Learning inertial odometry for a new platform usually requires ground-truth trajectories. We show that an inertial odometry model can instead be trained from scratch by a myopic teacher that integrates the IMU over each short window independently, a scheme we call integration self-supervision. Interestingly, although supervised only by short-window velocity changes, the student learns to directly predict velocity, and does so more accurately than its integration teacher. Pose graph distillation then refines the model with sequence-consistent targets from a pose graph over many windows. Training on a new platform therefore no longer requires collecting ground-truth trajectories, but only calibrating the IMU and recording data during normal operation. Across six datasets spanning pedestrians, drones, quadrupeds, and a humanoid, the same recipe with no velocity or position labels trains models that come within 0.07~m/s average velocity error of a same-architecture supervised baseline.
Figures & tables
Fig. 2 : Integration Self-Supervision trains the motion network to match the body-frame velocity change predicted by the frozen integration teacher under a squared velocity-change loss, supervising only the velocity mean. Pose Graph Distillation (PGD) builds a full-sequence pose graph from the teacher’s IMU factors and the network’s velocity factors. The graph posterior vopt then supervises the network through an L2 loss. Velocity Regularization , Lreg , is used over detected static frames.
Fig. 3 : Motion network architecture used for both stages of self-supervised training and the supervised AirIO † baseline.
TLIO
EuRoC
Blackbird
MIPO
GrandTour
Humanoid
Method
AVE ↓
ATE ↓
AVE ↓
ATE ↓
AVE ↓
ATE ↓
AVE ↓
ATE ↓
AVE ↓
ATE ↓
AVE ↓
ATE ↓
Supervised Learning
AirIMU [ 22 ]
20.83
–
13.42
–
46.95
–
1.958
39.07
2.448
334.8
0.618
–
RoNIN [ 17 ]
0.160
2.420
0.661
9.011
0.948
16.49
0.157
0.429
0.102
3.894
0.163
1.989
TLIO [ 19 ]
0.161
2.872
0.656
8.541
0.972
15.54
0.160
0.487
0.106
4.327
0.163
1.737
AirIO [ 23 ]
0.081
1.890
0.335
3.720
0.318
6.100
0.120
0.310
0.068
3.190
0.087
1.260
TABLE II : Self-Supervised Learning from Scratch. We report AVE (m/s) and ATE (m) across six datasets. Supervised rows are trained with ground-truth labels and serve as reference; ours are trained from random initialization without velocity or position labels.
Dataset
Time
Length
Max Speed
TLIO [ 19 ]
31.6 h
42.3 km
6.0 m/s
EuRoC [ 6 ]
22.5 min
892 m
2.9 m/s
Blackbird [ 1 ]
5.0 h
28.4 km
8.7 m/s
MIPO [ 28 ]
58.6 min
1.4 km
6.6 m/s
GrandTour [ 12 ]
5.65 h
10.3 km
1.5 m/s
Humanoid
1.31 h
1.9 km
3.9 m/s
TABLE I : Evaluation datasets. Six datasets on six distinct embodiments, spanning different IMU grades and motion patterns.
Table 5
Fig. 5 : Supervision window ablation on TLIO. Velocity-change error ∥Δv^ijnet−Δvij⋆∥2 [m/s] over evaluation and training horizons, where Δvij⋆ is the ground-truth body-frame velocity change; red cells underperform and blue cells outperform the integrator.
Fig. 6 : Cross-dataset transfer over five datasets. Each cell reports AVE for a model trained on one dataset and evaluated on another. Models trained on the target dataset outperform transferred models, showing the importance of training on the target distribution. The traces show a representative TLIO–Blackbird mismatch.
Figure 8
Dataset (source)
Reference tilt
Estimated tilt
Tilt error
TLIO (gyro)
0.109 / 3.21
0.133 / 3.81 (+22%)
<1∘
MIPO (AHRS)
0.182 / 1.63
0.235 / 1.41 (+29%)
0.9∘
TABLE IV : Attitude source in the teacher. Test AVE (m/s) / ATE (m) of Ours with the reference tilt and with a tilt estimated from the IMU alone: the gyroscope integrated from the sequence start on TLIO, the onboard AHRS on MIPO. Parentheses: AVE change relative to the reference column.
Fig. 9 : Raw-IMU Stationary Window Detector performance vs. GT stationary windows for two ANYmal GrandTour sequences, ETH-3 (top) and EIG-1 (bottom).
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Parameter
TLIO
EuRoC
Blackbird
MIPO
GrandTour (quadruped)
Humanoid
(pedestrian)
(drone)
(drone)
(quadruped)
IMU-only
joints
Batch size
512
64
512
256
256
256
512
Supervision horizon [s]
5.12
1.22
2.43
0.41
3.65
3.65
1.22
Learning rate
1e-3
1e-3
1e-3
5e-5
5e-5
5e-5
1e-3
Epochs
40
150
120
240
120
60
128
Regularization weight λreg
0.25
0.1
0.1
0.01
1.0
0.1
1.0
Appendix
TABLE VI : Integration self-supervision training configuration. Supervision horizon is the velocity-increment window length used during integration self-supervision.
Keyframes N
Dense (Cholesky)
Sparse (BSR + cuDSS)
Speedup
128
63.4 ms
35.3 ms
1.8×
512
OOM
167.9 ms
n/a
1024
OOM
438.7 ms
n/a
Appendix
TABLE VII : Pose-graph solve cost at batch size B=128 on a single NVIDIA RTX 5090 ( 32 GB), mean over 20 solves. Both backends return the same solution; OOM marks configurations whose dense memory exceeds the device.
Fig. 12 : Position and velocity on six sequences from the GrandTour Test set.
Fig. 13 : Position and velocity on six sequences from the TLIO Test set.
Fig. 14 : Position and velocity on six sequences from the EuRoC Test set.
The training of learned inertial odometry depends on dense, high-precision position ground truth from motion capture, visual-inertial odometry or SLAM, which is costly and hard to acquire at scale. We propose DINS-IO, which learns inertial odometry directly from raw IMU streams without position labels. Our key insight is that the strapdown INS velocity recursion is a strong, fully differentiable consistency prior: the predicted velocity, rotated into the navigation frame, must agree with the integrated specific force up to an unknown initial velocity and a constant accelerometer bias. We cast this constraint as a sliding-window least-squares problem with a globally shared bias, solve it in closed form, and use the solver residual as a self-supervised loss whose gradient flows back to the network through the analytic solution. To supply this per-sample constraint, we design a high-frequency network that emits dense body-frame velocity at the IMU rate. Since the self-supervised network learns consistent motion but its velocity is not yet metrically calibrated, we calibrate it to true metric velocity from a few labeled trajectories by directly supervising the predicted body-frame velocity and adapting only low-rank (LoRA) patches. On standard benchmarks, DINS-IO pretrained self-supervised and fine-tuned with a small fraction of labels matches or surpasses fully supervised baselines.
Hao Qiao, Yan Wang, Jian Kuang +1
School of Geodesy and Geomatics, Wuhan University · GNSS Research Center, Wuhan University · Hubei Luojia Laboratory
Inertial measurement units (IMUs), which provide high-frequency linear acceleration and angular velocity measurements, serve as fundamental sensing modalities in robotic systems. Recent advances in deep neural networks have led to remarkable progress in inertial odometry. However, the heavy reliance on ground truth data during training fundamentally limits scalability and generalization to unseen and diverse environments. We propose KISS-IMU, a novel self-supervised inertial odometry framework that eliminates ground truth dependency by leveraging simple LiDAR-based ICP registration and pose graph optimization as a supervisory signal. Our approach embodies two key principles: keeping the IMU stable through motion-aware balanced training and keeping the IMU strong through uncertainty-driven adaptive weighting during inference. To evaluate performance across diverse motion patterns and scenarios, we conducted comprehensive experiments on various real-world platforms, including quadruped robots. Importantly, we train only the IMU network in a self-supervised manner, with LiDAR serving solely as a lightweight supervisory signal rather than requiring additional learnable processes. This design enables the framework to ensure robustness without relying on joint multi-modal learning or ground truth supervision. The supplementary materials are available at https://sparolab.github.io/research/kiss_imu.
Jiwon Choi, Hogyun Kim, Geonmo Yang +2
Electrical and Computer Engineering, Inha University, Incheon, South Korea
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
Yuyang Leng, Renyuan Liu, Shaohan Hu +4
Richard · George Mason University Fairfax, VA, USA · Global Technology Applied Research, JPMorgan Chase New York, NY, USA