IMU Odometry Challenge

Announcements

May 30, 2026: The official dataset release hub is set to /imuchallenge/data/.

May 30, 2026: Platform pages (Car, Drone, Quadruped, Handheld) are published.

May 30, 2026: Initial IMU Odometry Challenge website structure is now live under the /imuchallenge/ namespace.

What This Challenge Is

The IMU Odometry Challenge is a CMU AirLab competition and benchmark for neural inertial odometry.

Participants train and evaluate models on shared train/validation splits and are ranked on held-out test sets. The benchmark is motivated by the Tartan IMU direction: large-scale pretraining, efficient adaptation, and robust generalization across platforms.

This challenge is held in conjunction with the IROS’26 Workshop: “Beyond Exteroception: Interoceptive Perception for Resilient Robotics”.

Intro Video: Learning IMU Odometry

All official dataset packages, download links, and the live metadata explorer are published on /imuchallenge/data.

Try the Live Demo

Run our reference TartanIMU specialist models on your own IMU sequences directly in the browser — no setup required. Launch the Hugging Face Space demo →

Challenge Goal

Build IMU odometry models that generalize across robot platforms and beat current state-of-the-art performance under a shared benchmark protocol, surfacing research questions for the workshop discussion on interoceptive robot perception.

Problem Formulation

Input: Raw 6-DOF IMU measurements — accelerometer (a_x, a_y, a_z) in m/s² and gyroscope (ω_x, ω_y, ω_z) in rad/s — sampled at 200 Hz. Each input window spans 1 second (200 samples); models receive sequences of 10 consecutive windows (10 s total).

Output: Per-window 3D body-frame velocity predictions v = (v_x, v_y, v_z) in m/s. Positions are derived by the organizers via integration with ground-truth orientation.

Evaluation Metrics

Submissions are ranked on held-out test trajectories across all platforms using:

Velocity RMSE (primary ranking metric)
Macro-averaged velocity RMSE — mean of per-platform RMSEs — so platform size imbalance cannot be gamed.

ATE — Absolute Trajectory Error (secondary metric)
Position RMSE over 5 m drift-corrected segments, computed by the organizers by integrating predicted velocities with ground-truth orientation.

Benchmark Structure

Train: development data for model fitting and ablation.

Validation: public split for model selection and error analysis.

Test: held-out benchmark split for official ranking.

Start Here

BibTeX

@inproceedings{zhao2025tartanimu,
  title={Tartan IMU: A Light Foundation Model for Inertial Positioning in Robotics},
  author={Zhao, Shibo and Yagnyatinskiy, Maxim and others},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2025},
  url={https://openaccess.thecvf.com/content/CVPR2025/papers/Zhao_Tartan_IMU_A_Light_Foundation_Model_for_Inertial_Positioning_in_CVPR_2025_paper.pdf}
}

@misc{imuchallenge2026,
  title={IMU Odometry Challenge: Cross-Platform Inertial Positioning Benchmark},
  author={CMU AirLab and Super Odometry Group},
  year={2026},
  howpublished={\url{https://superodometry.com/imuchallenge}},
  note={Dataset and benchmark challenge page}
}