TartanIMU operating across ground vehicle, quadruped, drone, and human motion platforms

CVPR 2025

TartanIMU

A Light Foundation Model for Inertial Positioning in Robotics

One inertial model learns motion across ground vehicles, quadrupeds, drones, and humans, then adapts efficiently to new robots and operating conditions.

Live Inference Code Paper Dataset

Shibo Zhao1†* · Sifan Zhou1†* · Raphael Blanchard1 · Yuheng Qiu1 · Wenshan Wang1 · Sebastian Scherer1

1Carnegie Mellon University   †Equal contribution   *Corresponding authors

Poster BibTeX
Pretrain
Generalize
Adapt
Deploy
Live TartanIMU

Run TartanIMU inference online

Explore IMU sequences and inspect model predictions directly in your browser.

Launch Data Explorer

Put cross-platform inertial models to the test. The IROS 2026 IMU Odometry Challenge provides a platform-blind benchmark across four embodiments.

View the challenge
01 / Results

One foundation model, four motion domains

TartanIMU transfers motion knowledge across distinct embodiments instead of training an isolated model for every platform.

Training data
100+ hrs
Platforms
4
ATE improvement
36%
Online speed
200 FPS
Benchmark overview

Cross-model IMU odometry comparison

Source: HERO
Qualitative pose comparison across specialist baselines and the shared IMU pretrained model. Video provided by the HERO project.
Interactive benchmark

Detailed single-trajectory comparison

Compare platform-specific expert models with the shared IMU pretrained model on the same car, quadruped, human, and drone trajectories.

Source: HERO
Ground vehicle used for the HERO car trajectory comparison
Trajectory 01

Car trajectory

Expert model
Car Expert Model

AI-IMU

Expert model
Dog Expert Model

RNIN-VIO

Expert model
Human Expert Model

TLIO

Expert model
Drone Expert Model

IMO

Shared pretrain
IMU Pretrain Model

TartanIMU (ours)

Science Robotics Movie 5.1

TartanIMU predictions by platform

Ground-truth trajectories and network predictions across four robot platforms.

Ground vehicle ยท Science Robotics Movie 5.100:30
02 / Overview

Generalizable inertial positioning

Most learned IMU odometry systems overfit one dataset or one robot. TartanIMU learns shared motion structure from diverse platforms, then adapts without discarding that general representation.

The system combines large-scale pretraining, parameter-efficient fine-tuning, and online test-time adaptation. This progression supports both zero-shot transfer and rapid specialization when a deployment introduces an unseen motion pattern or domain shift.

01

Cross-robot pretraining

A shared backbone captures motion knowledge across ground vehicles, quadrupeds, drones, and humans.

02

Efficient transfer

LoRA adapts the model to new tasks with 1.1 million trainable parameters while retaining prior knowledge.

03

Continuous adaptation

An adaptive memory buffer supports real-time updates as motion and operating conditions change.

Overview of TartanIMU across multiple robot embodiments and adaptation stages
TartanIMU is an open-source, cross-robot foundation model for state estimation from IMU measurements.
03 / Method

Three stages from pretraining to deployment

The same model moves from broad motion learning to lightweight task adaptation and finally to online updates in the field.

TartanIMU architecture showing pretraining, efficient fine-tuning, and online adaptation
The TartanIMU pipeline: shared pretraining, adapter-based fine-tuning, and online adaptation with an adaptive memory buffer.
1

Pretrained IMU model

A shared backbone learns general motion patterns from more than 100 hours of multi-platform inertial data.

t-SNE visualization of TartanIMU feature clusters across robot platforms
Learned features separate platform dynamics while preserving a shared inertial representation.
2

Efficient fine-tuning

Low-Rank Adaptation freezes the foundation model and trains compact adapters for new tasks, enabling positive transfer with limited data and compute.

TartanIMU offline fine-tuning evaluation Comparison showing retained performance after LoRA adaptation
LoRA improves adaptation performance while reducing catastrophic forgetting.
3

Online adaptation

A motion-aware memory buffer selects diverse recent samples for stable, real-time updates when speed, terrain, or motion patterns shift.

Online TartanIMU adaptation on an unseen trajectory
TartanIMU progressively adapts during deployment using balanced samples from recent motion segments.
04 / Dataset

Diverse motion at a consistent interface

More than 100 hours of synchronized inertial data span four embodiment families and a wide range of indoor and outdoor motion.

Ground vehicle trajectory and inertial result overview

Ground vehicles

Driving, off-road, and subterranean motion.

Quadruped trajectory and inertial result overview

Quadrupeds

Walking, turning, stairs, and rough terrain.

Drone trajectory and inertial result overview

Drones

Three-dimensional flight and rapid maneuvers.

Human motion trajectory and inertial result overview

Human motion

Handheld and body-worn locomotion patterns.

Browse dataset and checkpoints Read setup and training instructions
05 / Limitations

Generalization still has boundaries

TartanIMU generalizes across vehicles, drones, humans, and legged robots, but it does not yet support every possible embodiment. New platform families may require a dedicated motion head or a mixture of existing experts.

06 / Citation

Cite TartanIMU

CVPR 2025, pages 22520–22529

@inproceedings{zhao2025tartan,
  title={Tartan IMU: A Light Foundation Model for Inertial Positioning in Robotics},
  author={Zhao, Shibo and Zhou, Sifan and Blanchard, Raphael and Qiu, Yuheng and Wang, Wenshan and Scherer, Sebastian},
  booktitle={Proceedings of the Computer Vision and Pattern Recognition Conference},
  pages={22520--22529},
  year={2025}
}

Build on the official TartanIMU implementation

Models, training code, and project updates are available in the public repository.

Open TartanIMU Code