[regularfry]

Some of it was pre-computed. The middle layer, if you like. The Boston Dynamics group had walking gait of a sort nailed in the 80's; the trotting-on-the-spot that BigDog did was essentially a continuation of those mechanics and that's all based on a conceptually simple balance problem which is intrinsically reactive and not pre-planned. So that's what was going on at the lowest level.

At the top level you have the actual environment, with those meme videos of the robot trotting through a car park, getting kicked off balance, and recovering. The whole point of those tests was to demonstrate how robust their tech was to non-precomputed disturbances.

And between the two you've got the direction and planning layer, telling the robot to go from A to B with some set of suitably convoluted parameters that nobody but the operators would have understood. That planning layer might do all sorts of pre-computation and simulation but it needs to do it in the context of a noisy and possibly adversarial environment. That's equally true for Atlas as much as it was for BigDog, even when there's nobody actually kicking it. What I suspect the precompute and simulation is doing at that layer is a) checking for physical viability of the requested route, and b) parameter tuning in response to sensor readings over a number of runs. Not telling the robot the exact sequence of motions. But I'm nowhere near those teams (oh, I wish) to comment on whether that's true - maybe someone else round here is.

[Animats]

Right. Early thinking on this was zero moment point (ZMP) control. The zero moment point is where, if you land from a stride there, you continue going in the same direction at the same speed. You can speed up by displacing the landing point backwards a bit, and slow down by displacing it further, and turn by displacing it left or right.

When a foot or feet are in contact with the ground, the goal is to stabilize the posture. When in flight, the goal is to hit the zero moment point. Reactive controllers do that. The planner picks the next landing point.

ZMP robots tend to start moving by running in place, then slowly transition to fast forward motion. You see a lot of robot videos like that.

The next step up is to control takeoff. Bend knees, fall forward, launch. Launch angle is a planning problem. Speed and direction changes become much faster. Basic gymnastics are almost within reach. The planner controls launch angle and landing point, and the lower level reactive controls do the millisecond-level control.

It's possible to do all that from first principles of classical dynamics. That level of performance was reached before machine learning. That's what you're looking at in older BD videos.