EIGENROOMLAB WORKSHEET 14ANSWER KEY · FOR INSTRUCTORSCart-pole · State space
Control from estimated state
Goal
Compare control from an estimated state with full-state LQR under the same weights and nudge.
Setup
Lab eigenroom.com/lab?lesson=lqg · short link eigenroom.com/lqg
Open it on a phone or a computer; there is nothing to log in to or install. The link loads run A's settings:
- Initial angle
- 6°
- Controller type
- LQG
- Position cost
- 10
- Angle cost
- 100
- Force cost
- 1
- Process variance Q
- 0.00001
- Sensor variance R
- 0.001
- Sensor noise
- 0.3
- Disturbance events
- Nudge cart 1.2 N·s at 1.4 s
- Recording
- 8 s
- APreset, exactly as the link opens it.
- BChange only: Controller type → LQR.
- CChange only: Sensor noise → 0. (ext.)
1. Predict
Does estimated-state feedback recover like full-state LQR after the same kick?
LQG feeds the Kalman estimate into the LQR gain. It needs both a plant model and uncertainty assumptions; saturation and nonlinear motion break textbook optimality guarantees.
2. Observe
| Quantity | Where to read it | A | B | C ext. |
|---|---|---|---|---|
| Peak angle (°) | Measured response | 17.86° | 18.43° | 17.42° |
| Peak cart travel (m) | Measured response | 1.97 m | 0.77 m | 1.96 m |
| Final cart position (m) | Measured response | 0.01 m | 0.00 m | 0.00 m |
| Control effort (N·s) | Measured response | 3.93 N·s | 2.62 N·s | 3.74 N·s |
Values computed by DC motor and cart-pole engine 1.5.0 at the settings above. A replay readout can differ in the last digit if the slider is not exactly at the stated time; values read from a plot are approximate.
3. Explain
With the same weights and the same nudge, which run travels further and uses more control effort? Why?
LQG (run A): cart travel 1.97 m and effort 3.93 N·s; LQR (run B): 0.77 m and 2.62 N·s. The estimate lags the true state and carries sensor noise, so LQG acts on a delayed, noisy picture of the state.
Both runs keep the pole up. Which property of full-state LQR does LQG not inherit, and what in this model breaks the optimality claim for both?
Full-state LQR comes with guaranteed gain and phase margins; LQG keeps the stability of the linear design (the separation principle) but not those margins. In both runs the force limit and the nonlinear pendulum mean neither controller is optimal in the textbook sense.
Extension (optional)
Set the sensor noise to 0 and run LQG again (run C). How close does it come to LQR?
Cart travel becomes 1.96 m and effort 3.74 N·s, against 0.77 m and 2.62 N·s for LQR. The remaining difference comes from the estimate's lag after the nudge, not from noise.