We present IMUDINO, the first foundation model (FM) for inertial sensing to achieve competitive performance across diverse sensing tasks including human activity recognition (HAR), health, motion capture, and odometry. Prior inertial FMs often focus on one subfield of inertial sensing and sometimes underperform state-of-the-art supervised methods (e.g., in HAR), making their scope considerably narrower than FMs for vision and natural language. Moreover, they often require fixed input length and sampling rate, and generalize poorly to unseen sensor locations. To build a general-purpose foundation model for inertial sensing, we adapt the DINO self-supervised learning (SSL) paradigm to inertial data through a set of semantics-preserving motion transformations that simulate variations in motion pattern and sensor setup. For model architecture, we use a plain transformer encoder with frequency-aware positional encoding that natively supports varying input length and sampling rate. This simple setup proved remarkably effective for learning global and dense features generalizable across users, sensor setups, tasks, and, in the case of robot odometry, embodiments. We hope IMUDINO can serve as a backbone for inertial sensing and kickstart the development of general-purpose inertial FMs.
Figures & tables
Figure 1: An overview of the IMUDINO training objective.
Type
Method
MoSense
HAPT
HHAR
DSADS
WEAR
OPPO.
Shoaib
Average
Supervised Baselines
DeepConvLSTM
93.39
82.00
91.90
83.25
75.43
40.88
97.13
80.57
A&D
95.79
81.95
94.17
86.12
77.10
46.29
97.79
82.74
IF-ConvT.
95.81
85.35
95.01
88.82
N/A
50.29
98.28
85.59
TSF
97.34
85.89
9 5.81
89.70
N/A
52.44
99.35
86.76
Self-supervised HAR FMs
SimCLR ∗
96.08
80.56
87.56
91.56
88.48
43.50
99.29
83.86
Yuan et al.
93.41
64.66
80.41
83.37
83.10
18.28
97.30
74.36
Table 1: Results on seven HAR datasets. IMUDINO consistently outperforms both supervised and SSL baselines. ∗ Implemented with IMUDINO data augmentation.
Model
AnDy
UNIPD
CIP
Natural Motion
Average
SIP ↓
Ang ↓
Pos ↓
SIP ↓
Ang ↓
Pos ↓
SIP ↓
Ang ↓
Pos ↓
SIP ↓
Ang ↓
Pos ↓
SIP ↓
Ang ↓
Pos ↓
DIP ( Huang et al., 2018 )
11.39
5.73
4.34
12.02
5.47
4.22
19.13
8.61
6.86
33.43
11.91
13.33
18.99
7.93
7.19
TransPose ( Yi et al., 2021 )
12.15
6.29
4.91
15.11
6.05
4.82
20.06
8.75
6.86
30.62
11.26
11.99
19.49
8.09
7.15
TIP ( Jiang et al., 2022 )
10.11
4.55
3.56
9.85
4.06
2.78
13.05
5.67
4.30
22.06
7.90
7.92
13.77
5.55
4.64
PIP ( Yi et al., 2022 )
9 .49
4.09
3.29
8 .90
3.59
2.66
12.68
5.52
4.12
19.62
7.49
6.93
12.67
5.17
4.25
DynaIP ( Zhang et al., 2024b )
8.93
3.45
3 .41
7.29
2.77
2.21
11.42
4.54
3.69
15.78
7 .18
5.83
10.86
4.49
3.79
Table 2: Inertial MoCap results. We report SIP error ( ∘ ), global angular error ( ∘ ), and joint position error (cm). Average denotes the macro-average over AnDy, UNIPD, CIP, and Natural Motion. † Backbone initialized from the pretrained IMUDINO checkpoint and finetuned.
Figure 2: Qualitative comparison of inertial HPE on a natural motion sequence from Virginia Natural Motion. From left to right: ground truth, DynaIP, and the three IMUDINO variants of Table 2 (MLP probe, transformer probe, and MLP probe with finetuned backbone). Numbers below each mesh denote the per-frame mean root-aligned global rotation error ( ∘ ). See Appendix K for more cases.
Model
Car
Dog
Drone
Human
Average
AVE ↓
ATE20 ↓
AVE ↓
ATE20 ↓
AVE ↓
ATE20 ↓
AVE ↓
ATE20 ↓
AVE ↓
ATE20 ↓
Score ↓
TartanIMU ∗
0.852
2.273
0.425
1.901
1.374
2.772
0.235
1.877
0.721
2.206
0.872
TartanIMU †
0 .199
0.999
0.129
0 .644
0.966
2.006
0 .107
0.814
0.350
1.116
0.429
IMUDINO-85M+MLP ∗
0.200
0 .998
0 .123
0.707
0 .543
1 .304
0.132
1.039
0 .249
1 .012
0 .333
IMUDINO-85M+Transformer ∗
0.141
0.760
0.092
0.501
0.499
1.285
0.105
0 .872
0.209
0.854
0.280
Table 3: Results on the TartanIMU Challenge. ∗ Platform-agnostic. † Platform expert.
Figure 3: Visualized results on the TartanIMU Challenge. ▲ denotes the starting point of each trajectory. IMUDINO+Transformer (green line) traces the ground truth (black dashed line) most closely. The TartanIMU results (blue line) are from the corresponding expert model for each platform.
Type
Setting
HAR
MoCap (MLP Probe)
Robot Odometry (MLP Probe)
mF1
SIP ↓
Ang ↓
Pos ↓
AVE ↓
ATE20 ↓
Score ↓
Training Objective
w/o patch-level objective
89.41
15.21
5.77
4.85
0.293
1.045
0.373
w/o KoLeo
89.40
14.50
5.31
4.58
0.252
1.004
0.335
Data Augmentation
w/o scaling
89.63
14.87
5.35
4.74
0.257
1.016
0.340
w/o rotation
88.66
15.67
5.52
4.93
0.253
1.036
0.339
w/o resampling
88.86
14.64
5.25
4.59
0.255
1.012
0.338
Table 4: Result of the ablation study. The results are averaged over multiple datasets and correspond to the Average columns in Tables 1 , 2 and 3 .
Figure 4: How CLS and patch tokens responds to perturbation. Our invariance-motivated view transformations induced in the CLS tokens the desired robustness to rotation and resampling. We further demonstrate in the rotation case that CLS robustness does not expand to patch tokens, which remain sensitive to rotation. Results aggregated over 700 examples from 7 HAR datasets.
Appendix figures & tables15 assets
Supplementary material from the paper’s appendix.
Appendix
IMUFormer-2.7M
IMUFormer-6M
IMUFormer-25M
IMUFormer-85M
Patch size
16
16
16
16
Token dim
192
256
512
768
Num heads
6
8
16
24
Num layers
6
8
8
12
Register Tokens
4
4
4
4
MLP type
GELU MLP
GELU MLP
GELU MLP
GELU MLP
Appendix
Table 5: The 4 IMUFormer models used in this paper.
Figure 5: IMUDINO feature visualization. Samples are colored by label. Activities from WEAR (left) and MotionSense (middle) form distinct clusters. Furthermore, for the walking class in MotionSense (right), each user forms a distinct cluster, indicating that IMUDINO is capable of capturing the differences in gait between individuals. Further results on gait recognition are in Appendix F .
Datasets
MotionSense
HAPT
HHAR
DSADS
WEAR
OPPORTUNITY
Shoaib et al. (2014)
Sampling Rate (Hz)
50
50
100 ∗
25
50
30
50
Window Size (s)
2.56
2.56
2.00
5.00
2.00
1.60
0.96
Overlap (s)
1.28
1.28
1.00
0.00
1.00
0.80
0.48
LR
1e-3
1e-3
1e-3
1e-4
3e-4
3e-4
3e-4
Epochs
100
100
100
100
50
50
50
Appendix
Table 6: Data processing protocol, LR, and the number of training epochs for each HAR dataset. We use the final epoch checkpoint for test set evaluation. ∗ Phones with different sampling rate are resampled to 100 Hz.
Backbone / Condition
Data Util.
Collab. Labeling Acc.
Activities Discovered
HAR mF1
W/o Personalization
-
-
-
62.6
Supervised ResNet
67.4
9 8.8
8.3 / 8.7
89.5
IMUDINO-6M
7 6.5
98.7
8.4 / 8.7
9 5.3
IMUDINO-85M
77.3
99.3
8.4 / 8.7
96.8
Appendix
Table 7: Results on ActivitySeeker. Using IMUDINO off-the-shelf improves all metrics, demonstrating its potential as a general-purpose inertial feature extraction model.
Method
GOTOV
UniMiB-SHAR
Daphnet FOG
whuGAIT
RMSE ↓
R 2
HAR Acc
Fall Acc
Spec.
Sens.
Acc
EER ↓
Official Baseline
1.35
0.45
56.58
97.05
0.869
0.781
93.75
-
IMUDINO-6M+MLP
0 .96
0 .58
8 3.96
9 9.59
0 .908
0 .671
9 6.55
3 .24
IMUDINO-85M+MLP
0.94
0.59
84.78
99.81
0.936
0.551
97.41
2.92
Appendix
Table 8: Results on four health and gait analysis datasets. Official baseline results are from the original dataset paper, whose experiment protocols are followed in IMUDINO experiments.
Dataset
Task
Protocol
whuGAIT
gait authentication
fixed split (20 unseen subjects)
UniMiB-SHAR
17-way HAR / fall detection
LOSO
Daphnet
freeze of gait (FOG) detection
LOSO, 4 s/0.5 s windows
GOTOV
energy expenditure (EE) regression
LOSO, 120 s → 10 s bins
Appendix
Table 9: Summary of the gait analysis and health evaluation. All experiments use frozen IMUDINO backbones with lightweight MLP probes.
Window
25 Hz
50 Hz
100 Hz
200 Hz
1.28 s
97.40
97.56
97.60
97.86
2.56 s
98.26
98.54
98.67
98.72
5.12 s
98.89
98.91
98.92
98.81
10.24 s
99.16
99.34
99.33
99.24
Appendix
Table 10: MotionSense mF1 of IMUDINO-85M with the 2-layer MLP probe for different input window lengths and sampling rates (resampled from the native 50 Hz). Supervised SOTA: 97.34 (2.56 s, 50 Hz, Zhang et al. (2026) ).
Figure 6: How IMUDINO models of different sizes perform in HAR, MoCap, and robot odometry.
Window strategy
AVE ↓
ATE20 ↓
Score ↓
MLP
Transformer
MLP
Transformer
MLP
Transformer
1 s, non-overlapping
0.283
0.262
1.103
0.992
0.373
0.341
1 s, 0.2 s stride (default)
0.249
0.209
1.012
0.854
0.333
0.280
Appendix
Table 11: Window-strategy ablation on the TartanIMU validation split (frozen IMUDINO-85M, macro-averaged over platforms). Denser training windows help both probes.
Figure 7: Results on a squatting sequence from the UNIPD test set.
Figure 8: Results on a throwing sequence from the UNIPD test set.
Figure 9: Results on a jumping sequence from the UNIPD test set.
Figure 10: Results on a lifting sequence from the AnDy test set.
Figure 11: Results on a circuit sequence from the CIP test set.
Figure 12: Results on a walking sequence from the UNIPD test set.
Time-series foundation models are pretrained sequence models designed to transfer across forecasting, classification, anomaly detection, and other temporal tasks.