IONext: Next Era of Inertial Odometry

• IONext is an emerging research program that leverages learning-based, geometry-aware methods to overcome traditional inertial odometry limitations from noisy IMU data.
• It employs windowed motion estimation, on-manifold preintegration, and dynamic feature extraction to reduce drift and improve performance on benchmark datasets.
• Advanced uncertainty-aware fusion and cross-modal integration techniques adapt inertial localization to various platforms such as smartphones, UAVs, and wearables.

Next Era of Inertial Odometry (IONext) can be understood as the emerging research program that recasts inertial odometry from open-loop double integration of noisy IMU streams into a family of learning-based, geometry-aware, uncertainty-aware, and modular estimators. Across recent work, the defining elements are windowed motion estimation, on-manifold preintegration, explicit orientation handling, symmetry-aware canonicalization, adaptive spectral or dynamic-convolutional feature extraction, and probabilistic fusion with EKF, ESKF, or factor-graph back ends (Chen et al., 2018, Khorrambakht et al., 2020, Sun et al., 2021, Jayanth et al., 2024, Cui et al., 28 Feb 2026).

1. Origins and problem setting

IONext emerged from the recognition that classical strapdown mechanization is fundamentally brittle when driven by low-cost MEMS IMUs. In the canonical formulation, gyroscope and accelerometer outputs are integrated to propagate orientation, velocity, and position; in practice, bias instability, scale-factor errors, misalignment, and noise produce unbounded drift. A central quantitative observation is that a $1^\circ$ attitude error induces approximately $0.1712\ \mathrm{m/s^2}$ of spurious horizontal acceleration, yielding approximately $1.7\ \mathrm{m/s}$ velocity error and approximately $8.56\ \mathrm{m}$ position error within $10$ seconds (Chen et al., 2018). This error growth is especially severe when device placement is unconstrained, zero-velocity updates are unavailable, and environmental aiding is absent.

The first major conceptual break was to stop treating inertial odometry as continuous integration and instead formulate it as windowed sequence estimation. IONet segments the IMU stream into fixed-length windows $W_t$ and learns a mapping

$f_\theta(W_t)\rightarrow(\Delta l_t,\Delta\psi_t),$

where $\Delta l_t$ is horizontal displacement and $\Delta\psi_t$ is heading change over the window. With window length $L=200$ frames, stride $0.1712\ \mathrm{m/s^2}$0 frames, and a 2-layer Bi-LSTM with $0.1712\ \mathrm{m/s^2}$1 hidden units per layer, IONet replaced open-loop state propagation with pseudo-independent local predictions stitched into a trajectory, thereby slowing drift accumulation and avoiding step detection or explicit bias estimation (Chen et al., 2018).

This reformulation was made practically important by the appearance of large, attachment-diverse inertial datasets. OxIOD provided 158 sequences totaling 42.587 km and 14.72 h of data, with IMU sampled at 100 Hz, magnetometer also recorded, four attachments, four consumer phones, five users, and both Vicon-labeled room-scale motion and large-scale office-floor trajectories (Chen et al., 2018). That dataset established a benchmark regime in which inertial odometry had to survive handheld, pocket, handbag, and trolley motion rather than narrowly constrained foot-mounted or laboratory settings.

2. From raw streams to geometry-aware motion summaries

A second defining move in IONext was to compress raw IMU streams into representations that preserve motion geometry. “IMU Preintegrated Features for Efficient Deep Inertial Odometry” introduced a 9-dimensional input

$0.1712\ \mathrm{m/s^2}$2

where $0.1712\ \mathrm{m/s^2}$3, $0.1712\ \mathrm{m/s^2}$4, and $0.1712\ \mathrm{m/s^2}$5 are the manifold preintegrated rotation, velocity, and position increments over an interval $0.1712\ \mathrm{m/s^2}$6 (Khorrambakht et al., 2020). This representation reduced sequence length by roughly an order of magnitude while retaining the Lie-group structure used in visual-inertial back ends. On KITTI, PI features improved test-average relative translation error from 11.53% with raw IMU to 6.35%, and relative rotation error from 2.66 deg/100 m to 1.05 deg/100 m; on OxfordIO, relative translation error improved from 4.35% to 4.14% (Khorrambakht et al., 2020). The same work demonstrated embedded feasibility on an STM32F407ZET microcontroller, with preintegration taking 0.2 ms per feature and CNN inference 5 ms.

Accurate preintegration then pushed this idea further by replacing first-order discrete approximations with piecewise-constant-in-body-frame integration. “Deep Inertial Odometry with Accurate IMU Preintegration” explicitly described the resulting pipeline as a fusion of model-driven and data-driven approaches and reported that accurate preintegration improved dynamic odometry on KITTI from 16.96% relative translation error for raw IO-Net inputs to 8.30%, compared with 9.18% for standard Forster-style preintegration (Khorrambakht et al., 2021). On OxfordIO, where motion is milder, the gap between accurate and standard preintegration was small but still measurable: 3.66% versus 3.71%, both better than 4.35% for raw IMU (Khorrambakht et al., 2021). In IONext, geometry is therefore not discarded by learning; it is increasingly used to compress redundancy, regularize the representation, and improve conditioning before the learned stage.

3. Orientation, frames, symmetries, and eventized representations

A central theme in IONext is that representation choice is itself an observability decision. IDOL made this explicit for smartphone pedestrian localization by identifying inaccurate phone orientation as the dominant downstream failure mode. The paper reports that conventional smartphone orientation estimates can exceed $0.1712\ \mathrm{m/s^2}$7 of error on an iPhone 8 and proposes a two-stage pipeline in which an LSTM-based OrientNet predicts absolute orientation and covariance, an EKF fuses those predictions with gyroscope dynamics on the quaternion manifold, and a second recurrent network then estimates position from world-frame IMU signals (Sun et al., 2021). Orientation RMSE dropped from 0.39 rad for CoreMotion to 0.08 rad on Building 1, and replacing API orientation with IDOL’s orientation improved RoNIN-LSTM on Building 1 from 18.41/6.78/0.52 to 7.03/2.71/0.35 in ATE/T-RTE/D-RTE, nearly matching the 6.53/2.33/0.28 obtained with true orientation (Sun et al., 2021). In this line of work, orientation is not merely a preprocessing nuisance; it sets the performance ceiling of the odometry stack.

EqNIO generalized that insight by embedding the physically correct roto-reflective symmetries of gravity-aligned IMU data directly into the network. It identifies the relevant subgroup as rotations around gravity and reflections with respect to planes parallel to gravity, with equivariance enforced through

$0.1712\ \mathrm{m/s^2}$8

The pipeline estimates an equivariant gravity-aligned frame, canonicalizes IMU sequences into that frame, applies an off-the-shelf odometry backbone, and maps predicted displacement and covariance back into the original frame so that both outputs transform equivariantly (Jayanth et al., 2024). The result is not merely data augmentation by yaw rotation; it is a symmetry-constrained architecture that improves TLIO- and RONIN-style models on TLIO, Aria, RIDI, and OxIOD (Jayanth et al., 2024).

AirIO reached a complementary conclusion for quadrotors: transforming raw IMU into a global frame can reduce rather than improve observability. Its claim is that preserving body-frame IMU with gravity retained exposes attitude-dependent structure more directly for UAV dynamics. The reported effect is an average 66.7% increase in accuracy across three datasets from body-frame representation alone, followed by an additional 23.8% improvement when explicit attitude encoding is added (Qiu et al., 26 Jan 2025). This directly challenges the common practice, inherited from pedestrian pipelines, of early world-frame conversion.

The representational agenda extends even to sampling itself. “Neural Inertial Odometry from Lie Events” replaces fixed-rate raw IMU windows with asynchronous Lie events triggered when the norm of the SE(3) preintegration change exceeds a threshold in the Lie algebra, each event carrying a normalized Lie polarity. The stated motivation is robustness to input-rate changes and trajectory-profile variation, and the paper reports up to 21% reduction in downstream trajectory error with minimal preprocessing (Jayanth et al., 14 May 2025). IONext therefore includes not only new encoders, but new notions of what the inertial signal should be.

4. Architectural evolution beyond raw recurrent models

The architectural history of IONext does not follow a simple progression from RNNs to Transformers. An early and influential edge-oriented branch replaced recurrent sequence models with causal convolutions. L-IONet used an 8-layer WaveNet-style architecture with dilated causal convolutions, gated activations, residual and skip connections, and direct regression of $0.1712\ \mathrm{m/s^2}$9 from 2 s IMU windows (Chen et al., 2020). On OxIOD, WaveNet (32 filters) achieved MSE = 0.0069, had about one quarter of the parameters of a 1-layer Bi-LSTM-128 baseline, trained about 10 times faster, and ran at 3.70 ms per forward pass on a Huawei Mate 8 and 56.78 ms on a Sony Smartwatch 2 (Chen et al., 2020). This established that edge-ready inertial odometry need not sacrifice learned temporal structure.

Later models broadened the representational toolkit. FTIN argued that time-domain CNNs struggle to capture long-term dependencies and introduced a Frequency–Time Integration Network that combines a ResNet-1D backbone, frequency-domain learning with DFT-based complex MLP processing, and a Scalar LSTM for time-domain stability (Zhang et al., 22 Jul 2025). On RoNIN, FTIN reduced ATE from 6.799 to 3.875 and RTE from 3.523 to 3.061 relative to RoNIN ResNet, corresponding to a 43.0% reduction in ATE and a 13.1% reduction in RTE (Zhang et al., 22 Jul 2025). The same study reported consistent gains across RIDI, RNIN, OxIOD, TLIO, and IMUNet, while also showing through ablations that the frequency and time branches each contribute independently.

At the same time, several 2025 works challenged the assumption that Transformer-style global modeling requires attention. “IONext: Unlocking the Next Era of Inertial Odometry” proposed a purely convolutional 1D hierarchy built from the Dual-wing Adaptive Dynamic Mixer and the Spatio-Temporal Gating Unit, reporting normalized RNIN improvements from 2.22/1.41/13.64 to 1.21/0.75/11.35 in ĀTE/R̄TE/ĀLE relative to iMOT (Zhang et al., 23 Jul 2025). DWSFormer followed a parallel route with the Star Operation, collaborative channel–temporal attention, and Multi-Scale Gated Convolution Units, achieving state-of-the-art results across six datasets with 2.76M parameters and 25.12M FLOPs and reporting ATE reductions ranging from 2.26% to 65.78% relative to RoNIN-ResNet (Zhang et al., 22 Jul 2025). Architecturally, IONext is therefore plural: recurrent, frequency–time, dynamic-CNN, and symmetry-aware pipelines all remain competitive when their inductive biases align with the motion regime.

5. Uncertainty-aware fusion and cross-modal expansion

A hallmark of IONext is that learned motion is increasingly treated as a probabilistic measurement rather than a self-sufficient trajectory output. In AI-IO, the network predicts body-frame velocity and a diagonal covariance, and the EKF uses that covariance directly as measurement noise in the update. The generic pattern is

$1.7\ \mathrm{m/s}$0

with $1.7\ \mathrm{m/s}$1 supplied by the learned model rather than fixed a priori (Cui et al., 28 Feb 2026). This formulation allows learned odometry to update full state estimates, including cross-correlated velocity, position, orientation, and biases, while retaining the consistency machinery of classical filters.

Quadrotor work has made the observability argument especially explicit. AI-IO adds rotor speed as the “minimal yet informative” sensing signal needed to restore observability of body-frame velocity under rotor-induced aerodynamics, then combines temporal CNN denoising, a lightweight transformer, and an uncertainty-aware EKF. The paper reports a 36.9% improvement in velocity prediction accuracy from adding rotor speed, a further 22.4% gain from replacing a TCN with the transformer, and 8.9 ms average network inference time on a Radxa Zero3W with EKF at 20 Hz (Cui et al., 28 Feb 2026). This is not an auxiliary convenience variable; it is the missing physical quantity in the aerodynamic measurement model.

Human-motion tracking on wearables shows a different but analogous expansion. MARIO introduces a learned IMU-inferred pose prior grounded in lower-body kinematics and then fuses magnetometer, barometer, and secondary IMUs already present on AR glasses. The paper reports that the pose prior alone reduces positional drift by up to 36% on Nymeria and that multimodal fusion raises the reduction to up to 42%; AirIO +All reaches 133.6M FLOPs and 315 FPS on an A40 GPU (Li et al., 2 Jun 2026). Here the learned prior is not a generic displacement regressor but a kinematic regularizer derived from human pose.

Cross-platform generalization has likewise become an explicit design objective. X-IONet routes 1 s IMU windows through a learned human-versus-quadruped classifier and platform-specific dual-stage attention experts, predicts both displacement and full covariance, and fuses the result through an EKF. The reported reductions are 14.3% in ATE and 11.4% in RTE on pedestrian data, and 52.8% and 41.3% on quadruped data (Shen et al., 11 Nov 2025). The implication is direct: a single inertial prior does not transfer cleanly across platforms with distinct gait and dynamics.

Hybridization with exteroceptive systems follows the same logic. Adaptive-LIO combines adaptive segmentation, LO↔LIO switching under IMU saturation or faults, and adaptive multi-resolution voxel maps; on Robot Town A/B/C it reports end-to-end errors of 1.21 m, 1.36 m, and 1.38 m, outperforming its own no-multi-resolution variant and most baselines (Zhao et al., 7 Mar 2025). ALIVE-LIO goes further by predicting body-frame velocity and fusing it into an ESKF only along degenerate directions when LiDAR observability collapses, yielding the most competitive results in 22 out of 32 sequences (Kim et al., 3 Apr 2026). In VIO/SLAM, inertial guidance has also been used at the correspondence level: inertial-guided uncertainty estimation for feature matches improved average 10-second RPE on CVG-ZJU from 45.56/1.275 to 42.11/1.225 mm/deg, then propagated that covariance into the visual residual weighting (Yoon et al., 2023). IONext thus increasingly supplies selective corrections, uncertainty models, and observability-aware priors inside larger multimodal estimators rather than existing only as a standalone IMU-only frontend.

6. Benchmarks, contested assumptions, and future directions

The empirical base of IONext is now broad enough to expose recurring patterns rather than isolated demonstrations. Across recent benchmark-oriented studies, public datasets include RoNIN at 42.7 h, RIDI at 11.6 h, RNIN at 27.0 h, TLIO at 60.0 h, OxIOD at 14.7 h, and IMUNet at 9.0 h (Zhang et al., 22 Jul 2025). OxIOD remains especially important because its 158 sequences, four attachments, four phones, five users, and office-scale trajectories force methods to confront device heterogeneity and everyday handling variability rather than curated laboratory motion (Chen et al., 2018).

Several recurrent assumptions are directly challenged by this literature. First, orientation is not a solved preprocessing detail: IDOL shows that phone-API orientation can be the dominant downstream failure mode and that replacing it alone can move RoNIN- and TLIO-style systems close to their ground-truth-orientation ceiling (Sun et al., 2021). Second, stronger global modeling does not imply that Transformers are categorically superior: both the CNN-based IONext backbone and DWSFormer report state-of-the-art or near-state-of-the-art results through dynamic convolutional designs that retain strong local inductive biases (Zhang et al., 23 Jul 2025, Zhang et al., 22 Jul 2025). Third, the idea of a universally transferable inertial prior is repeatedly contradicted by platform-specific observability results: body-frame modeling matters for UAVs, rotor speed matters for quadrotors, kinematic pose priors matter for head-mounted human tracking, and expert routing matters for quadrupeds (Qiu et al., 26 Jan 2025).

Future directions recur with notable consistency. IONet’s synthesis emphasizes physics-informed self-/semi-supervised learning, uncertainty-aware odometry, domain adaptation across devices and attachments, multi-sensor fusion, adaptive windowing, and standardized benchmarks (Chen et al., 2018). IDOL identifies broader cross-building generalization, data augmentation, and joint optimization of orientation and localization as open problems (Sun et al., 2021). Lie-event sampling proposes extending eventized Lie-group sensing beyond IMUs to other modalities (Jayanth et al., 14 May 2025). MARIO points toward tighter integration of human kinematics, reliability-aware multimodal fusion, and adaptation to individual biomechanics (Li et al., 2 Jun 2026). Taken together, these directions suggest that IONext is less a single architecture than a design doctrine: preserve the physics that sharpen observability, learn the priors that are hard to model analytically, and fuse them with uncertainty in a way that respects geometry, platform specificity, and operational constraints.