IROS 2026 IMU Odometry
Official Hugging Face repository ID:
xiaojx/2026-tartaimu-challenge.
This repository contains our entry for the IROS 2026 Learning IMU Odometry Challenge. The system first learns temporal IMU representations through masked reconstruction, then finetunes one shared residual mixture-of-experts model to predict mean body-frame velocity for each one-second window. The same model is used for car, human, quadruped, and drone data; platform labels are not required at inference, and the final submission uses neither ensembling nor test-time adaptation.
The selected model uses 10-sample patches, 40% masking, and geometry/platform auxiliary supervision during velocity finetuning. Its ten-second bidirectional context means that it is not a strictly causal estimator.
Results
Kaggle team: JX0005.
| Evaluation | Score | ATE20 (m) | AVE (m/s) | RTE@5s (m) |
|---|---|---|---|---|
| Public leaderboard (50%) | 0.32898 | — | — | — |
| Official scoring service (89 sequences) | 0.26486 | 0.68557 | 0.21683 | 0.97086 |
Masked reconstruction reduced the public score from 0.53189 for the matching random-initialization baseline to 0.33271, a 37.4% improvement. Adding the finetuning auxiliary objectives produced the selected public score of 0.32898. These are single-seed experiments, and aerial trajectories remain the largest source of error.
The final submitted prediction file is
checkpoints/submissions/submission_primary.csv,
with MD5 ae8ff897f40fe407c765551fd4dd485f. The official result tables are included in the
technical report.
Environment setup
The workspace supports Python 3.10 and 3.11. Dependencies are locked with
uv.lock; the resolved GPU build is PyTorch 2.7.1 with CUDA 12.8.
Clone the official Hugging Face repository and create the environment from its root:
git clone https://huggingface.co/xiaojx/2026-tartaimu-challenge
cd 2026-tartaimu-challenge
uv sync --frozen
The Hugging Face repository includes the upstream TartanIMU source. Challenge data are distributed separately and are not included in the repository.
Documentation
research/README.md: method, research results, data layout, training, validation, ablations, and prediction from research runs.checkpoints/README.md: generate a submission with the released checkpoint and identify the exact final submission artifact.TartanIMU/README.md: upstream challenge library and scorer documentation.report/IMU_Challenge_Report_Template/main.tex: technical report with the full ablation and per-sequence results.