The contact layer · learned representation

A tactile embedding, trained live in your browser

No hand-coded ridge or bore rules. A small network learns to represent touch straight from raw reconstructed depth patches — watch the embedding cluster by texture as it trains, then it recognizes a new press. The interpretable-scale cousin of tactile foundation models.

Step

Train loss

—

Accuracy

—

Params

—

Learned embedding — 12-D → PCA-2D · colored by texture

Training — loss ↓ & accuracy ↑

Recognize a new touch

A 100→24→12→5 MLP with real backprop, trained on labeled example presses. Self-supervised transformers (T3, AnyTouch) are the field frontier; this is the on-device, legible version — the point is learned features, not hand-tuned rules.