violent twist to the right and throws out the right front foot. It needs a foot position within the traction limits to provide the roll moment needed to recover balance, and it has a good enough planner to find one. Look at that sequence and ask yourself first "where does the foot need to be to get traction", then "where does the foot need to be to recover balance". Then you'll understand how it works.

Big Dog has, finally, true gaitless locomotion. Decades of locomotion research have focused on gait, foot sequence, "central patten generators", and similar mechanisms that deal with the easy cases. Wrong answer. The right answer is to think of legs as assets that can be deployed to maintain slip and stability criteria. It's very clear that Big Dog does this; it can use its feet (and knees!) as necessary. It's not constrained to a gait pattern at all.

There's a true dynamics predictor and planner in there. This is not just a reactive robot, like Brooks' little machines. Nor is it a straightforward ZMP ("zero moment point" [wikipedia.org]) stabilization system, like Asimo. (Think of ZMP as a generalization of center of gravity to include momentum.) There's a planner with a horizon of (I think) about two foot placements ahead, and it has "what if" internal simulation capability. That's why this robot moves so well. It can predict, at least approximately, what's going to happen for its next move, and plans on that basis. That's why its movement are so smooth. Without that, you'