RelCon: Relative Contrastive Learning for a Motion Foundation Model

Summary

RelCon (Apple, ICLR 2025) introduces relative contrastive learning — a self-supervised approach that learns a learnable distance measure capturing motif similarity and rotation invariance, then uses that distance as a soft semantic similarity signal to train a motion foundation model. Trained on 1 billion segments from 87,376 participants, RelCon achieves SOTA on HAR and gait metric regression. This is the largest-scale wrist-accel FM known as of mid-2025.