Research map / Architectures and efficiency / RNNs and normalization
EqNIO: Subequivariant Neural Inertial Odometry
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 EqNIO, which predicts an O(2)/O_g(3)-subequivariant yaw frame that canonicalizes IMU inputs for neural displacement priors, applied to TLIO and RONIN for better generalization across IMU orientations and motion directions.
Problem
Neural inertial odometry learns displacement priors from a single IMU, but must ignore variations from IMU mount orientation and motion direction, which hampers generalization; existing methods only approximately enforce this via augmentation or auxiliary losses.
Approach
Predict a learnable gravity-aligned yaw frame with a subequivariant network (equivariant under roto-reflections preserving gravity), canonicalize IMU data into that frame, run an off-the-shelf displacement network, and map outputs back; angular rates are decomposed into two acceleration-like vectors so that accelerations and rates transform consistently under reflections.
Key results
- With TLIO, the O(2) equivariant frame reduces MSE* on Aria from 15.248 to 2.304 (10^-2 m^2) vs. vanilla TLIO, and ATE* on TLIO dataset from 3.079 to 2.382 m.
- On RONIN, the O(2) frame gives ATE* of 0.70 m on RIDI cross-subject vs 1.67 m for the 100% data baseline, and 1.28 m on OxIOD vs 3.46 m.
- Frames predicted by a non-equivariant MLP overfit and generalize worse on Aria than yaw-augmented TLIO.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use O(2) Eq. Frame instead of TLIO | TLIO · TLIO Dataset | ATE*: 3.079 -> 2.382 | Helped |
| use O(2) Eq. Frame instead of TLIO | TLIO · Aria Dataset | ATE*: 4.560 -> 1.850 | Helped |
| use O(2) Eq. Frame instead of RONIN with 50% data | RONIN · RONIN-U | ATE*: 5.57 -> 4.42 | Helped |
| use O(2) Eq. Frame instead of RONIN with 50% data | RONIN · RONIN-S | ATE*: 3.54 -> 3.32 | Helped |
| use O(2) Eq. Frame instead of RONIN with 50% data | RONIN · RIDI-T | ATE*: 1.19 -> 0.82 | Helped |
| use O(2) Eq. Frame instead of SO(2) Eq. Frame | RONIN · RIDI-C | ATE*: 0.63 -> 0.70 | Hurt |
| use O(2) Eq. Frame instead of RONIN + 50% data + TTT | RONIN · OxIOD | ATE*: 2.92 -> 1.28 | Helped |
| add yaw rotation augmentation (ablation) | TLIO · Aria Dataset | MSE*: 15.248 -> 5.322 | Helped |
| increase residual block depth from 2 to 3 (ablation) | TLIO · Aria Dataset | ATE*: 1.864 -> 2.541 | Hurt |
| predict the frame with a non-equivariant MLP (ablation) | TLIO + yaw augmentation · Aria Dataset | ATE*: 1.864 -> 2.071 | Hurt |
| use PCA-generated frames (ablation) | TLIO · Aria Dataset | ATE*: 1.864 -> 4.635 | Hurt |
| replace frame-based method with fully equivariant CNN (ablation) | TLIO · Aria Dataset | ATE*: 1.850 -> 6.916 | Hurt |
Surprising findings
- A non-equivariant learned frame performs worse than yaw augmentation alone, indicating that equivariance is needed for the frame network.
- A deeper TLIO with comparable parameter count does not match the equivariant methods.
Limitations
- Frame network adds compute (638.5 vs 35.5 MFLOPs for TLIO) and parameters.
- Fully equivariant CNN underperforms the frame-based approach, suggesting full equivariant redesign is restrictive.
- Gravity estimation errors are a source of perturbation requiring augmentation for robustness.
Experimental setup
Datasets: TLIO Dataset RONIN Dataset
Benchmarks: TLIO Dataset Aria Everyday Activities RONIN RIDI OxIOD
Metrics: MSE ATE RTE AYE
Baselines: TLIO TLIO + rot. aug. RONIN RIO NDI
Builds on: TLIO RONIN Vector Neurons RIO
Closest papers by meaning
- Tartan IMU: A Light Foundation Model for Inertial Positioning in Robotics (CVPR 2025 · similarity 0.87)
- iMoT: Inertial Motion Transformer for Inertial Navigation (AAAI 2025 · similarity 0.87)
- Adaptive VIO: Deep Visual-Inertial Odometry with Online Continual Learning (CVPR 2024 · similarity 0.86)
- From Variance to Veracity: Unbundling and Mitigating Gradient Variance in Differentiable Bundle Adjustment Layers (CVPR 2024 · similarity 0.85)
- Leveraging Consistent Spatio-Temporal Correspondence for Robust Visual Odometry (AAAI 2025 · similarity 0.84)
- MoE-Gyro: Self-Supervised Over-Range Reconstruction and Denoising for MEMS Gyroscopes (NeurIPS 2025 · similarity 0.83)
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.