Research map / Vision-language and multimodal / Cross-modal learning
iMoT: Inertial Motion Transformer for Inertial Navigation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes iMoT, a Transformer encoder-decoder for inertial odometry that adds progressive series decoupling, adaptive positional encoding, and learnable query motion particles scored by a dynamic scoring mechanism to estimate velocity segments and reconstruct trajectories.
Problem
Inertial odometry must estimate velocity segments from IMU sequences while handling motion uncertainty and the different characteristics of acceleration and angular-rate modalities, so that integrated trajectories drift less.
Approach
Treats each variate's full temporal series as a token, decomposes seasonal and trend-cycle components inside each encoder layer (PSD), uses modality-adaptive positional scaling (APE), and models motion uncertainty with learnable query motion particles whose final set is pooled by a learned dynamic score (DSM) instead of entropy-regularized averaging.
Key results
- On RoNIN unseen subjects, iMoT reaches ATE 5.31 m versus 6.77 m for TLIO and 6.89 m for CTIN.
- On RoNIN, the full model reduces ATE from 6.48 m (baseline) to 5.31 m in the ablation study.
- On IDOL unseen, T-RTE improves to 2.85 m from 3.37 m for RoResNet18 (best reported T-RTE for iMoT).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use iMoT instead of RoResnet18 | inertial odometry · RoNIN Unseen | ATE (m): 6.89 m vs. 5.31 m | Helped |
| Add PSD (ablation) | iMoT without modules · RoNIN Unseen | ATE (m): 6.48 -> 5.58 | Helped |
| Add APE (ablation) | iMoT without modules · RoNIN Unseen | ATE (m): 6.48 -> 5.76 | Helped |
| Add ASC (ablation) | iMoT without modules · RoNIN Unseen | ATE (m): 6.48 -> 6.00 | Helped |
| Add query particles with Jvel and Jent, without DSM (ablation) | iMoT without modules · RoNIN Unseen | ATE (m): 6.48 -> 7.10 | Hurt |
| Add PSD and particles without DSM (ablation) | iMoT · RoNIN Unseen | ATE (m): 5.58 -> 7.03 | Hurt |
| Add DSM to the particles-only configuration (ablation) | iMoT with particles · RoNIN Unseen | ATE (m): 7.10 -> 6.13 | Helped |
| Add DSM to PSD, APE, and ASC (ablation) | iMoT · RoNIN Unseen | ATE (m): 5.54 -> 5.31 | Helped |
| Add query particles and DSM to PSD, ASC, and APE (ablation) | iMoT · RoNIN Unseen | ATE (m): 5.54 -> 5.31 | Helped |
| Increase particle count beyond 128 with Jvel/Jent and average pooling (ablation) | iMoT decoder · RoNIN | ATE | Hurt |
| Use 128 particles with DSM (ablation) | iMoT decoder · RoNIN | ATE | Helped |
| Use iMoT instead of RoResnet18 | inertial odometry · IDOL Seen | ATE (m): 2.70 -> 2.22 | Helped |
Surprising findings
- Directly adding query motion particles with the entropy-based loss worsens ATE (e.g., 7.10 m vs 6.48 m baseline), and only the dynamic scoring mechanism makes the particles beneficial.
- RoResNet18 achieves slightly better T-RTE and comparable D-RTE than iMoT on some RIDI and RoNIN splits, despite iMoT's higher ATE advantage.
Limitations
- Larger parameter count (about 14.5M) than several baselines such as RoResNet18 and RoTCN.
- Gains are smaller or absent on OxIOD and on T-RTE for some splits, where per-sequence attachments reduce motion variability and the query particles are underused.
Experimental setup
Datasets: RIDI RoNIN OxIOD IDOL
Benchmarks: RIDI RoNIN OxIOD IDOL
Metrics: ATE T-RTE D-RTE PDE
Baselines: SINS PDR RIDI RoLSTM RoTCN RoResNet18 CTIN TLIO
Builds on: DETR Autoformer-style series decomposition CTIN RoNIN TLIO
Closest papers by meaning
- EqNIO: Subequivariant Neural Inertial Odometry (ICLR 2025 · similarity 0.87)
- Accurate and Steady Inertial Pose Estimation through Sequence Structure Learning and Modulation (NeurIPS 2024 · similarity 0.85)
- Tartan IMU: A Light Foundation Model for Inertial Positioning in Robotics (CVPR 2025 · similarity 0.85)
- MODA: Motion-Drift Augmentation for Inertial Human Motion Analysis (CVPR 2025 · similarity 0.84)
- Dynamic Inertial Poser (DynaIP): Part-Based Motion Dynamics Learning for Enhanced Human Pose Estimation with Sparse Inertial Sensors (CVPR 2024 · similarity 0.82)
- Adaptive VIO: Deep Visual-Inertial Odometry with Online Continual Learning (CVPR 2024 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.