Transfer between sensor layouts¶

OpenTac-Transfer (opentac.transfer) maps a calibrated normal-force distribution from one sensor layout to another using their physical geometry. The sensors can differ in outline, taxel pitch and taxel count. This supports reuse of recorded or simulated contact data across layouts, subject to assumptions about spatial resolution, calibration and contact placement.

from opentac import manifest
from opentac.transfer import transfer

a = manifest.load("skin_a.json")
b = manifest.load("skin_b.json")
frame_b = transfer(frame_a, a, b)      # frame_a in newtons per taxel

examples/transfer.py simulates a press on a 4 mm skin, transfers it to a 2.5 mm skin, and compares it with the same press simulated on that skin.

Three frames at one millimetre scale: a simulated ball press on a 4 mm grid, a bright 2 by 2 block; the same press transferred onto a 2.5 mm pitch board, a dimmer, finer spot; and the press simulated directly on that board, a similar spot. — Example from `examples/transfer.py`: a 10 mm ball at 3 N simulated on skin A (4 mm pitch), its transfer onto skin B (2.5 mm pitch), and the same press simulated on B. Each taxel shows the newtons on its own area, so B's smaller taxels read less each.

Input data and placement¶

frame: a Frame from skin A in newtons per taxel (units == "N"). ADC-count frames are rejected because the resampling operation sums force contributions. Convert counts with the source sensor's calibration first. Unread taxels (NaN) are treated as zero.
a, b: the two skins' manifests (a Manifest, a manifest.json path, a design .zip, or a dict). Each skin's taxels must sit on one square grid; the two grids can have different pitches.
center=True (default): the contact's centre of pressure goes to the centroid of B's taxels, or to target=(x, y) in B's millimetres. The centre uses only taxels above 10 % of the frame's peak, so noise does not pull it.
center=False: A's millimetres are used as B's, moved by shift_mm.
angle_deg: turn the contact by this angle.
kernel: "box" (default) or "tent" (smoother when B is finer than A).

Output and force conservation¶

A Frame on B: newtons per taxel in B's pixels order, units == "N", the input's t. imu is left empty, because it described A's board.

The resampling conserves force when the contribution outside the target is accounted for: force assigned to B plus force outside B equals the force on A. To see the lost part, or to apply one placement to many frames, use Transfer directly:

from opentac.transfer import Transfer

tr = Transfer.centered(a, b, frame_a.taxels)   # or Transfer(a, b, angle_deg=..., shift_mm=...)
fb = tr.force(frames_a)                        # (n_frames, n_A) -> (n_frames, n_B), N
tr.lost(frame_a.taxels)                        # N that landed outside B's taxels

Resample in physical coordinates¶

The transfer model assigns each taxel a square cell, centred on the taxel and one pitch wide. It treats the measured force as uniformly distributed over that cell. This piecewise-constant pressure approximation provides the basis for resampling in physical coordinates:

Place A on B. Rotate A's taxel positions by angle_deg, then shift them, so the contact sits where you want it on B.
Resample by area. Cut each A cell into small sub-cells, each carrying an equal share of that taxel's force. Each share goes to the B cell it lands in. This is one fixed matrix per placement: F_B = M @ F_A.
Drop what B cannot see. Shares that land outside B's taxels (past the outline, or in a gap) are dropped and counted in lost().

Left: the 4 mm cells of skin A, turned 20 degrees and laid over the smaller 2.5 mm cells of skin B. Right, two of A's cells close up: each holds a 7 by 7 pattern of share points, coloured by the B cell each lands in; in the cell that hangs over B's edge, the points outside B are crosses. — The three steps for the two skins of `examples/transfer.py`, with A turned 20°. The share points are the ones `Transfer` bins; of the cell over B's edge, 47% lands on B.

Evaluation and limitations¶

The paper evaluates transfer using hand-applied contacts on four skins with 409, 502, 96 and 174 taxels, at 2.0 and 4.0 mm pitch. Five objects were evaluated at forces from 1 to 8 N. Every method was given the target contact pose. Shape F1 compares thresholded contact regions, using approximately one quarter of each frame's peak; it evaluates contact shape and location rather than force-magnitude accuracy.

predicting skin B's frame from	median shape F1	frames
another press on skin B itself (repeatability limit)	0.80	1857
skin A, with `opentac.transfer`	0.75	5290
the same, scored only where A covers B	0.79	5286
skin A's counts, area-averaged onto B (no calibration)	0.71	5291
skin A's counts, stretched by row and column (no manifest)	0.63	5077

Contact shape and spatial extent. In a separate demo, three objects were pressed freehand on skins of 4.2, 3.1 and 2.0 mm pitch. An allen key read 10.3, 9.7 and 9.5 mm wide on the three.

A block pressed onto three skins of different shape and pitch. The manifests place all three recordings on a common physical coordinate grid in millimetres.

The mapping has several limitations:

Detail finer than A's taxels. Resampling a 4 mm grid onto a 2 mm grid cannot recover structure below the source resolution. The "tent" kernel smooths the result.
Contact area at coarse resolution. A partly covered taxel reports its whole area. In the demo, the 4.2 mm skin read 28–54 % more contact area than the 2.0 mm skin for the same object (tape roll: 1504 against 976 mm²). Transfer cannot recover the unresolved contact boundary within a coarse source cell.
Uncalibrated signal magnitude. The same objects peaked at about 200 counts per mm² on the 4.2 mm skin and 620–780 on the 2.0 mm skin. Counts depend on the design and the velostat, which is why transfer works in newtons.
Saturation. A taxel that hit the ADC maximum on A is a lower bound, and stays one on B.
Contact placement. The mapping does not infer the object's target pose. It centres the contact unless you give a target, shift or angle.