Custom Data

Upload a recording with columns t, x1, x2, ... and u1, u2, .... A linear model x[k+1] = A x[k] + B u[k] is fitted to it by least squares, and once it is found your data becomes a live simulation you can push and steer with an LQR. Nothing leaves your browser.

1 · Arrive Turn a recording into a model, then drive the model as a live simulation with a controller you designed. about 10 min

2 · Inform What the page does

Least squares picks the A and B that best predict each sample from the one before. The dashed lines show those one-step predictions against your data; the fit percentage says how close. The poles show the model's own motions: any outside the unit circle is unstable.

The LQR trades state error (Q) against input effort (R) and gives a feedback u = −K(x − xref) plus the input that holds the reference. Once the model is found, the bottom plot becomes a simulation of it, running live: drag the reference, push the states, or switch the controller off to see the model on its own.

Some states may not be reachable. The input only moves a state directly or through the other states. If nothing connects a state to the input, no controller can steer it, and a recording that never excited it cannot show how it responds. The page counts the directions the input can move (reachable: r of n) and tells you when some are out of reach. The same goes for tracking: a state can be followed only if some steady input holds it at a constant value.

The recording is a plain-text CSV, described below.

CSV format

A header row, then one row per sample. Three kinds of column, named like this:

t,x1,x2,x3,x4,u1
0.00,0.00000,0.00000,0.08000,0.00000,-1.9060
0.05,0.00031,0.00112,0.08140,0.01560,-2.0420
0.10,0.00112,...
ColumnRule
tTime in seconds. One column, strictly increasing.
x1, x2, ...The states, measured at each sample. A single state can be just x.
u1, u2, ...The inputs, as commanded at each sample. A single input can be just u. The input in row k acts between row k and row k+1.

Columns with any other name are ignored. Commas, semicolons or tabs separate the columns, and lines starting with # are skipped. Values are plain numbers in your own units; the LQR gain comes out in those units. A row with an empty cell is dropped. You need at least 20 complete rows. Evenly spaced times are best: if the spacing varies by more than 5%, the data is resampled onto the median spacing. The signals should wiggle around one operating point, since the model is linear.

Recording

LQR weights

Simulation

Result


    
3 · Process

Try it

  1. The page opens on an example: a cart-pole balanced by an LQR with a little dither, recorded through noisy sensors. How well does the model predict each state, noise and all?
  2. Look at the poles: is the model unstable on its own? Where do the poles go with the LQR?
  3. Watch the bottom plot: it is your model running live under the LQR. Drag the reference and the tracked state follows it.
  4. Press Push the states: how fast does the controller recover? Raise a Q weight and it recovers faster, at the cost of more input.
  5. Switch the controller off. Does the model sit still or run away? Compare with the poles.
  6. Change the tracked state. Some states may have no steady state of their own, or may not be reachable at all: read the reachable count above the weights.
  7. Download the example CSV, rename a state column (say x3 to y3) so it is ignored, and see what the model loses. Then try your own recording; Load the example brings the cart-pole back.