lihua_imu_iros2026_v1

TartanIMU IROS2026 — Unified Single-Model Inference

Public, reproducible inference for the Tartan IMU Challenge (IROS 2026).

Task. Given 6-axis IMU windows (acc 3 + gyro 3, length 200) covering a motion window plus its temporal context, predict body-frame velocity (vx, vy, vz) for all four platforms (car / dog / drone / human) with a single unified model.

Model. lihua_imu_iros2026_v1: a shared ResNet1D encoder with an FFT branch, followed by a Transformer over a triplet of consecutive windows (prev / curr / next), and a 4-expert Mixture-of-Experts head. The expert gate is a soft router driven purely by IMU-derived features — experts are learned motion regimes (general / slow / fast / agile), not platforms. A single frozen checkpoint (seed 0) is shipped.

Submitted artifact

• Kaggle submission CSV MD5: 3e664d57cf8e2075d801cf42c2fb7be9
• Kaggle Public LB Score: 0.46609
• Official scoring service: TartanIMU Score 0.37390, AVE 0.29972 m/s, ATE20 1.00837 m, RTE@5s 1.35469 m
• Model: lihua_imu_iros2026_v1, checkpoint weights/seed0.pt, seed 0

Reported validation metrics

Repository layout

.
├── inference.py          # entry point (test submission)
├── model.py              # lihua_imu_iros2026_v1 definition
├── requirements.txt      # fixed runtime
├── weights/
│   ├── seed0.pt          # frozen checkpoint
│   └── manifest.json     # sha256 integrity manifest
└── submissions/
    └── submission.csv    # 30644-row leaderboard submission

Usage

Requirements: Python 3.10+, a CUDA-enabled torch, single GPU <=16 GB VRAM, official competition data at <competition_root>.

pip install -r requirements.txt

# Reproduce the challenge submission (anonymous test set):
python inference.py --data <competition_root> --out submissions/submission.csv

inference.py first verifies the SHA-256 of the checkpoint against weights/manifest.json (freeze-integrity), loads the frozen model, runs the triplet-window forward pass in fp16 autocast, and writes window_id,vx,vy,vz in the exact index order.

Compliance notes

• Unified model: one checkpoint, one architecture, soft IMU-driven expert gate; no per-platform routing and the platform auxiliary head is not used for prediction.
• Inference limits: single GPU, ~3.2M-param model, single pass over the test set (~minutes), no network calls.
• Inputs: only raw 6-axis IMU (the window triplet is temporal context, not a platform label). No orientation/position/velocity ground truth at inference.
• Anonymized test: we never read or route on a platform label; windows are streamed from test/{traj_id}.npz in index order.

Reproduction target

Re-running inference.py --data <competition_root> on the official data reproduces submissions/submission.csv up to fp16 GPU rounding noise (relative diff ~1e-4; same magnitude as run-to-run non-determinism of the reference notebook). The shipped CSV is the authoritative submission.