IF-D: A High-Frequency, General-Purpose Inertial Foundation Dataset for Self-Supervised Learning

We present IF-D, a large-scale inertial dataset designed to enable self-supervised and foundational learning for IMU time series. IF-D comprises continuous, long-duration multichannel recordings (accelerometer, gyroscope, magnetometer) sampled at 200Hz using a UM7 IMU mounted inside a 3D-printed spherical enclosure that promotes diverse, free rotations during vehicle traversal. The collection covers approximately 135min of trajectory, yielding on the order of millions of samples across nine sensor channels. We describe the data acquisition setup, preprocessing and calibration procedures (six-orientation accelerometer calibration, stationary gyroscope bias estimation, and ellipsoid fitting for magnetometer hard-/soft-iron correction), and provide quantitative calibration results. IF-D is intended to reduce platform-specific motion bias and to expose models to both physical dynamics and typical measurement noise, facilitating robust representation learning and downstream tasks such as event detection, motion-mode recognition and inertial navigation. The calibration sequence is published in its entirety and also divided into the accel_faces, gyro_static and mag_rotation subsequences, to facilitate the execution of the methodologies.