Can One IMU Model Recognize Any Activity Across Any Sensor?

An inertial measurement unit sits inside every phone, watch, and fitness band. Researchers have spent years trying to build a single model that understands all of them.

The problem splits in two. Sensors differ in sampling rate and axis count. They also sit at different spots on the body. Labels differ too — one dataset calls it "walking," another "ambulating," a third "locomotion." HALO attacks both at once.

Heterogeneity baked into the tokenizer

Stage one pretrains an IMU encoder with three moving parts. Adaptive-pooling tokenization resamples any input stream to a fixed token length regardless of original frequency. Channel-independent feature extraction processes each axis separately so a three-axis watch and a six-axis research node share the same backbone. Contextualized sensor conditioning injects a free-text description of each channel — "left wrist accelerometer X" — directly into the embedding.

No hand-crafted alignment. No per-dataset adapter. The model learns to map heterogeneous raw signals into a common latent space.

Language alignment without hard labels

The second stage pairs the frozen encoder with a text encoder. Synonym-aware soft contrastive learning pulls matching IMU-text pairs together while pushing mismatches apart. "Walking" and "ambulating" receive similar gradients because their embeddings sit close in the language space. At inference, HALO computes cosine similarity between a test window and any candidate label — no classifier head, no fine-tuning.

Open-set recognition becomes a retrieval problem.

Thirty-five million parameters against three hundred forty-one

HALO carries roughly 35 million trainable parameters. MOMENT, the previous IMU foundation model, carries 341.2 million. On eight aggregate metrics across seven held-out datasets, HALO beats five state-of-the-art baselines on all eight. Under baseline-matched inputs it still leads on three of four settings.

Zero-shot open-set accuracy across all 87 training labels rises 13.7 percentage points.

Where the curve breaks

Two additional datasets exhibit severe distribution shift. Every model collapses zero-shot, HALO included. The paper does not quantify how many devices or subjects remain in the wild beyond the seventeen datasets used. A video demonstration shows real-time performance on a wrist-worn unit.