tartanimu-iros2026

TartanIMU Challenge (IROS 2026) submission model

Single unified model (one set of weights) for car / dog / drone / human. Input: 1 s windows of 200 Hz 6-axis IMU (body frame, gravity retained); output: mean body-frame velocity per window.

• run: ad_dgate1_trainval (last.pt) — SHA-256 19ebaaee391ad91d76ca01847cb20aa439b8120ebe404ead382e475acde6bd44
• architecture: 1-D ResNet window encoder + 2-layer bi-GRU over 32-window chunks + context-gated linear (physics) head (lin)
• training config: config.json (all CLI arguments of train.py); trained on train,val for 8000 steps
• inference: python predict.py --ckpt model.pt --split test --chunk 64 --overlap 16 --out submission.csv (deterministic chunked processing; no TTA, no ensembling, no weight averaging)
• Kaggle submission 56320418: public 0.31780, full test (organisers' scorer) 0.23790; submission CSV md5 344d58a8571d29b05b3f5f8a48456c2f
• training/inference code: https://github.com/BharathSShankar/tartanimu-iros2026 at commit 83a79a858a1d91762f87b3b5782adc8fd625af64 (private repository; access for the organisers on request)
• the technical report (report.pdf) and the full experiment log (EXPERIMENTS.md) are included in this repo