calibration-algorithm-iteration

name	calibration-algorithm-iteration
description	Iteration playbook for improving calibration and registration pipelines without losing baseline comparability. Use this when asked to optimize a calibration algorithm, propose next iterations, run ablations, or improve quality step by step rather than making a risky one-shot rewrite.

Calibration Algorithm Iteration

Use this skill to improve calibration algorithms in controlled, evidence-backed steps.

When to use

Use this skill when the task involves:

• improving an existing calibration baseline
• planning iterative upgrades instead of a rewrite
• choosing among candidate enhancements
• refining one concrete hypothesis at a time

Use calibration-benchmark-and-ablation for multi-method comparisons and calibration-sota-upgrade for paper-driven method scouting.

Core policy

Iteration must be:

1. Baseline-preserving
   • keep the trusted baseline runnable and comparable
2. Single-hypothesis
   • change one major idea at a time
3. Measurement-locked
   • keep metrics and artifacts stable while iterating
4. Failure-aware
   • record when the change helps, hurts, or only shifts failure modes

The iteration loop

For each round, always produce:

1. Hypothesis
   • what is expected to improve and why
2. Minimal change
   • the smallest implementation that exercises the idea
3. Comparison
   • baseline vs candidate on the same inputs and outputs
4. Failure review
   • what new breakage or brittleness appeared
5. Decision
   • keep, revert, or gate behind an option

Preferred upgrade order

Prefer this ladder unless strong evidence says otherwise:

1. Better data screening and sample weighting
2. Better initialization or correspondence pruning
3. Better robust loss and coarse-to-fine scheduling
4. Better local refinement or bundle adjustment style solve
5. Better global optimization or joint optimization
6. More ambitious modeling assumptions such as continuous-time or learned priors

This order usually yields more reliable gains than jumping straight to a sophisticated optimizer.

Common upgrade levers

Good iteration ideas include:

• stricter acceptance gates for weak samples
• observability-aware sample selection
• robust loss tuning
• multi-stage initialization
• coarse-to-fine registration
• holdout-based early rejection
• better uncertainty or conditioning diagnostics
• explicit fallback from strong candidate to trusted baseline

LiDAR-to-LiDAR iteration order

For lidar2lidar, prefer this order:

1. better window selection and scene sufficiency gating
2. better pairing inside static windows
3. better representative-transform or consensus selection across windows
4. only then better local refinement or loop-closure optimization
5. only after that, stronger initialization methods

If repeated reruns jump between distinct solution families, stop calling it convergence. Treat it as an observability or pairing problem and change the earlier stages first.

Required output

When proposing or reviewing an iteration, include:

1. Current baseline
2. Proposed change
3. Expected upside
4. Expected new failure modes
5. Comparison plan
6. Rollback or fallback rule

Reference canon

Use concrete references such as:

• Ceres Solver docs for iterative nonlinear optimization hygiene
• Kalibr for staged calibration rather than one-shot magic
• Open3D / PCL for coarse-to-fine registration pipelines
• TEASER++ for strong initialization under heavy outliers
• GICP and NDT literature for alternative refinement behavior
• continuous-time calibration papers when motion distortion or asynchronous sensing is the actual bottleneck

Anti-patterns

Avoid these iteration mistakes:

• replacing the baseline before the candidate wins repeatedly
• changing the optimizer and the metric at the same time
• merging several experimental ideas into one patch
• accepting a candidate because one demo looks better
• removing diagnostics to make the new method look cleaner

Output style

Present iterations as a ladder:

• Round N baseline
• Round N+1 candidate
• Expected gain
• Decision metric
• Abort condition

Make the next step obvious and falsifiable.