Tactile Data Collection Framework

Tactile Data Collection Framework

Collection starts with the downstream claim

A tactile dataset is evidence from a physical measurement process. The collection plan should state what the model must estimate, what object and contact population the claim covers, which failures matter, and what ground truth can support the claim.

Collecting frames before defining the grouping and split logic creates a high risk of leakage and wasted runs.

Minimum run record

Ground truth

Force labels may require a calibrated force-torque sensor. Pose may require markers, motion capture, or another reference. Slip labels may come from motion evidence or a physical model, each with its own failure modes.

The Sparsh dataset instructions provide concrete examples. Its force collection uses an ATI Nano17 reference, several indenter shapes, paired tactile frames, and normal and shear loading. Its pose collection synchronizes DIGIT images with tagged object pose and robot state.

Those examples are not a universal standard. A new task needs a ground-truth design that is valid for its range, timing, and physical consequence.

Pilot before scale

A pilot should cover the full path from sensor exposure through raw storage, labels, metadata, split assignment, reload, and one baseline model. Inspect images and traces together. Confirm units, coordinate frames, timestamp direction, dropped-frame behavior, and calibration state.

The pilot should intentionally create a few known failures, such as missed contact, overload, occlusion, or timestamp offset. The pipeline must preserve and label those events rather than silently cleaning them away.

Split before training

Group sequential frames by contact trajectory or run. If the claim concerns new objects, hold out objects. If it concerns replacement sensors, hold out sensor units. If it concerns durability, hold out later sessions or surface ages.

Store the split assignment as versioned data. Do not regenerate it opportunistically after viewing model results.

Dataset statement

The release record should describe intended uses, excluded uses, sensors, collection procedure, labels, known errors, missing populations, rights, maintenance, and contact for corrections. The Datasheets for Datasets paper provides a general documentation model that can be adapted to tactile measurement data.