uradkr/tartanimu-urad-unified

URAD — TartanIMU Challenge (IROS 2026) single unified model

One network, one set of weights, all four platforms (car, quadruped, drone, handheld). Raw 200 Hz IMU in, one body-frame velocity per one-second window out.

Kaggle team / submission	URAD / 56405420 (F6_soup5_s8_x102.csv, 2026-09-20 20:20:17 UTC)
Kaggle public score	0.35850
Organizers' scoring service, full test	0.27980 (macro AVE 0.23189 m/s, macro ATE20 0.70628 m) — submission md5 6a599b20e9d83a39c54784821ff41c6a
Checkpoint	model.pt, 4,785,872 parameters, SHA-256 da34721e56601a790d00915d922970b1e58b5eff06c35cc1875b87f4ce1e2d9d
Reference predictions	submission.csv, SHA-256 94ad1f2600fe1ff4f78b65d107a360a1ecade0d69c2db152dc72728b3c863898

Run

pip install -r requirements.txt            # or requirements-cpu.txt
python inference.py --data <competition data root> --checkpoint model.pt \
       --reference submission.csv --out out --device cuda

<competition data root> must contain index/test_windows.csv, sample_submission.csv and test/*.npz. No network access is needed. The script writes out/submission.csv and a JSON audit with the maximum absolute difference to the reference (our CPU re-run: 3.97e-6 m/s, see runtime_cpu.json).

What the model is

A small residual 1-D CNN, the organizers' pretrained TartanIMU ResNet-1D trunk used as a second feature branch, and window / gyro-rotation / whole-trajectory motion descriptors feed a two-layer bidirectional GRU over 64 windows; a learned softmax gate mixes four jointly trained velocity heads inside the same network.

Disclosures

• Checkpoint averaging: model.pt is the parameter-wise mean of five fine-tuning runs from one common checkpoint.
• Inference-time processing: overlapping 64-window contexts of this one checkpoint are averaged at stride 8, and every output vector is multiplied by 1.02. Both settings are stored in the checkpoint config (infer_stride, output_scale) and applied by inference.py. No input-transforming TTA, smoothing, filtering or clipping.
• Platform handling: internal learned gate only. No platform label, trajectory identity, pose or ground truth enters the forward pass.
• Training data: the challenge train and val splits. No external data. Test IMU was never used for fitting.
• Pretrained weights: the organizers' Tartan-IMU/TartanIMU checkpoints/unified.pt (revision 556a9e89c85d35ad83b7d05c852194165a6601cd), fine-tuned and embedded in model.pt; see EXTERNAL_PRETRAINING.json.
• Inference is offline and non-causal (whole-trajectory descriptors), and each trajectory is processed on its own.

Licence and attribution

Model and code definitions derive from TartanIMU (copyright 2026 Shibo Zhao, Apache-2.0; licence text in TartanIMU-LICENSE). The embedded weights derive from the organizers' released checkpoint, whose model card restricts use to research / non-commercial purposes; this repository carries the same restriction. See NOTICE.