Conversation

Replying to
The slight hesitation between steps in an otherwise revealingly human-looking demo strengthens my suspicion this is a fairly conservative maneuver automaton based strategy. It was newish when I was finishing grad school, but old now.
Quote Tweet
Replying to @0knaomi
These don’t look weird enough to be like deep learning. Looks too human. I suspect they did motion capture on a human gymnast and are using a mix of nonlinear and linear model reference control patched together with a maneuver automaton. 90s vintage published research.
2
36
Deep learning would produce much weirder and more unstable movements. And they would not look so human (just as AlphaGo did not play like a human despite the same “rules” — here the rules are natural dynamics of the body). This much anthropomorphism has to be motion capture.
2
52
When they do get around to using deep learning you can expect weird zombie, Orc, or exorcist style movements. And a ton of falling since stability wouldn’t be explicitly engineered.
3
31
The basic idea is not complex. You do it when driving. Straight arc segments and steady turns are what are known as trim trajectories. You patch them with discontinuous but not unstable maneuvers (turning steering wheel to new positions while braking/accelerating to a profile).
1
14
In a car, the 3 inputs get translated to 2 steering angles, and differential turn rates with fixed formulas, so 3 inputs control like a dozen variables. Here it is more complex. At least 12 joint angles controlled at once. But still the same principle.
2
15
The stable trim trajectory segments would be parametrized gaits. Like jog, run, skip, etc. Then the patching. Probably the 12 angles couple and correlate down to 3-4 most of the time.
1
11
The hardest bits would be unstable transitions (one foot landing on an inclined plane is basically a hard inverted pendulum problem) but if it’s in a human-validated mocap sequence, you’d only need to solve it for a fraction of a second to transition to next maneuver.
1
16
If this is what they’ve done, it’s extremely impressive. Though known in the literature, these are hardly routine techniques like PID and LQG (which mostly won’t work here due to nonlinearity). Much more sophisticated. Still, not defensible.
2
14
Replying to
The quadrotor drones are pretty sophisticated, as are kiva style robots, and lots of industrial control things. I’d guess the fielded control systems in the world are 70% PD or PID, 25% LQG, 5% everything else. But the 5% would be very high-value applications.