Wrist IMU benchmark

IMU-Bench: Benchmarking Wrist IMU Representations Across Kinematic, Semantic and Clinical Tasks

A benchmark of ten tasks from eight datasets for studying IMU multi-task performance and cross-task transfer, with InrtlWrist, a reusable pretrained wrist-IMU encoder.

Yiquan Li*, Aidan Bradshaw*, Chenfeng Gao, Karan Ahuja

Northwestern University · *Equal contribution

arXiv PDF Code Leaderboard Documentation InrtlWrist weights
10tasks
8datasets
6task families
6matched SSL methods
3evaluation protocols
3,855 hof IMU data across 18 datasets
Overview

From single-task evaluation to a map of what a wrist representation carries

Abstract. Wrist-worn inertial measurement units (IMUs) provide a common sensing modality for tasks ranging from activity and gesture recognition to pose estimation, inertial odometry and clinical motor assessment. Existing wearable benchmarks evaluate performance across multiple tasks, but leave open how supervision for one task affects transfer to others and whether these tasks benefit from a shared representation. We present IMU-Bench, a benchmark of ten tasks from eight datasets for studying IMU multi-task performance and cross-task transfer. We compare six self-supervised methods pretrained on unlabeled recorded and simulated wrist motion, evaluating classification, pose and velocity estimation, and IMU–language retrieval. We find that our masked autoencoder, InrtlWrist, achieves the best or joint-best frozen-probe performance on nine of ten tasks, and fine-tuning improves on training from scratch on nine of ten tasks, including full-body pose, odometry and retrieval. Across transfer and joint-training experiments, supervision that benefits one target can reduce performance on others, with supervised sources transferring within kinematic estimation while, under frozen probing, none approaches the pretrained encoder on language retrieval, and joint-training gains that come at a cost to the paired task. These findings distinguish broad downstream usefulness from complementary task supervision. Code, task cards, the InrtlWrist weights and a leaderboard are publicly available.

Why this is hard. Different downstream tasks reveal distinct aspects of the same underlying motion, each offering only a partial view. Self-supervised wearable models are usually evaluated on activity recognition or health classification, under their own splits, decoders and pretraining corpora, while pose estimation and inertial odometry live in separate literatures with their own architectures and metrics. Whether these tasks draw on a common representation, and whether supervision for one is complementary to another, cannot be read off the existing literature.

Our approach. IMU-Bench fixes the backbone, one light decoder per task, the data splits, the training budget and the model-selection rule, and audits every pretraining corpus against every evaluation set. Six self-supervised objectives are compared on the same corpus under frozen probing, fine-tuning and training from scratch, beside retrained task-specific systems. Supervised source-task transfer and joint training then measure whether task supervision is complementary.

Overview of IMU-Bench in three panels: a sunburst of the ten evaluations, an encoder–decoder pretrained on label-less wrist and simulated data and then probed or fine-tuned, and a shared representation fanning out to odometry, activity, pose, gesture, text alignment and clinical tasks.
Figure 1. Overview of IMU-Bench. Ten tasks from eight datasets evaluate representations learned from single-wrist accelerometer and gyroscope signals. Pretraining uses unlabeled recorded and simulated motion. Frozen probing and fine-tuning measure downstream performance, while source-task transfer and joint training examine the effects of supervision across tasks.
  • Oct 2026Project website is live.
  • Oct 2026Paper on arXiv: 2610.19481.
  • Oct 2026InrtlWrist checkpoint released on the Hugging Face Hub; code, task cards and per-cell results on GitHub.
The benchmark

One controlled space over tasks, encoders, protocols and data

Every number in the paper comes from one registry-driven harness: a task card fixing data, split, input window, decoder, recipe, selection rule and scorer, and an encoder card fixing checkpoint, architecture, input normalization and pretraining corpus.

01

Tasks

Ten tasks from eight datasets in six families: pose, odometry, microgesture, IMU–text retrieval, Parkinson's disease and activity recognition.

02

Encoders

Six self-supervised objectives on one 12-layer, width-384 ViT (21 M parameters) and one corpus of 197,094 unlabeled windows.

03

Protocols

Frozen probing, fine-tuning and training from scratch, each with the same decoder, optimizer, budget, split and selection rule per task.

04

Audit

Every pretraining corpus checked against every evaluation set; duplicated and mirrored participants merged; sensor units verified; evaluation sets excluded from pretraining.

Tasks and datasets
TaskDatasetHoursOutputMetricSplit
Arm poseNymeria3005 right-arm joint rotations per 0.32 s patchposition error, cm ↓recording session
Full-body poseNymeria30023 SMPL joint rotations per 0.32 s patchroot-relative MPJPE, mm ↓recording session
Inertial odometryNymeria300pelvis velocity in the wrist frame, integrated to a trajectoryATE, m ↓recording session
MicrogestureOpenWatch1.95 gestures + background, voted per clipclip macro-F1 ↑43 canonical subjects
IMU–text retrievalHumanML3D (simulated wrist)28.6alignment with CLIP ViT-B/32 text embeddingsmean Recall@1, galleries of 100 ↑official test split
Parkinson's diseasePADS35.9healthy / Parkinson's / differential diagnosis per participantparticipant macro-F1 ↑469 subjects, disjoint
Activity (6 classes)HHAR4.82.56 s window labelmacro-F1 ↑subject-disjoint
Activity (12 classes)MHEALTH1.92.56 s window labelmacro-F1 ↑subject-disjoint
Activity (5 classes)PRISM16.32.56 s window labelmacro-F1 ↑subject-disjoint
Activity (60 classes)TransfHAR2.02.56 s window labelmacro-F1 ↑80/20 in time

Hours are the sizes of the released datasets. Pretraining sources (released hours): UT-Watch 45.3, WISDM 44.7, RecoFit 43.0, PAMAP2 7.1, SAMoSA 6.7 and Shoaib 6.5 for activity recordings; OCDetect 2,600 and EatingSpeed 525.9 for free-living and meal sessions; Motion-X++ 40.4 and BONES-SEED 144.2 for wrist motion simulated from motion capture. The pretraining set uses 197,094 windows of 10.24 s at 50% overlap, about 280 hours. Summed over their released sizes, the ten pretraining sources hold 3,464 hours and the eight evaluation datasets 391 hours (Nymeria counted once), 3,855 hours in all.

Pretraining objectives
Masked reconstructionPredict masked accelerometer and gyroscope patches. InrtlWrist alternates frame-wise attention across the six channels with global attention over all tokens; the decoder is discarded after pretraining.
InrtlWristLIMU-BERT
Reconstruction + contrastiveReconstruct from neighbors in a learned similarity space while pulling augmented views together.
SimMTM
ContrastiveContrastive, orthogonality and temporal losses (FOCAL) or a relative contrastive loss over motion neighbors (RelCon).
FOCALRelCon
Transformation predictionClassify which augmentations were applied to a window, the SSL-Wearables recipe, with cross-entropy.
SSL-Wearables
ControlsA never-trained random frozen encoder, a raw-input control with zero transformer layers, and the same backbone trained from scratch per task.
random frozenraw inputscratch
ReferencesA released pretrained model run unchanged, and task-specific systems retrained under each task's protocol with their own inputs and recipes.
NormWearMobilePoserAir-IOMix-TokenUniMTSDeepConvLSTMBOSS + CatBoost
Default setup
one wrist6 channels: acc m/s², gyro rad/sforward–left–up frameper-channel z-score (released encoders; submissions choose their own normalization)0.32 s patches1.28–10.24 s crops12-layer ViT, width 384, 21 M24 octahedral rotations in pretraininglinear heads for classificationwidth-192 attention decoders for kinematicsclass-weighted cross-entropyone A100 / H100 per run
Findings

What the benchmark reveals

Among the six matched methods, one pretrained encoder is broadly useful. Which supervision is compatible with which target is a separate question, and the benchmark answers it pair by pair.

4.1Activity, gesture and clinical recognition

Frozen InrtlWrist leads the six matched methods on all four activity datasets, with 57.8 mean macro-F1 against 56.2 for LIMU-BERT and 53.3 for training from scratch. Fine-tuning raises the mean to 63.9, with the largest gains on TransfHAR (72.7 to 86.1) and PRISM (27.9 to 36.8), while HHAR favors frozen probing. A random frozen encoder reaches 26.2 under the same readout.

On OpenWatch gestures, frozen InrtlWrist reaches 64.7 clip F1, ahead of every fine-tuned matched method; fine-tuning it drops the score to 54.2 while SimMTM leads at 61.9. On PADS, frozen InrtlWrist reaches 55.2 participant F1, below SSL-Wearables at 57.4 and only 3.5 points above the random frozen encoder, a weaker signal of pretraining benefit than on activity and gesture.

57.8 → 63.9activity mean F1, frozen → fine-tuned
64.7frozen gesture F1, above every fine-tuned method
55.2Parkinson's participant F1, frozen (scratch 29.9)
Radar charts of frozen-probe and fine-tune performance across seven task axes for the six matched methods and training from scratch.
Figure 2. Downstream performance across task families under frozen probing (left) and fine-tuning (right), with training from scratch in both. Each axis is normalized to the best score across both panels; larger radii are better.

4.2Kinematic estimation

Frozen InrtlWrist beats the random encoder on all three kinematic targets but trails training from scratch on full-body pose and odometry. Fine-tuning reduces full-body MPJPE from 97.9 to 88.1 mm and odometry ATE from 0.200 to 0.154 m, past the scratch baseline at 92.7 mm and 0.162 m, and gives the lowest error among the six matched methods on all five fine-tuned metrics. Arm pose changes little: frozen InrtlWrist and LIMU-BERT tie at 8.14 cm, and fine-tuning reaches 8.01 cm.

Retrained MobilePoser and Air-IO reach 91.9 mm and 0.180 m with their own inputs and recipes; MobilePoser also receives wrist orientation. These results use session-held-out Nymeria recordings with recurring participants, and odometry uses the ground-truth initial orientation over short windows.

88.1 mmfull-body MPJPE, fine-tuned (scratch 92.7)
0.154 modometry ATE, fine-tuned (scratch 0.162)
8.01 cmarm position error, fine-tuned (scratch 8.21)
EncoderArm, cm ↓probe / fine-tuneBody MPJPE, mm ↓probe / fine-tuneOdometry ATE, m ↓probe / fine-tune
InrtlWrist (ours)8.14 / 8.0197.9 / 88.10.200 / 0.154
LIMU-BERT8.14 / 8.1898.5 / 89.70.208 / 0.158
SimMTM8.47 / 8.44103.6 / 92.90.224 / 0.163
FOCAL8.71 / 8.66108.4 / 95.90.240 / 0.177
RelCon8.50 / 8.55104.3 / 92.10.215 / 0.167
SSL-Wearables8.42 / 8.57101.7 / 90.60.220 / 0.166
Random encoder, frozen8.46105.00.233
Same backbone, scratch8.2192.70.162
Task-specific system9.74 (IMUPoser)91.9 (MobilePoser)0.180 (Air-IO)

4.3IMU–language retrieval

InrtlWrist has the highest mean Recall@1 among the six matched methods under both protocols: 27.9 frozen against 24.9 for LIMU-BERT, 23.2 for training from scratch and 9.5 for the random frozen encoder, and 38.8 after fine-tuning. The advantage persists in the full 3,701-clip gallery, where fine-tuned InrtlWrist reaches 5.1 IMU-to-text and 4.4 text-to-IMU Recall@1 against 2.1 in both directions for scratch.

All encoders receive paired motion–text supervision during downstream training, so these numbers measure how well a pretrained representation supports learning alignment, not zero-shot language alignment.

27.9frozen Recall@1 (scratch 23.2, random 9.5)
38.8fine-tuned Recall@1, galleries of 100
5.1 / 4.4full-gallery Recall@1, IMU→text / text→IMU

4.4Cross-task transfer

Each source encoder is trained from scratch on one of seven tasks and evaluated on the other six with a new decoder under the target's split and budget. Under frozen probing, supervised sources beat a random frozen encoder on 28 of 42 pairs, with 15 gains above 10%. Kinematic targets benefit most from kinematic sources: full-body supervision reduces odometry ATE from 0.233 to 0.186 m. All six sources raise PRISM macro-F1 from 16.1 to 21.3–27.4 but fall below the random encoder on PADS, and none approaches InrtlWrist on gesture or IMU–text. Overall, 35 of 42 frozen transfers stay below target training from scratch.

With fine-tuning, 22 pairs improve over scratch and 20 decline. Every source improves IMU–text and PADS, with PRISM supervision reaching 58.5 PADS F1 against 51.9 for direct InrtlWrist fine-tuning. Four of six sources improve full-body pose, and odometry initialization reaches 87.9 mm MPJPE. No source improves odometry, only odometry supervision improves arm pose, and every source reduces gesture F1.

Heatmap of frozen cross-task transfer, sources in columns and targets in rows, colored against a random frozen encoder.
Figure 3a. Frozen transfer, colored by relative improvement over a random frozen encoder.
Heatmap of cross-task transfer with fine-tuning, sources in columns and targets in rows, colored against training from scratch.
Figure 3b. Transfer with fine-tuning, colored by relative improvement over target training from scratch. Green improves, orange-pink degrades, saturated at ±40%. MAE denotes InrtlWrist.

4.5Decoder capacity, pretraining data and joint supervision

Decoder capacity. With a linear readout, pretraining is worth 9.7 mm of full-body MPJPE over a random frozen encoder, and 7.1 mm with the benchmark decoder. Increasing decoder width changes the gap little, while adding temporal layers reduces it below 1 mm; a 12-layer odometry decoder lets the random encoder outperform every pretrained one. The benchmark therefore uses light readouts.

Pretraining data and motion coverage. Holding architecture, probes and budget fixed, uniform draws from 0.1% to 100% of a 1,959,288-window pool perform alike at 50,000 updates, and the 0.01% subsets trail the 0.1% subsets at every coverage level. Restricting pretraining to subsets of 1,000 motion clusters hurts most when few clusters are available. These findings are conditional on the budget and sampling scheme.

Joint task supervision. Fine-tuning InrtlWrist jointly on full-body pose and a partner task improves odometry ATE from 0.154 to 0.149 m and gesture F1 from 54.2 to 65.0, but increases body MPJPE and lowers frozen IMU–text Recall@1.

Full-body MPJPE of a frozen pretrained and a frozen random encoder as decoder capacity grows; the gap shrinks once temporal layers are added.
Figure 4. Decoder capacity. Frozen-encoder MPJPE (mm, lower is better); Δ is pretrained minus random.
Eight-probe mean against pretraining volume for different numbers of allowed motion clusters, after 50,000 updates.
Figure 5. Motion coverage. Eight-probe mean after 50,000 updates; per-probe changes from the full pool are normalized by its gain over random and clipped to ±100%.
Paper

Read & cite

The arXiv and PDF links will be added when the preprint is public.

arXiv PDF Code InrtlWrist weights

Every number on this page is produced by the released harness from task cards and encoder cards; the per-cell result files, secondary metrics and the corpus audit are part of the release. 

@article{imubench2026,
  title   = {IMU-Bench: Benchmarking Wrist IMU Representations
             Across Kinematic, Semantic and Clinical Tasks},
  author  = {Li, Yiquan and Bradshaw, Aidan and Gao, Chenfeng and Ahuja, Karan},
  journal = {arXiv preprint arXiv:2610.19481},
  year    = {2026}
}