Differentiable Filtering for Inertial Odometry

arXiv:2611.13517 · cs.RO, cs.CL · Submitted 2026-11-10 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: We're looking at the paper: "Differentiable Filtering for Inertial Odometry".

Paper summary: Rosa: So we talked about the core idea of "Differentiable Filtering for Inertial Odometry", which is using a neural network, IncNet, to predict displacement increments and measurement covariances inside an error state Kalman filter framework. The authors claim this approach embeds the update stages of that filter directly into a computational graph.

Dev: Essentially, they start with raw IMU measurements driving the analytical error state Kalman filter propagation and then feed those measurements into IncNet to predict the three dimensional displacement increment and its associated covariance within that local gravity aligned frame.

Taro: The key is that after initial standalone pre-training, they use posterior position supervision through differentiable measurement updates to refine both the displacement prediction head and the covariance prediction head of IncNet simultaneously.

Rosa: They are optimizing these heads jointly using this feedback loop based on where the filter actually ends up being, which is a step beyond just training the network to predict displacement in isolation.

Dev: This entire process is validated on public pedestrian motion datasets that include both calibrated and uncalibrated inertial measurements, showing tangible performance gains over baselines like TLIO.

Taro: The results they reported are that DFIO reduces mean absolute and one second relative translation errors by thirteen point two percent and thirty nine point six percent relative to the TLIO baseline on a pedestrian dataset.

Rosa: And they also tested its effectiveness on uncalibrated smartphone inertial data, where it successfully mitigates translational drift while maintaining trajectory alignment under highly noisy sensor conditions.

Dev: The method is structured around this idea of having the analytical IMU propagation model of the error state Kalman filter remain deterministic, while the learnable parts are that neural measurement predictions and their uncertainties.

Taro: This structure lets them keep the known kinematics of how motion propagates while only learning how to best predict those noisy increments and uncertainties.

Rosa: It’s about combining a strong analytical foundation with a data-driven correction mechanism that learns how to handle the noise and bias better during the filtering process.

Dev: The methodology involves this two stage training strategy, moving from minimizing mean squared error against ground truth displacement in phase one to using marginal negative log likelihood in phase two for refinement.

Taro: That coupling between the supervised learning of displacement and covariance heads is what lets them achieve the lowest mean absolute translation error, which was one point six hundred ninety-nine meters on their test set.

Rosa: So, to sum up this paper, DFIO introduces a framework where a temporal network predicts motion and uncertainty that feeds back into a differentiable filter update process during training.

Conclusion: Rosa: Looking at "Differentiable Filtering for Inertial Odometry" by Yaru Gu and Ting Zou, the core idea is embedding the error state Kalman filter's update stages directly into a computational graph using a neural network called IncNet.

Dev: The authors claim this framework jointly optimizes the displacement and covariance heads using posterior position supervision through differentiable measurement updates after an initial standalone pre-training phase.

Taro: They demonstrate that this approach reduces mean absolute and one second relative translation errors by thirteen point two percent and thirty nine point six percent when tested on pedestrian motion datasets.

Rosa: More importantly, they show that it handles noisy, uncalibrated inertial measurements well, maintaining alignment even when the initial gyroscope and accelerometer biases are set to zero because there was no prior calibration data.

Dev: The implication is that for systems needing accurate tracking of movement in real time, you can leverage a learning component to actively correct the filter state based on how well the prediction matches reality.

Taro: It suggests that this learned inertial measurement model provides a way to extract information about motion itself from data when traditional methods rely heavily on unconstrained integration.

Rosa: And they are making the source code public, which is always beneficial for others who want to explore this specific approach to inertial odometry.

Dev: So, in simple terms, DFIO is a learning framework that improves inertial odometry by using a neural network to predict motion and uncertainty that directly influences the filter update process during training.

Yaru Gu, Ting Zou

Department of Mechanical and Mechatronics Engineering, Memorial University of Newfoundland

cs.RO, cs.CL

Submitted: 2026-11-10

Updated: 2026-11-10

Comments: 19 pages, 4 figures

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 74/100

The gist: Accurate inertial odometry is critical for autonomous systems but suffers from unbounded drift caused by the integration of sensor noise and biases The proposed system is validated on public…

Key concepts

Error-State Kalman Filter (ESKF)
The ESKF is a mathematical tool used to estimate the state of a system, like an IMU's position and orientation. It works by propagating the system's state using known physics (IMU kinematics) while simultaneously correcting for errors caused by sensor noise and biases. This allows it to track movement more accurately over time.
IncNet
IncNet is a temporal neural network designed to predict two things: how much the system is displaced in 3D space and the uncertainty (covariance) of that prediction. It takes raw IMU data and predicts these values in a gravity-aligned frame, helping to model complex measurement errors that traditional filters might miss.
Posterior-Position Supervision
This training technique involves using the estimated position from the filter as a target for training the neural network. The network is refined by minimizing a loss function based on this corrected position, ensuring that the neural predictions align closely with the actual movement derived from the filter's state estimation.
Gravity-Aligned Frame (G)
This is a specific reference frame used to simplify how the neural network processes IMU measurements. By transforming raw sensor data into this frame, IncNet can learn more general and robust patterns for predicting displacement and uncertainty, making the model less sensitive to the specific orientation of the device.

Terminology

Summary

Accurate inertial odometry is critical for autonomous systems but suffers from unbounded drift caused by the integration of sensor noise and biases 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. p. 02% and 39.6% relative to the TLIO baseline 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

Framework Architecture

DFIO combines an error-state Kalman filter (ESKF) with IncNet, a temporal network that predicts three-dimensional displacement increments and diagonal measurement covariances in a local gravity-aligned frame The raw IMU measurements drive analytical ESKF propagation and are transformed into a local gravity-aligned frame G for neural measurement model IncNet The network is first trained using displacement supervision Its displacement and covariance heads are then refined together using posterior-position supervision through differentiable filter updates, while the pretrained temporal encoder is held fixed The propagation model retains the deterministic IMU kinematics; the learnable quantities are the neural measurement predictions and their uncertainties

Training Strategy

The training process is structured into a two-stage curriculum Phase 1 involves standalone measurement pre-training where the network is first trained for 10 epochs minimizing the Mean Squared Error (MSE) against the ground-truth window displacement The loss transitions to the marginal NLL for the remaining epochs Phase 2 involves filter-coupled refinement where gradients from posterior-position supervision pass through the error-state correction and covariance update The coupled system minimizes LPhase2 = Lpos-NLL + λNEESLNEES

Performance Gains

The complete DFIO model achieved a mean ATE of 1.699 m and a mean RTE-1s of 0. p. 0084 m on the TLIO test set These correspond to reductions of 13. p. 62% and 39.6% relative to the TLIO baseline The refinement primarily improved translation accuracy while retaining the heading accuracy of the pretrained system Joint refinement in DFIO produces the lowest mean ATE, 1. p. 0699 m, compared with 1.897 m for Tune Cov and 1. p. 9795 m for Tune Inc

Sensor Robustness

The supplementary OxIOD case study provided a supplementary illustration on a smartphone walking sequence to show DFIO’s effectiveness in noisy and uncalibrated IMU data Due to the lack of prior calibration data in this dataset, the initial gyroscope and accelerometer biases for the ESKF were initialized to zero DFIO effectively mitigates translational drift, maintaining tight alignment with the ground-truth on OxIOD sequences

Conclusion

The present evaluation concerns pedestrian motion Future work will investigate a broader range of robotic motions and sensor conditions, together with refinement over longer filter sequences The DFIO source code will be made publicly available The TLIO and OxIOD datasets analysed in this study are publicly available at https://github.com/CathIAS/TLIO and https://deepio. p. 0cs.ox.ac.uk.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Joint refinement of displacement and covariance heads in Phase 2 allows for the two output heads to be refined according to their effect on the posterior position and covariance using posterior-position supervision through differentiable filter updates. This enables a tighter coupling between the neural measurement model and the error-state Kalman filter, leading to lower mean Absolute Translation Error (ATE) of 1.699 m compared to baselines.

  2. Implementation of two-stage training allows for learning a competent baseline displacement estimator first (Phase 1: Standalone Measurement Pre-training) before applying filter-output supervision in Phase 2. This structured approach avoids instability by isolating the initial learning task from the complex joint optimization required for filter refinement.

  3. The system can effectively mitigate translational drift and maintain trajectory alignment under highly noisy sensor conditions, as demonstrated on uncalibrated smartphone inertial data in the OxIOD subset evaluation. This capability is crucial for robust deployment on low-cost or non-calibrated mobile platforms where traditional calibration methods are insufficient.

  4. The model can reduce mean Absolute Translation Error by 13.2% and mean Relative Translation Error by 39.6% relative to the TLIO baseline when compared against a model trained independently of the filter (independently trained IncNet–ESKF). This demonstrates that filter-aware refinement significantly improves short-interval translation accuracy while retaining heading performance.

  5. The system can adapt measurement weighting to improve local translation accuracy and final-position drift without refining the displacement head, as shown by Tune Cov configuration achieving the lowest mean and median Translation Drift Rate (TDR) of 1.479%.

Sources

Related papers