CKANIO: Learnable Chebyshev Polynomials for Inertial Odometry

Basic Information

• Paper ID: 2507.16865
• Title: CKANIO: Learnable Chebyshev Polynomials for Inertial Odometry
• Authors: Shanshan Zhang, Siyue Wang, Tianshui Wen, Liqin Wu, Qi Zhang, Ziheng Zhou, Ao Peng, Xuemin Hong, Lingxiang Zheng, Yu Yang
• Category: cs.RO (Robotics)
• Publication Date: October 16, 2025 (arXiv v2)
• Paper Link: https://arxiv.org/abs/2507.16865

Abstract

Inertial odometry (IO) relies solely on inertial measurement unit (IMU) signals for localization, offering a promising approach for consumer-grade positioning. However, accurately modeling the nonlinear motion patterns inherent in IMU signals remains a primary limitation on IO accuracy. To address this challenge, the authors propose CKANIO, an IO framework that integrates Chebyshev-based Kolmogorov-Arnold Networks (Chebyshev KAN). Specifically, a novel residual architecture is designed that leverages the nonlinear approximation capabilities of Chebyshev polynomials within the KAN framework to more effectively model the complex motion characteristics inherent in IMU signals. To the authors' knowledge, this is the first work applying interpretable KAN models to IO. Experimental results on five public datasets demonstrate the effectiveness of CKANIO.

Research Background and Motivation

Problem Definition

Inertial odometry (IO) is a technique that uses only IMU sensors for position and orientation estimation, with important applications in autonomous driving, robot navigation, and other fields. When GPS signals are unavailable or external sensors are limited, IO can serve as a supplementary or alternative positioning solution.

Core Challenges

1. Error Accumulation Problem: Traditional Newtonian mechanics-based IO methods suffer from error accumulation over time due to inherent IMU measurement noise, severely affecting long-distance positioning accuracy
2. Difficulty in Nonlinear Motion Modeling: Complex nonlinear motion patterns in IMU signals (such as rotations) are difficult to model accurately
3. Limitations of Fixed Activation Functions: Existing data-driven methods rely on fixed activation functions and cannot sufficiently capture the complex dynamics of IMU signals

Research Motivation

Inspired by the excellent performance of Kolmogorov-Arnold Networks (KAN) in nonlinear function approximation, the authors propose using Chebyshev polynomial-based KAN to replace traditional fixed activation functions, thereby better modeling the complex nonlinear features in IMU signals.

Core Contributions

1. First Introduction of Chebyshev KAN to IO Domain: Designs a residual network architecture that leverages the nonlinear approximation capabilities of Chebyshev polynomials to enhance the model's ability to characterize nonlinear dynamics
2. Development of Efficient Kernel-based Self-Attention Module (EKSA): Enhances contextual motion modeling for more comprehensive utilization of IMU signals, while reducing computational complexity from O(L²) to O(L)
3. Provides First Quantitative Analysis of Gravity's Impact on IO Accuracy: Quantitatively analyzes the specific impact of gravity on IO performance in the global coordinate system

Methodology Details

Task Definition

Given IMU signals X ∈ ℝ^(C×L) within a time window (where C=6 represents three-axis acceleration and three-axis angular velocity, and L is the number of sampling points), predict velocity vector v̂ while minimizing mean squared error with true velocity v.

Model Architecture

1. Residual Chebyshev KAN (ResCKAN)

Input Preprocessing:

X' = arccos(tanh(X)) ∈ (0,π)^(C×L)

The tanh function constrains inputs to the range (-1,1), and arccos is then applied to obtain a numerically stable phase representation.

Chebyshev Polynomial Definition:

T_n(x) = cos(n·arccos(x))  (x ∈ [-1,1])

Learnable Activation Function:

Y = Σ(i=0 to n) W_i × T_i(tanh(X))

where W_i are learnable parameter tensors and n=4 is the polynomial order. This design enables the activation function to dynamically adapt to IMU signal characteristics.

2. Efficient Kernel-based Self-Attention (EKSA)

Core Concept: Uses kernel functions to approximate the attention matrix, avoiding explicit computation of the full attention matrix.

Pearson Correlation Coefficient Similarity:

ρ(q,k)² = (q̄ᵀk̄ / (||q̄|| ||k̄||))²

where q̄ = q - μ(q) and k̄ = k - μ(k). This metric is robust to offset and gain noise commonly found in IMU signals.

Exponential Kernel Function:

K_F(q,k) = exp(ρ(q,k)²)

Taylor Expansion Approximation:

K_F(q,k) = Σ(n=0 to ∞) (q̃²ⁿ/√n!) · (k̃²ⁿ/√n!)ᵀ

EKSA Computation:

EKSA(Q,K,V) = f_Q(f_K^T V)

Reduces complexity from O(C_R L_R²) to O(L_R C_R²).

Technical Innovations

1. Adaptive Nonlinear Modeling: Learnable combinations of Chebyshev polynomials dynamically adapt to different IMU signal characteristics
2. Computational Efficiency Optimization: The EKSA module achieves linear time complexity for global dependency modeling
3. Noise Robustness: The invariance of Pearson correlation coefficients to offset and gain noise improves system robustness

Experimental Setup

Datasets

Five public inertial datasets are used:

• IMUNet, RoNIN, RIDI, RNIN, TLIO
• Datasets are randomly split 8:1:1 into training, validation, and test sets
• A specially constructed TLIO (w/o gravity) dataset is used for gravity impact analysis

Evaluation Metrics

1. Position Drift Error (PDE): End-point drift error
2. Relative Trajectory Error (RTE): Local trajectory accuracy
3. Absolute Trajectory Error (ATE): Global positioning accuracy

Comparison Methods

• RoNIN (ResNet, TCN, LSTM variants)
• IMUNet
• TLIO network
• RNIN network

Implementation Details

• GPU: NVIDIA RTX A40 (48 GB)
• Learning rate: 10⁻⁴
• Batch size: 512
• Training epochs: 100
• CUDA version: 11.3
• PyTorch version: 1.11

Experimental Results

Main Results

Cross-dataset performance comparisons on five datasets show that CKANIO consistently outperforms comparison methods on most datasets:

RoNIN Dataset Performance:

• ATE: 5.365m → 3.814m (compared to RoNIN ResNet)
• RTE: 3.390m → 3.272m

Overall Performance: CKANIO achieves the best or near-best ATE and RTE values across all test datasets.

Gravity Impact Analysis

Experiments reveal that removing gravitational acceleration in the global coordinate system significantly reduces errors for nearly all algorithms:

• Gravity acts as a constant bias in the globally aligned coordinate system
• Removing the gravity component allows accelerometer measurements to better reflect the vehicle's translational motion
• Improves learning effectiveness and numerical stability

Ablation Studies

• CKANIO (w/o EKSA): Using ResCKAN alone already outperforms RoNIN ResNet
• Complete CKANIO: Adding EKSA further reduces ATE and RTE, improving velocity prediction accuracy

Visualization Analysis

Trajectory comparisons show:

1. In simple scenarios, RoNIN ResNet gradually deviates due to accumulated heading errors, while CKANIO demonstrates stronger noise robustness
2. In complex scenarios, RoNIN ResNet exhibits significant drift after nonlinear maneuvers such as rotations, while CKANIO better captures these dynamic characteristics

Related Work

Traditional IO Methods

Primarily rely on Newtonian mechanics for position and orientation estimation, but suffer from error accumulation due to IMU noise.

Data-Driven IO Methods

Learn motion patterns from large-scale IMU data, outperforming traditional methods in accuracy and applicability, but still have limitations in modeling complex nonlinear motion.

KAN Networks

Kolmogorov-Arnold Networks replace traditional fixed activation functions with learnable ones, demonstrating excellent nonlinear approximation capabilities across various domains.

Conclusions and Discussion

Main Conclusions

1. CKANIO successfully introduces Chebyshev KAN to the IO domain, significantly improving nonlinear motion modeling capabilities
2. The EKSA module effectively balances performance and computational efficiency
3. Gravity removal has a significant positive impact on IO performance

Limitations

1. Validation is limited to pedestrian IMU signals, lacking evaluation on other vehicle types
2. Computational complexity, while improved, remains higher than traditional methods
3. Handling capability for extreme motion patterns requires further verification

Future Directions

The authors propose extending evaluation to non-pedestrian IMU signals as a promising research direction.

In-Depth Evaluation

Strengths

1. Strong Technical Innovation: First introduction of KAN to the IO domain with pioneering significance
2. Solid Theoretical Foundation: Nonlinear approximation theory of Chebyshev polynomials provides a solid basis for the method
3. Comprehensive Experimental Design: Comparative experiments on five datasets with detailed ablation studies
4. High Practical Value: Significantly improves accuracy while maintaining computational efficiency

Weaknesses

1. Limited Applicability: Validation limited to pedestrian scenarios; generalization for industrial applications remains to be verified
2. Insufficient Theoretical Analysis: Lacks theoretical explanation for why Chebyshev polynomials are particularly suitable for IMU signals
3. Inadequate Computational Cost Analysis: While complexity is reduced, actual runtime comparisons are lacking

Impact

1. Academic Impact: Introduces a new network architecture paradigm to the IO domain
2. Practical Value: Provides a new technical pathway for consumer-grade positioning devices
3. Reproducibility: Detailed experimental setup facilitates subsequent research

Applicable Scenarios

1. Indoor positioning for consumer mobile devices
2. Navigation systems in GPS-denied environments
3. Robot systems requiring high-precision inertial navigation

References

The paper cites 26 relevant references covering important works in inertial navigation, deep learning, KAN networks, and other domains, providing a solid theoretical foundation and comparison benchmarks for this research.