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

Yiquan Li; Aidan Bradshaw; Chenfeng Gao; Karan Ahuja

cs.ROOct 14, 2026

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

Authors: Yiquan Li, Aidan Bradshaw, Chenfeng Gao, Karan Ahuja

Organizations: Northwestern University

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 and a leaderboard are publicly available.

Figures & tables

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. Representation visualizations are schematic.

Pretraining			Evaluation
Regime	Dataset	Hours	Task	Dataset	Hours
HAR/Everyday	UT-Watch	45.3	Arm pose	Nymeria	300
	WISDM	44.7	Body pose	Nymeria	300
	RecoFit	43.0	Odometry	Nymeria	300
	PAMAP2	7.1	Gesture	OpenWatch	1.9
	SAMoSA	6.7	Retrieval	HumanML3D	28.6

Table 1: Pretraining sources and evaluation datasets. Hours give the size of each released dataset, from which pretraining samples about 280 hours (Section 3.1 ). Parentheses give class counts.

Figure 2: Downstream performance across task families. Frozen probing (left) and fine-tuning (right), with training from scratch shown in both panels. Activity averages macro-F1 across four datasets. Each axis is normalized to the best score across both panels, including scratch, using score/best for F1 and Recall@1 and best/error for error metrics. Axis labels give these reference values in the original units. Larger radii indicate better performance; the displayed range is 0.20–1.00.

Frozen probe
	Activity recognition					Gesture	PADS
Encoder / system	HHAR	MHEALTH	PRISM	TransfHAR	Mean	OpenWatch	F1	BA
InrtlWrist (Ours)	62.1	68.6	27.9	72.7	57.8	64.7	55.2	60.0
LIMU-BERT	61.2	67.6	25.1	71.1	56.2	61.7	50.5	58.4
SimMTM	57.5	61.4	26.7	66.8	53.1	55.9	48.7	59.3
FOCAL	44.9	63.0	21.9	47.0	44.2	48.1	40.7	53.9

Table 2: Activity, gesture, and clinical recognition. Macro-F1 and balanced accuracy (BA), reported as percentages with higher values better. Scores are evaluated per window for HAR, per clip for OpenWatch and per participant for PADS. Mean averages the four HAR datasets. Bold marks the best of six matched methods within each protocol, including ties. Gray rows show variants and external references excluded from ranking;

\ddagger

denotes released NormWear. Task-specific references are retrained DeepConvLSTM for HAR, Mix-Token for OpenWatch and the PADS authors’ BOSS pipeline with CatBoost.

	Arm pose				Full-body pose		Odometry
	Position (cm)		Angle ( ∘ )		MPJPE (mm)		ATE (m)		Vel. MSE ( $10^{-2}$ )
Encoder / system	Probing	Fine-tuning	Probing	Fine-tuning	Probing	Fine-tuning	Probing	Fine-tuning	Probing	Fine-tuning
InrtlWrist (Ours)	8.14	8.01	9.89	9.86	97.9	88.1	0.200	0.154	2.732	1.824
LIMU-BERT	8.14	8.18	9.86	9.97	98.5	89.7	0.208	0.158	2.856	1.885
SimMTM	8.47	8.44	10.27	10.29	103.6	92.9	0.224	0.163	3.064	1.946
FOCAL	8.71	8.66	10.70	10.69	108.4	95.9	0.240	0.177	3.334	2.218

Table 3: Kinematic estimation on Nymeria. Lower values are better for all metrics. Bold marks the best of six matched methods per column, including ties. Spanning cells report each reference under its stated training condition. The gray row reports retrained MobilePoser for pose and Air-IO for odometry, using their own inputs and training recipes. Velocity MSE is reported in units of

10^{-2}

	Frozen probing					Fine-tuning
	Gallery of 100			Full gallery		Gallery of 100			Full gallery
Encoder / system	Mean	I $\rightarrow$ T	T $\rightarrow$ I	I $\rightarrow$ T	T $\rightarrow$ I	Mean	I $\rightarrow$ T	T $\rightarrow$ I	I $\rightarrow$ T	T $\rightarrow$ I
InrtlWrist (Ours)	27.9	28.3	27.5	3.6	3.4	38.8	39.5	38.1	5.1	4.4
LIMU-BERT	24.9	25.3	24.5	3.1	2.5	36.6	37.2	36.0	4.4	3.9
SimMTM	17.2	17.4	17.0	1.7	1.2	32.4	32.7	32.1	4.0	3.4
FOCAL	8.6	9.1	8.0	0.7	0.7	22.4	22.7	22.0	2.1	1.9

Table 4: IMU–language retrieval on HumanML3D. Recall@1 (%

\uparrow

), with I and T denoting IMU and text. Mean averages both retrieval directions over 200 sampled galleries of 100 candidates; the full gallery contains 3,701 clips.

Panel of Figure 3 — Figure 3: Cross-task transfer. Source encoders (columns) are evaluated on each target (rows). Cells give raw scores; color indicates relative improvement over the first column, using a frozen random encoder in (a) and target training from scratch in (b). Green indicates improvement and orange-pink degradation, with color saturated at $\pm 40\%$ . MAE denotes InrtlWrist, frozen in (a) and fine-tuned in (b). Dashes omit same-task evaluations. HAR denotes PRISM; PADS uses participant macro-F1.

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Frozen probe
Encoder	pos err (cm) $\downarrow$	ang err ( ∘ ) $\downarrow$
InrtlWrist (Ours)	8.14	9.89
LIMU-BERT	8.14	9.86
SimMTM	8.47	10.27
FOCAL	8.71	10.70
RelCon	8.50	10.41

Appendix

Table 5: Arm pose: Nymeria real right-wrist IMU

\rightarrow

5-joint right-arm chain rotations. One training seed per cell. Task-specific systems are listed for orientation and are not ranked (Appendix B ). MobilePoser and IMUPoser also read wrist orientation, and must predict the ancestor joints that our arm-pose metric supplies.

Frozen probe
Encoder	MPJPE (mm) $\downarrow$
InrtlWrist (Ours)	97.85
LIMU-BERT	98.55
SimMTM	103.65
FOCAL	108.41
RelCon	104.29

Appendix

Table 6: Full-body pose: Nymeria real right-wrist IMU

\rightarrow

23-joint SMPL body (root excluded). One training seed per cell. Task-specific systems are listed for orientation and are not ranked (Appendix B ). MobilePoser and IMUPoser also read wrist orientation.

Frozen probe
Encoder	ATE (m) $\downarrow$	vel MSE $\downarrow$
InrtlWrist (Ours)	0.1999	0.02732
LIMU-BERT	0.2078	0.02856
SimMTM	0.2240	0.03064
FOCAL	0.2404	0.03334
RelCon	0.2153	0.02893

Appendix

Table 7: Odometry (wrist frame): Nymeria real right-wrist IMU

\rightarrow

pelvis velocity in the wrist (FLU) frame. Task-specific systems are listed for orientation and are not ranked (Appendix B ).

Frozen probe
Encoder	clip F1 (test) (%) $\uparrow$	clip acc (test) (%) $\uparrow$	win F1 (test) (%) $\uparrow$	win acc (test) (%) $\uparrow$
InrtlWrist (Ours)	64.7	67.8	49.4	53.6
LIMU-BERT	61.7	60.9	49.8	51.9
SimMTM	55.9	57.5	48.3	50.6
FOCAL	48.1	47.1	36.9	38.3
RelCon	38.5	39.1	39.9	42.5

Appendix

Table 8: Microgesture recognition on OpenWatch (five gestures plus background, real recordings, held-out test split), our encoder against training from scratch and the task-specific systems.

Frozen probe
		Gallery of 100				Full gallery
Encoder	R@1 mean $\uparrow$	i2t R@1 $\uparrow$	t2i R@1 $\uparrow$	i2t MRR $\uparrow$	t2i MRR $\uparrow$	i2t R@1 $\uparrow$	t2i R@1 $\uparrow$	i2t MRR $\uparrow$	t2i MRR $\uparrow$
InrtlWrist (Ours)	27.9	28.3	27.5	42.5	41.8	3.6	3.4	9.0	8.2
LIMU-BERT	24.9	25.3	24.5	39.2	38.4	3.1	2.5	7.8	6.6
SimMTM	17.2	17.4	17.0	30.3	29.8	1.7	1.2	5.0	4.2
FOCAL	8.6	9.1	8.0	19.9	18.6	0.7	0.7	2.3	2.2

Appendix

Table 9: IMU–text retrieval on HumanML3D (simulated wrist IMU, test split), where the task-specific row is a wrist-only UniMTS retrained from scratch whose text-to-IMU direction is at chance because of hubness.

Frozen probe
	Per subject		Per window
Encoder	macro-F1 $\uparrow$	bal-acc $\uparrow$	bal-acc $\uparrow$	macro-F1 $\uparrow$	acc $\uparrow$
InrtlWrist (Ours)	55.2	60.0	48.3	44.1	47.9
LIMU-BERT	50.5	58.4	49.0	43.8	46.7
SimMTM	48.7	59.3	48.0	40.3	41.6
FOCAL	40.7	53.9	39.2	31.3	31.5

Appendix

Table 10: Parkinson’s classification on PADS (healthy, Parkinson’s or differential diagnosis, all values in %), with the dataset authors’ BOSS classifier [ 65 ] re-run under this protocol as the task-specific row.

Frozen probe
	HHAR			MHEALTH			PRISM			TransfHAR			Mean
Encoder	macro-F1 $\uparrow$	bal-acc $\uparrow$	acc $\uparrow$	macro-F1 $\uparrow$	bal-acc $\uparrow$	acc $\uparrow$	macro-F1 $\uparrow$	bal-acc $\uparrow$	acc $\uparrow$	macro-F1 $\uparrow$	bal-acc $\uparrow$	acc $\uparrow$	macro-F1 $\uparrow$
InrtlWrist (Ours)	62.1	62.7	61.9	68.6	70.0	68.3	27.9	27.2	67.0	72.7	72.2	73.0	57.8
LIMU-BERT	61.2	60.8	59.9	67.6	68.1	66.3	25.1	25.1	66.0	71.1	70.3	71.4	56.2
SimMTM	57.5	57.8	56.8	61.4	63.1	62.5	26.7	26.3	66.9	66.8	67.2	68.7	53.1
FOCAL	44.9	46.7	45.8	63.0	63.0	63.4	21.9	23.2	65.9	47.0	48.1	48.9	44.2

Appendix

Table 11: Activity recognition (macro-F1, %) on the HHAR, MHEALTH and PRISM, which are subject-disjoint, and TransfHAR [ 10 ] , which is split in time, with their mean.

Participant ids in the release	50
Distinct identities after merging duplicate ids	43
Duplicate-identity groups (ids)	39, 41, 42, 46, 47, 48, 49; 40, 43
Byte-identical files / groups / ids involved	254 / 102 / 9
Recordings before / after exact de-duplication	2,330 / 2,178
Augmented ids that are mirrored copies of real ids	35 of 35 (ids 51–85; 32 distinct sources)
Clip macro-F1 (%)	release split / canonical split

Appendix

Table 12: OpenWatch release audit. Duplicates are exact: files are grouped by content hash. Bottom: clip macro-F1 (%) on the release’s participant-id split, where test id 48 is a byte-identical copy of training id 46, and on the canonical-identity split used by IMU-Bench. The two splits have different test participants, so the difference mixes leakage with test-set composition (per-participant analysis in the text).

Target	Frozen encoder	metric	gain	metric	gain
		session split (Fig. 3 )		participant split
Arm pose (cm $\downarrow$ )	never-trained encoder	8.46	—	8.82	—
	InrtlWrist	8.14	$+3.8$	8.68	$+1.6$
	source: full-body pose	8.13	$+3.8$	8.62	$+2.3$
	source: odometry	8.26	$+2.4$	9.00	$-2.0$
Full-body pose (mm $\downarrow$ )	never-trained encoder	105.0	—	107.9	—

Appendix

Table 13: The kinematic block of Figure 3 repeated on three disjoint thirds of the Nymeria participants, one per task, so that no encoder has seen a participant that another task is evaluated on. Frozen probes only. Gain is the improvement over the never-trained encoder of the same split ; the raw metrics of the two splits are not comparable, because each third carries a third of the data and every cell, source encoders included, is retrained inside it.

Target	Scratch	MAE alone	MAE $+$ pose
Odometry (ATE, m $\downarrow$ )	0.162	0.154	0.149
Microgesture (clip F1 $\uparrow$ )	60.8	54.2	65.0

Appendix

Table 14: Joint fine-tuning with full-body pose. Both MAE conditions start from the same pretrained encoder. Scores use per-task checkpoints. Bold: best per row.

Figure 6: t-SNE [ 64 ] of the reference InrtlWrist encoder’s frozen embeddings (cosine distances): 12,961 windows from eighteen sources, at most 800 each. The 9,139 windows with one of 261 activity labels are colored by semantic family and labeled by family and, where names fit, by activity; gray is the rest, including OCDetect and EatingSpeed, two pretraining sources without activity labels. The projection is illustrative; every statement in the text rests on neighborhood statistics.

Figure 7: Each probe of the grid against the number of motion modes at five pretraining volumes on the five-source pool, every model trained for 50,000 steps; Figure 5 is its average. Modes are added largest first ( $K{=}1$ is 98 % EatingSpeed). Activity, Parkinson’s (window macro-F1) and gesture panels use the benchmark probes on evaluation sets outside the pool; the arm, full-body and odometry panels use the grid’s own lighter probes and recipe, so they show trends within the grid only.

Wearable Sensing

Wearable sensing uses body-worn devices to measure motion, physiology, and interaction for health, activity, and assistive applications.

See all 129 papers

Physical Self-Supervised Learning: IMU Sensing without Manual Labels
Jul 20, 2026Yuyang Leng, Renyuan Liu, Shaohan Hu +4Self-Supervised Learning Physics-Informed ML
Richard · George Mason University Fairfax, VA, USA · Global Technology Applied Research, JPMorgan Chase New York, NY, USA
AnyMo: Geometry-Aware Setup-Agnostic Modeling of Human Motion in the Wild
May 21, 2026Baiyu Chen, Zechen Li, Wilson Wongso +5Human Activity Recognition Wearable Sensing
The University of New South Wales · The Hong Kong University of Science and Technology (Guangzhou) · The Hong Kong University of Science and Technology
ArmPoser: Real-Time, Calibration-Free Arm Pose Estimation from Smartwatch IMU
Sep 8, 2026Bishnu Dev, Vasco Xu, Xi-Aan Loh +3Wearable Sensing Human Pose Estimation
Mohamed bin Zayed University of Artificial Intelligence · University of Chicago · Northwestern University

Jul 20, 2026J/K move · Enter open · S save

Physical Self-Supervised Learning: IMU Sensing without Manual Labels

Yuyang Leng, Renyuan Liu, Shaohan Hu, Peijun Zhao+3

Richard · George Mason University Fairfax, VA, USA · Global Technology Applied Research, JPMorgan Chase New York, NY, USA

Self-Supervised Learning Physics-Informed ML Inertial Navigation

Open paper AskPDF arXiv

Deep neural networks have become a promising approach for IMU-based sensing, but their scalability is fundamentally limited by costly labeled data and poor robustness to heterogeneous devices, placements, and users. Existing unsupervised and self-supervised methods reduce but do not remove this dependence, still requiring labeled data for domain adaptation and largely ignoring known physical structure. We propose physical self-supervised learning, an autoencoder-style paradigm for label-free IMU sensing. We replace the conventional neural decoder with an auto-adaptive physics decoder, a learnable family of kinematic equations that enforces explicit physical structure while adapting across environments, and adopt a hybrid two-stage IMU encoder with reconstruction in a structured latent space to mitigate sensor noise. Our framework further introduces probabilistic frequency-spatial constraints to disentangle sensor and object motion, a multi-view kinematic tree to exploit sparse physical self-supervised signals, and an uncertainty-aware formulation to handle the inherent ambiguity of IMU inference. Evaluated on inertial tracking and full-body motion capture over public datasets and realistic deployments, physical self-supervised learning reduces errors by up to 5x for tracking and 4x for motion capture in challenging generalization scenarios, consistently outperforming state-of-the-art supervised and self-supervised baselines without any labels. Our code is available at https://github.com/YuyangLeng/physical-ssl-imu-label-free

Abstract

Wearable Sensing

Explore similar work

Physical Self-Supervised Learning: IMU Sensing without Manual Labels

AnyMo: Geometry-Aware Setup-Agnostic Modeling of Human Motion in the Wild

ArmPoser: Real-Time, Calibration-Free Arm Pose Estimation from Smartwatch IMU