acceptodds
Under review as a conference paper at ICLR 2027

InertialBridge: Efficient Structure-Aware Denoising of IMU Trajectories via Variational Autoencoder Conditional Bridge Matching

Abstract

Dense inertial motion estimation requires accurate per-sample predictions at a manageable computational cost. We present InertialBridge (IB), a conditional bridge-matching (BM) framework that predicts 100 samples of 3D velocity v and angular rate ω after observing a complete one-second IMU window, using a single network evaluation. Conditioning the drift on the source window makes every interior bridge target algebraically identifiable. A boundary-weighted scheduler πρ = ρ δ0 + (1 − ρ) U [0, 1] allocates training mass to the deployed query at t = 0. Our population analysis establishes one-step exactness of the optimal field and characterizes the boundary drift that determines its prediction. A matched ten-epoch, single-seed sweep yields 2.7–3.9 times higher position ATE under uniform scheduling than under the evaluated positive-boundary-mass schedules. IB parameterizes the drift with a Mamba3 state-space network conditioned on a frozen variational autoencoder (VAE), and uses a structured kinematic loss. Under our common RoNIN protocol, IB achieves the lowest mean position ATE and RTE among the evaluated methods on both subject splits; paired ATE tests do not detect a significant difference from EqNIO. IB uses 1.95 M parameters and 0.87 GFLOPs per window, compared with 5.03 M and 0.84 GFLOPs for EqNIO, while returning an estimate at every input sample. These results support efficient dense reconstruction of observed IMU windows.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
IMU, flow matching, bridge matching, inertial odometry, navigation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_XppYOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes InertialBridge, which using complete one-second IMU widow information to predict velocity and angular-rate sequences. And the method uses boundary-weighted scheduler and a Mamba3 state-space network. The evaluation on different datasets demonstrates its competitive position accuracy with fewer parameters.

Critical strengths

This paper proposes a novelty method, which using a bridge-matching framework and Mamba3 network to reconstruct dense velocity and angular-rate reconstruction. In addition, it provides a good theoretical analysis and sufficient experiments.

Critical weaknesses
  1. The experiment is not sufficient, which show bridge-interior training over direct regression. Because, the direct-regression base in table 2 disables source stacking, which weaken the persuasion for isolating the effect of bridge-interior training.
  2. In Table 1, although IB has lower parameter count and memory usage comparing with EqNIO. However, its FLOPs and Lat.(EqNIO: 3.88, IB: 5.92) are higher. The author can analyze and explain this phenomenon.
Questions for authors
  1. Before identifying the specific methodological gap, the authors should state the problem being addressed more explicitly in the abstract.
  2. The introduction and method section should give a more clearer explanation of this paper's motivation.
  3. The first paragraph of the methodology section should provide a clearer overview of IB to help readers understand the subsequent technical details.
  4. The authors could consider moving selected experiments and the associated discussion from the appendix into the main text. Typos and Minor Issues:
  5. Please define FLOPs as “floating-point operations” at its first occurrence.
  6. Table 3 lists six OxIOD(T) test sequences, whereas Appendix A.1.8 refers to two. Please reconcile these counts and explain any filtering.
1 comment

Thank you for the detailed response and additional experiments. I appreciate the authors' efforts to address my concerns.

However, I will maintain my original rating. The new experiments show that bridge-interior training does not significantly improve trajectory accuracy compared with direct regression. The gains are mainly limited to small improvements in dense velocity estimation. Regarding efficiency, IB still has higher FLOPs and latency per window than EqNIO. While the authors have clarified the advantage of dense outputs, this benefit is limited to specific offline reconstruction settings.

Reviewer_a1pLOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

InertialBridge efficiently predicts many plausible motion trajectories from one IMU window. Existing methods force a trade-off: window-level models are accurate but provide only one motion estimate per second, while dense models output motion at all 100 time steps but are computationally expensive.

The key insight is to train more heavily on the difficult final bridge query. InertialBridge generates 100 IMU samples in one pass using differentiable bridge matching and a boundary-weighted scheduler. A continuous uniform distribution never samples exactly at (t=0), creating a mismatch between training and deployment. The authors therefore add a (\delta_0) boundary-weighted scheduler to train at the boundary.

The artificial bridge coordinate works roughly as follows:

  • (t=0)(t=0): the source boundary, where deployment begins.
  • (0<t<1)(0<t<1): intermediate bridge states used during training.
  • (t=1)(t=1): the target side containing the desired kinematic sequence. At deployment, only the drift at (t=0)(t=0) is queried. A single Euler step then predicts all 100 velocity and angular-rate time steps simultaneously.

Subsection 3.1 defines the linear bridge, source-conditioned drift, boundary-weighted scheduler, one-step inference, and structured loss. Subsection 3.2 shows that the ideal interior drift follows straight paths, while boundary prediction is the Bayes estimator. One Euler step recovers this estimator for the selected ideal field. Subsection 3.3 implements the drift using a bidirectional Mamba3 U-Net, time-step conditioning, and a frozen VAE to obtain a representation of the source window.

The experiments support competitive one-pass, network-level performance and provide quantitative evidence comparing the method with EqNIO.

Critical strengths
  • The motivation and trade-off are clear and well written.
  • The related-work section is coherent, the discussion is well written, and the conclusion does not overclaim.
  • The method efficiently produces dense predictions in one pass.
  • The abstract provides quantitative evidence comparing the method with EqNIO, and the experiments support competitive one-pass, network-level performance.
Critical weaknesses
  • For every (t>0)(t>0), ((xt,x0,t))((x_t,x_0,t)) determines (x1)(x_1). The bridge therefore acts mainly as a target-revealing auxiliary reconstruction task. Whether this reconstruction objective actually helps prediction at (t=0)(t=0) is an empirical regularization question, not a consequence of the propositions. A matched direct-regression control with the same source stacking and training budget could help clarify this.
  • It is also unclear how the interior and boundary losses do not interfere at finite capacity when using neural networks.
  • The evaluation does not assess the practical value of the learned prior in a complete inertial-odometry pipeline. It seems that integration with an EKF is not straightforward with the current learned-prior architecture. The current metrics do not test whether the per-sample error is lower with overlapping windows, as would occur when integrating with an EKF. The current evaluation uses non-overlapping windows, as in prior work, and analyzes only the learned-prior component.
Questions for authors

The paper is already very well written and is meaningful but the below pointers can further strengthen the contribution. Please correct me if any question seems unreasonable indicating my lack of understanding of this contribution. Good work!

  • Could the authors clarify the practical significance of predicting all 100 samples and the utility of the reported 2.7−3.9×2.7-3.9\times improvement in the abstract for improved accessibility?
  • Could the authors provide a simple illustration or example of pairs crossing and explain how InertialBridge eliminates this problem? This would improve accessibility and strengthen the intuition.
  • What are the roles of the frozen VAE and Mamba3? The VAE appears to contain no additional information, and the cross-attention appears to be a projection. Could the authors provide an ablation removing the VAE to clarify its importance?
  • Could the authors explain how the interior and boundary losses do not interfere at finite capacity when using neural networks?
  • Summed angular-rate error is only a heuristic proxy for orientation error because 3D rotations are noncommutative. Could the authors clarify the effect of this choice?
  • Jointly optimizing e−σL+σe^{-\sigma}L+\sigma becomes unbounded if an empirical component approaches zero unless σ\sigma is constrained or regularized. Is σ\sigma constrained or regularized?
  • What would the benefits look like with high-quality IMU data, such as the Aria Dataset used in EqNIO and recent papers in this field?
  • Does InertialBridge still improve trajectory accuracy when all learned priors are placed in the same TLIO-style EKF? My prior is that the EKF is very good at handling minor errors, resulting in less benefit from a better learned motion prior. The EKF is an important component in the practical deployment of inertial-estimation systems.
  • How would the covariance of the learned prior be estimated and calibrated? Would it be overestimated, as in traditional TLIO, for which the predicted uncertainty is scaled?
  • How are correlations among the 100 outputs, and their correlations with the raw IMU propagation, handled? The current evaluation uses non-overlapping windows.
  • Does the benefit shrink when the EKF already corrects small prediction errors?
  • Could the authors provide a multi-seed version of Table 5?
  • Could the authors provide visualizations of predictions for stationary trajectories and for the same path traveled at different speeds? The input window would differ at different speeds.
Reviewer_EGEAOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

InertialBridge estimates dense velocity and angular-rate sequences from complete one-second IMU windows using a single network evaluation. It combines a source-conditioned deterministic bridge-matching objective, explicit sampling mass at the inference boundary, a Mamba3 backbone, frozen-VAE conditioning, and a structured kinematic loss. The theoretical analysis separates algebraically identifiable interior targets from the boundary prediction and establishes one-step exactness at the population optimum. Evaluation includes RoNIN seen/unseen subjects and a pooled multi-benchmark setting. The method achieves competitive trajectory accuracy with fewer parameters than EqNIO, although their paired ATE differences are not statistically significant. Its outputs are dense within an observed window rather than causal estimates available at every incoming sample.

Critical strengths

The framework produces a prediction at each input sample using one network evaluation. On RoNIN, its trajectory accuracy is competitive with the strongest evaluated window-level baseline, while using substantially fewer parameters. This is a potentially useful trade-off for offline reconstruction or applications that tolerate window-level delay.

The paper explains why source conditioning identifies interior bridge targets and why the deployed boundary requires direct supervision. The matched boundary-rate sweep provides evidence for the importance of boundary supervision. The manuscript also explicitly reports the non-significant ATE comparisons with EqNIO, single-run training, and non-causal window processing. The supplementary implementation and scripts provide useful material for checking the proposed method.

Critical weaknesses

The motivation and positioning of the contribution are not sufficiently convincing. The paper presents window-level output as a limitation without clearly establishing when this output format is inadequate. Although offline reconstruction and delayed smoothing are identified as potential applications, the connection between these applications, the need for dense predictions, and the reported evaluation remains underdeveloped. This makes it difficult to assess the practical significance of the proposed method. In the absence of a clear methodological advance, I do not see substantial practical value in the proposed approach.

Questions for authors

IMU noise and bias instability are important practical challenges that degrade motion estimation and make reliable tracking difficult. However, the abstract opens by emphasizing accurate per-sample predictions at a manageable computational cost, and I find the connection between this objective and the practical challenges insufficiently clear. Could the authors clarify the primary problem they aim to address: estimation accuracy, computational efficiency, or temporal output resolution? In particular, why are dense per-sample predictions necessary for the intended applications, and which experimental results demonstrate their benefit over window-level estimates?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.