Fit parameters from a log¶
Draft
This guide is scaffolded. The outline below marks what it should cover.
Fit minimizes windowed prediction error against logged
controls + measurements to recover physical parameters.
To cover¶
- Marking a
Parameterpromotable with amanifold=(thruster gains,Mass.mass, mount transforms). - Assembling
Windows from a log (x0,u,z,dt). - Adding a
Priorfor MAP regularization. - Running
Fit(world, parameters={...}), readingFitResult.convergedand the recovered values. - Fitting the noise model instead with
NoiseFit(innovation-NLL σ) — and why σ can't be L2-fit. - Pitfalls: whiten sensors before fitting;
OP_OUTPUTmust snapshot.
Source material¶
- Reference: System identification
- Code:
manta/fit/ - Tutorial: System identification — drone