Predict 3-D body-frame velocity from 6-axis IMU alone - one unified model across car, dog, drone, and human.
Four robots. One inertial sense. Can a single model read them all? A car, a legged dog, a racing drone, and a walking person all carry the same cheap sensor — a 6-axis IMU. From one second of that raw signal alone, and nothing else, predict how fast the body is moving in 3-D. The twist: you must use one unified model for all four. The test set never tells you which platform a window comes from — there are no platform labels and the trajectory ids are anonymized — so a single model has to infer and adapt on its own. One model, four embodiments, one leaderboard. This is a cross-embodiment generalization benchmark, not four separate regressions.
Scored by the TartanIMU Score — 60 % per-window velocity error (AVE) plus 40 % 20 m-segment trajectory error (ATE20), each macro-averaged over the four platforms so no platform dominates and no degenerate strategy pays: a submission of all zeros scores exactly 1.000, and the organizers' released baseline scores 0.637. Top teams are invited to present at the IROS 2026 workshop Beyond Exteroception: Interoceptive Perception for Resilient Robotics.
To be eligible for the final ranking, in addition to submitting predictions, you must complete the Submission Form for each submission you want considered, attaching the prediction CSV you uploaded. You must also submit a technical report (PDF)within 7 days of the deadline, written with the official template, and publish your single unified model's weights and inference codeas a public Hugging Face model repository, within the same 7 days. Teams missing any of the three, or whose model cannot be re-executed, are removed from the final ranking regardless of leaderboard position.
The task is to estimate 3-D body-frame linear velocity (m/s) from raw IMU. The input is raw 6-axis IMU only— 3-axis accelerometer (m/s^2) and 3-axis gyroscope (rad/s), 200 Hz, in the body frame, gravity retained. No ground-truth orientation, position, or velocity is provided as model input— exactly what a robot has at runtime. The data consists of 136,289 windows across 564 trajectories, about 37.9 hours.