Research map / Large language models / AI and society

Inertia-1: An Open Exploration of Wearable Motion Foundation Models

NeurIPS 2026 · Large language models › AI and society · approach: AI for health and public decisions · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Inertia-1 is a large open exploration of wearable accelerometer foundation models, comparing 10 pretraining methods, sensing configurations, and model/data scale across 15 datasets (HAR, freezing-of-gait, disease prediction).

Problem

Wearable motion foundation models are fragmented: pretraining objectives, sensing configurations (sampling rate, window length, axes, placement), dataset scale, and downstream tasks are studied in isolation under inconsistent protocols, so it is unclear which design choices matter.

Approach

A controlled, unified framework that evaluates ten representative SSL and supervised methods under shared data, sensing, and evaluation settings, varying one design axis at a time over a large corpus of NHANES and UK Biobank accelerometry, and releasing the results as an open cookbook.

Key results

• Self-supervised pretraining consistently outperforms supervised training from scratch across task families, but no single objective dominates all tasks.
• SSL gains over supervised training are 4.0 AUROC points on HAR, 16.4 on FoG, and 19.7 on disease prediction.
• Triaxial inputs consistently outperform vector-magnitude inputs, and pretrained representations remain competitive down to 1Hz for HAR.

Surprising findings

• Scaling model capacity from ~5M to ~100M parameters largely plateaus linear-probe performance, and full finetuning of larger models can degrade results.
• Pretrained representations remain competitive at 1Hz sampling for HAR, unlike prior studies reporting sharp drops below 10Hz.

Limitations

• Downstream evaluation is classification-based; continuous health prediction, regression, and forecasting are not covered.
• Detailed ablations focus on a subset of top-performing objectives rather than all ten methods.
• UK Biobank data are only available as precomputed vector-magnitude five-second epochs, limiting the scale experiments to lower-resolution signals.

Experimental setup

Datasets: NHANES UK Biobank CAPTURE-24 HAR70+ HARTH HHAR MHEALTH OPPORTUNITY PAMAP2 RecoFit WISDM WEAR

Benchmarks: CAPTURE-24 HAR70+ HARTH HHAR MHEALTH OPPORTUNITY PAMAP2 RecoFit WISDM WEAR Daphnet FoG OdayFoG

Metrics: AUROC AUPRC

Baselines: ViT (supervised) CNN (supervised) PatchTST SimCLR DINO SSL-Wearables RelCon Self-PAB

Builds on: PatchTST DINO SimCLR SSL-Wearables RelCon Self-PAB

Closest papers by meaning

• Scaling Wearable Foundation Models (ICLR 2025 · similarity 0.89)
• Beyond Sensor Data: Foundation Models of Behavioral Data from Wearables Improve Health Predictions (ICML 2025 · similarity 0.89)
• RelCon: Relative Contrastive Learning for a Motion Foundation Model for Wearable Data (ICLR 2025 · similarity 0.88)
• UniMTS: Unified Pre-training for Motion Time Series (NeurIPS 2024 · similarity 0.87)
• OpenMHC: Accelerating the Science of Wearable Foundation Models (NeurIPS 2026 · similarity 0.86)
• MoPFormer: Motion-Primitive Transformer for Wearable-Sensor Activity Recognition (NeurIPS 2025 · similarity 0.86)