SPRIIA study of persistent information
Formation

Baxter

A grasp can reveal hardness—or shape.

Real tactile recordings test whether changing the shared relation changes which material or shape information becomes accessible.

Real tactile datarelation-dependent readout
Motion preview
Recorded tactile signalsData: Youssef Amine, Christian Gianoglio and Maurizio Valle · CC BY 4.0. Source & changes
Across independent experiences

What persists.

Hardness under one relation, or shape under the other.

Across distinct realizations

What changes.

The other factor, with a separate grasp interaction.

Experience → prediction

The task.

A separate tactile grasp history.

→ Future tactile representation.

Evaluated instance

The comparison.

Predictive encoder · Align + Cross.

Controls. Balanced Random relation.

Interpreting the evidence

What stays fixed in the test.

Query/donor marginals, architecture and training budget are matched. Hardness-sharing changes shape; shape-sharing changes hardness.

What this setting establishes.

Hardness-based pairing favors cross-shape hardness readout, whereas shape-based pairing favors shape readout.

Scope of the evidence

Hardness and shape are not both held constant. The source lacks pre-contact frames and action sequences; unavailable static-leakage tests are not approximated.

Full result record

Measurements and comparisons.

Reported results are kept with their own populations and conditions. Development, held-out and sealed comparisons remain separate.

01 / Study Table 19

Relation-selective tactile organization

Held-out grasps from 6 known configurations; 3 frozen source seeds

Cross-shape hardness accuracy; leave-one-hardness-level-out shape balanced accuracy ↑

Relation-selective tactile organization
Training relationHardness G_H↑Shape G_S↑
Random0.280 ± 0.0550.435 ± 0.032
R_H (same hardness)1.000 ± 0.0000.500 ± 0.000
R_S (same shape)0.333 ± 0.0001.000 ± 0.000

Bold marks hardness accuracy under hardness-sharing and shape accuracy under shape-sharing.

  • R_H (same hardness)Confirmation holds out grasps from six known configurations; it does not hold out all object configurations.

Frozen confirmation means ± sample SD across three sources. Hardness accuracy transfers across shape; shape balanced accuracy leaves out one hardness level. Confirmation holds out grasps from six known configurations, rather than all object configurations.

02 / Study Table D.3

Input recoverability and registered selectivity contrasts

Six known hardness × shape configurations; validation/confirmation contrasts positive in every source

Classification accuracy; relation-specific contrast

Input recoverability and registered selectivity contrasts
MeasurementInput/conditionReported value
Hardness accuracyEight input samples0.954 ± 0.003
Hardness accuracyFull peak1.000 ± 0.000
Shape accuracyBoth input lengths1.000
Registered hardness contrast C_HValidation and confirmation0.667
Registered shape contrast C_SValidation and confirmation0.500

Bold marks the two registered relation-selectivity contrasts.

The dashed rule separates input recoverability from representation selectivity.

  • Registered hardness contrast C_H · Validation and confirmationInput accuracy and relation contrasts are different measurements; their numerical scales are not a shared ranking.

Input recoverability and registered relation-selectivity contrasts are distinct measurements. Both contrasts are positive in every source on validation and confirmation within six known configurations. Forty-sample files are alternate windows from the same acquisitions.