Tartan IMU: A Light Foundation Model for Inertial Positioning in Robotics

Shibo Zhao, Sifan Zhou, Raphael Blanchard, Yuheng Qiu, Wenshan Wang, Sebastian A. SchererCVPR 2025 · DOI

Introduction

Inertial Measurement Units (IMUs) serve as the "inner ear" for robotic systems, providing critical data on acceleration and angular velocity. While essential for navigation in GPS-denied or low-visibility environments, translating raw IMU data into precise positioning—known as Inertial Odometry—remains a significant challenge. Traditional methods relying on double integration are prone to rapid error accumulation, leading to massive positional drift.

Recent deep learning approaches have attempted to mitigate these errors, yet they often suffer from a "specialization gap," where a model optimized for one platform (e.g., a walking human) fails when deployed on another (e.g., a high-speed drone). Tartan IMU addresses this by introducing the first foundation model for inertial positioning, designed to generalize across diverse motion types and adapt to new environments in real-time.

Methodology

The Tartan IMU framework is structured into three distinct stages to ensure broad generalization and rapid local adaptation:

  1. Pre-training (The Foundation): The authors developed a Heterogeneous Shared Backbone utilizing ResNet and LSTM layers. This architecture maps diverse IMU signals into a unified high-dimensional space. To accommodate the varying dynamics of different platforms, a multi-head architecture is employed, connecting the shared backbone to specific heads for cars, drones, humans, and quadrupeds.
  2. Efficient Fine-tuning (The Adapter): To transition to unseen platforms without the computational cost of full retraining, the framework utilizes Low-Rank Adaptation (LoRA). By freezing the foundation model and training only a compact adapter (1.1M parameters), the system achieves positive knowledge transfer while maintaining its core capabilities.
  3. Online Test-Time Adaptation: During deployment, the model continues to learn through a "Learning within SLAM" approach. A simultaneous localization and mapping (SLAM) system provides feedback to the IMU model. To maintain high efficiency (200 FPS), Gaussian Mixture Models (GMMs) are used to select a diverse buffer of motion samples, ensuring the model learns from varied dynamics rather than repetitive stationary data.
The three-stage framework of Tartan IMU: Pre-training, Fine-tuning with LoRA, and Online Adaptation.

Data and Cross-Platform Synergy

The model was trained on a massive, heterogeneous dataset comprising over 100 hours of real-world IMU data, including ground vehicles (TartanDrive, SubT-MRS), legged robots, human-held devices (IDOL), and high-speed drones (Blackbird, UZH).

A key insight from this research is that training on a variety of robots enhances the model's performance on any single platform. The "All-robot" foundation model outperformed specialized single-robot models by an average of 42%. This suggests that the model learns fundamental principles of motion that transcend specific mechanical configurations.

Visualization of the learned motion embedding space across different robotic platforms and motion types.

Results and Performance

Tartan IMU demonstrates significant improvements over state-of-the-art specialized models such as AI-IMU and TLIO:

  • Accuracy: The model achieved a 35.5% improvement in Absolute Trajectory Error (ATE) and a 41.0% improvement in Time-Relative Error.
  • Adaptability: When introduced to a completely new vehicle and environment, the model successfully adapted and reduced trajectory error within just 105 seconds.
  • Efficiency: Despite its robust architecture, Tartan IMU maintains an inference speed of 200 FPS, making it suitable for low-latency, real-time robotics applications.

Conclusion and Future Work

Tartan IMU represents a major step toward a universal inertial positioning system. By treating deployment as a continuous learning process, it bridges the domain gap between training environments and real-world conditions. Future research will explore truly robot-agnostic heads, the integration of simulated IMU data to further expand the model's knowledge, and more sophisticated buffer selection methods for handling chaotic motion patterns.

Summary updated

BibTeX
@inproceedings{zhao2025tartan,
  author = {Shibo Zhao and Sifan Zhou and Raphael Blanchard and Yuheng Qiu and Wenshan Wang and Sebastian A. Scherer},
  title = {{Tartan IMU: A Light Foundation Model for Inertial Positioning in Robotics}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2025},
  doi = {10.1109/cvpr52734.2025.02097},
}
Figure