Base-frame calibration: where the robot sits relative to the cell

[Onwcan]

What is missing

WP-06 named three calibration problems.

Two are implemented:

• calibrateToolPoint
• calibrateHandEye

The third, calibration of the robot base relative to the cell frame, is not implemented.

ADR-0011 records this explicitly.

Why it matters

Without base-frame calibration there is no direct way to express a target in cell coordinates.

Every target pose given to the arm must already be expressed in the robot's own base frame.

That means a fixture drawing or an externally measured location must currently be transformed manually or by an external tool.

Base-frame calibration is what turns "the part is at this location on the table" into a pose expressed in the robot's base frame.

It is missing while the two more difficult calibration primitives are already present.

Proposed approach

The problem uses the same algebra as hand-eye calibration.

Use A X = X B with:

• the robot's measured flange poses on one side
• external measurements of the same poses on the other

The quaternion-linear form from ADR-0011 Decision 1 applies unchanged:

(L(q_A) - R(q_B)) q_X = 0

The existing null-space-by-shifted-power-iteration approach can also be reused.

Where

• include/motionkit/core/calibration.hpp
• src/core/calibration.cpp — use calibrateHandEye as the implementation template

Acceptance criteria

• Add calibrateBaseFrame.
• Accept paired stations containing the flange pose measured in robot base coordinates and the same flange pose measured by the external device.
• Detect degenerate geometry as its own failure mode.
• Parallel rotation axes across all motions must be recognised as making base position along that axis unrecoverable.
• Report axis_spread so "barely solvable" can be distinguished from "not solvable".
• Keep DegenerateGeometry distinct from NotEnoughSamples.
• Report both RMS residual and worst-case residual.
• Use fixed-size normal equations.
• Keep the implementation allocation-free and assert this in a test.
• Add a noise-free synthetic recovery test that approaches machine precision.
• Add a deterministic-wobble test showing the reported residual is of the same order as the injected measurement error.
• If the implementation would substantially duplicate calibrateHandEye, factor the common solve instead.
• Document in the ADR whether the solve was shared or kept separate and why.
• Amend the ADR-0011 consequence section.

References

docs/adr/0011-calibration-refuses-what-it-cannot-determine.md, Consequences.