FORC · Lecture 1

Feedback, state, and control objectives

Simeon Nedelchev

Feedback loop

The plant evolves under the applied input.

The controller uses observations to choose the next input.

An estimator can supply quantities that sensors do not measure directly.

x=(q,v,a),\qquad y=g(x)

MuJoCo: qpos, qvel, act.

The first labs assume full-state observation.

f(x^\star,u^\star)=0

A hover equilibrium requires nonzero thrust.

A tracking reference must respect the dynamics.

m\ddot p=-k_p(p-p^\star)-k_d\dot p+d

At rest, the position offset is $d/k_p$ .

For $d=1$ N and $k_p=20$ N/m: 5 cm.

RMS and peak error, force limits, failures.

Same initial states and disturbance schedule for every controller.

A policy has low mean error and frequent saturation.

What would you test next?