No ground truth needed. Uses LiDAR-based ICP and pose graph optimization as lightweight supervisory signals for training.
GMM-based motion analysis identifies and rebalances underrepresented motion patterns, ensuring robust learning across all dynamics.
Uncertainty-aware adaptive weighting during inference dynamically adjusts sensor confidence for reliable pose estimation.
Inertial measurement units (IMUs), which provide high-frequency linear acceleration and angular velocity measurements, serve as fundamental sensing modalities in robotic systems. Recent advances in deep neural networks have led to remarkable progress in inertial odometry. However, the heavy reliance on ground truth data during training fundamentally limits scalability and generalization to unseen and diverse environments.
We propose KISS-IMU, a novel self-supervised inertial odometry framework that eliminates ground truth dependency by leveraging simple LiDAR-based ICP registration and pose graph optimization as a supervisory signal. Our approach embodies two key principles: keeping the IMU stable through motion-aware balanced training and keeping the IMU strong through uncertainty-driven adaptive weighting during inference.
To evaluate performance across diverse motion patterns and scenarios, we conducted comprehensive experiments on various real-world platforms, including quadruped robots. Importantly, we train only the IMU network in a self-supervised manner, with LiDAR serving solely as a lightweight supervisory signal rather than requiring additional learnable processes.
Overall Framework. Our inertial odometry framework follows the Keep IMU Stable and Strong philosophy through three components: (a) Self-supervised training combines an IMU network for correction and uncertainty prediction with a LiDAR registration module for geometric constraints. PGO fuses both modalities, followed by selective pseudo-label generation using symmetric overlap scores. (b) GMM analysis stabilizes IO via motion clustering with balanced reweighting. (c) Sensor confidence-aware PGO strengthens IO through adaptive weighting during inference.
In real-world deployment, the motion distribution of training sequences and test sequences can differ significantly. For instance, a model trained primarily on straight-line trajectories will struggle with sharp turns or sudden accelerations at test time.
As shown above, the distributions of normalized angular speed vary significantly across sequences, and lower Wasserstein distances indicate better motion similarity between training and test data. This distributional gap is a critical bottleneck for generalizable inertial odometry. To address this, KISS-IMU introduces a GMM-based motion analysis pipeline that identifies and rebalances under-represented motion components during training, ensuring stable learning across the full spectrum of motion dynamics.
We model the motion distribution using a Gaussian Mixture Model with Bayesian information criterion to determine the optimal number of components. Each training sample is assigned to its corresponding motion cluster, and under-represented components (e.g., sharp turns, sudden acceleration) receive higher sampling weights. This motion-aware reweighting prevents the model from being biased toward dominant motion patterns, enabling stable and generalizable inertial odometry learning.
Effect of Motion-Balanced Training. The t-SNE visualization demonstrates the impact of our GMM-based reweighting on the learned feature space. Without balancing, motion clusters overlap significantly, making it difficult for the network to learn motion-specific corrections. With GMM-balanced training, distinct motion patterns (left turn, right turn, straight) become clearly separated with tight intra-class compactness — even in unseen environments — confirming that balanced training leads to more discriminative and generalizable representations.
At inference time, Strong IO strengthens trajectory estimation through an uncertainty-driven adaptive pipeline. The trained IMU network outputs corrected measurements along with per-sample uncertainty estimates. These uncertainties are then used to construct an adaptive information matrix within pose graph optimization (PGO), where confident predictions receive higher weights while unreliable ones — caused by out-of-distribution motions or noisy readings — are automatically suppressed.
This sensor confidence-aware weighting allows KISS-IMU to self-assess its reliability at each time step, producing robust and accurate trajectories even in unseen environments without any additional fine-tuning or ground truth supervision.
We evaluate on three diverse datasets with increasing complexity:
Complexity: 5/5
Unitree B2 quadruped in crater-filled planetary terrain. Extreme motion variations; ground truth infeasible.
Quantitative comparison on Botanic Garden and DiTer++ datasets with varying training data percentages (100%, 60%, 20%). Our approach consistently achieves stable performance across different data scales and generalizes better to unseen environments. Notably, our method maintains robust performance with only 20% training data.
Qualitative comparison on Botanic Garden and DiTer++ datasets. Our method produces trajectories that closely align with the ground truth, while competing approaches exhibit noticeable drift in challenging segments.
Trajectory estimation across varying training ratios on DiTer++ dataset. Even with reduced training data, KISS-IMU maintains consistent trajectory quality, demonstrating robust generalization under data-scarce conditions.
@inproceedings{choi2026kissimu,
author = {Choi, Jiwon and Kim, Hogyun and Yang, Geonmo and Lee, Juhui and Cho, Younggun},
title = {KISS-IMU: Self-supervised Inertial Odometry with Motion-balanced Learning and Uncertainty-aware Inference},
booktitle = {IEEE International Conference on Robotics and Automation (ICRA)},
year = {2026},
}