Accurate inertial odometry is critical for autonomous systems but suffers from unbounded drift caused by the integration of sensor noise and biases. In approaches such as TLIO, neural pseudo-measurements are trained independently of their downstream filter. This paper introduces Differentiable Filtering for Inertial Odometry (DFIO), a learning framework that embeds the update stages of an error-state Kalman filter directly into a computational graph. Following standalone pre-training, DFIO jointly optimizes the displacement and covariance output heads using posterior-position supervision through differentiable measurement updates. The proposed system is validated on public pedestrian motion datasets featuring both calibrated and uncalibrated inertial measurements. Experimental results on a pedestrian dataset demonstrate that DFIO reduces mean absolute and one-second relative translation errors by 13.2% and 39.6% relative to the TLIO baseline, and by 7.7% and 17.4% relative to its pre-trained counterpart. Furthermore, evaluations on uncalibrated smartphone inertial data indicate that the proposed method effectively mitigates translational drift and maintains trajectory alignment under highly noisy sensor conditions. Source code will be made publicly available.
Figures & tables
Figure 1: Overview of DFIO. Raw IMU measurements drive analytical ESKF propagation and are transformed into a local gravity-aligned frame G for neural measurement model IncNet. The specific force aG and angular velocity ωG are processed by IncNet to predict a displacement increment δpG and its covariance R for the ESKF measurement update. The pairs (x^i∣i−1,P^i∣i−1) and (xi∣i,Pi∣i) denote the prior and posterior state estimates and error-state covariances, respectively, for measurement windows i . After standalone pretraining, posterior-position supervision through the differentiable update refines the displacement and covariance heads of IncNet. Blue dotted arrows indicate backpropagation during training
Method type
Model
ATE[m]
RTE-1s[m]
TDR[%]
AYE[deg]
RYE-1s[deg]
YDR[deg/h]
Open-loop
RoNIN-IncNet
2.657 / 2.197
0.244 / 0.218
2.852 / 2.985
12.903 / 10.983
1.723 / 1.333
295.233 / 218.982
TLIO
1.957 / 1.610
0.140 / 0.136
1.863 / 1.401
2.353 / 1.806
0.893 / 0.494
197.434 / 73.172
Independent learning
Phase 1 Only
1.840 / 1.856
0.102 / 0.102
1.760 / 2.019
2.120 / 1.920
0.231 / 0.237
46.757 / 45.102
Tune Cov
1.897 / 1.758
0.085 / 0.086
1.479 / 1.124
2.358 / 1.865
0.230 / 0.235
52.922 / 52.309
Tune Inc
1.795 / 1.631
0.085 / 0.085
1.673 / 1.540
1.998 / 1.484
0.229 / 0.231
46.214 / 37.220
Filter-aware learning
DFIO
1.699 / 1.645
0.084 / 0.082
1.656 / 1.700
2.007 / 1.370
0.231 / 0.233
43.598 / 37.015
Table 1: Results on the 15 held-out TLIO trajectories. Each metric cell reports mean / median values. The lowest mean and median across all six methods are bold.
Baseline
Metric
Improved / worsened
Mean Δ [m]
95% CI [m]
TLIO
ATE
6/9
−0.2576
[−0.9231,0.2861]
TLIO
RTE-1s
15/0
−0.0554
[−0.0718,−0.0395]
Phase 1 Only
ATE
8/7
−0.1409
[−0.3510,0.0528]
Phase 1 Only
RTE-1s
14/1
−0.0177
[−0.0225,−0.0125]
Table 2: Paired translation-error comparisons on the 15 TLIO test trajectories. Δ denotes DFIO minus the baseline; negative values favor DFIO. The 95% confidence intervals are percentile bootstrap intervals for the mean paired difference, obtained from 100,000 resamples of complete test trajectories
Figure 2: Cumulative distributions of error metrics on the TLIO test set for DFIO, TLIO, and RoNIN-IncNet baseline
Figure 3: Selected planar trajectories from the TLIO test set, comparing reference trajectories, DFIO, TLIO, and RoNIN-IncNet.
Figure 4: Position components and yaw estimation on the held-out OxIOD sequence.
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 navigation using low-cost MEMS sensors is plagued by rapid drift due to sensor noise and bias instability. While recent data-driven approaches have made significant strides, they often struggle with micro-drifts during stationarity and mode fusion during complex motion transitions due to their reliance on fixed-window regression. In this work, we introduce Gated Neural Inertial Odometry (GNIO), a novel learning-based framework that explicitly models motion validity and context. We propose two key architectural innovations: \ding{182} a learnable Motion Bank that queries a global dictionary of motion patterns to provide semantic context beyond the local receptive field, and \ding{183} a Gated Prediction Head that decomposes displacement into magnitude and direction. This gating mechanism acts as a soft, differentiable Zero-Velocity Update (ZUPT), dynamically suppressing sensor noise during stationary periods while scaling predictions during dynamic motion. Extensive experiments across four public benchmarks demonstrate that GNIO significantly reduces position drift compared to state-of-the-art CNN and Transformer-based baselines. Notably, GNIO achieves a 60.21% reduction in trajectory error on the OxIOD dataset and exhibits superior generalization in challenging scenarios involving frequent stops and irregular motion speeds.
Dapeng Feng, Yizhen Yin, Zhiqiang Chen +2
Sun Yat-sen University, Guangzhou, China. · The University of Hong Kong, Hong Kong SAR, China.
Neural inertial odometry increasingly uses networks as learned measurements inside filtering pipelines. Such measurements should transform consistently under arbitrary IMU mounting conventions: their mean must transform as a vector, and their covariance must transform congruently as a second-order tensor. We present GINIO, a geometric SO(3)-equivariant interface for neural inertial odometry under arbitrary rotations of the IMU measurement frame. Given calibrated IMU windows, our framework predicts a motion measurement and uncertainty obeying these tensorial laws. To support efficient sensor-frame learning, we introduce Last-Frame Alignment (LFA), a deterministic preprocessing step that is provably equivalent to world-frame training for SO(3)-equivariant predictors. The connected estimator tracks sensor-local states such as IMU bias, separating nuisance estimation from the geometric law enforced by the learned measurement. We instantiate the same interface in filter-connected NIO, AirIO-style recurrent aerial prediction, EqNIO-style full-SO(3) canonicalization, and ResNet-style temporal backbones. On TLIO, GINIO achieves 2.018 m ID/SO(3) ATE while EqNIO degrades to 76.389 m, using 11.6x fewer FLOPs. On NanoBench, our AirIO-style instantiation improves ATE from 5.579 m to 1.430 m without external attitude input, and our ResNet-style instantiation reaches 0.581 m ATE versus 0.645 m for ResNet1D. On Fetch, GINIO empirically reduces unseen physical-remount ATE from 8.15 m to 0.50 m without retraining, demonstrating robustness beyond the exact coordinate-frame guarantee. For uncertainty, spectral covariance reduces covariance-equivariance error by over three orders of magnitude compared with a diagonal head.
Chankyo Kim, Minghan Zhu, Tzu-Yuan Lin +2
University of Michigan Ann Arbor, MI, United States