YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
TartanIMU Challenge — HL Robotics Final Submission (v14)
Single unified model for IMU-only body-frame velocity estimation across four platforms (car / quadruped / drone / human). Leaderboard public 0.31621.
Contents
| File | Description |
|---|---|
mosaic_v14.pt |
Frozen weights of the single unified model (16.75M params, one weight set) |
infer.py |
Self-contained inference entry point (no internet, no external repo imports) |
stats_norm.npz |
Train-set normalisation constants for the 22-d spectral statistics |
submission.csv |
The exact prediction file that produced the claimed leaderboard score |
requirements.txt |
Pinned dependencies (torch / numpy / pandas only) |
Reproduce the submission
pip install -r requirements.txt
python infer.py --data_dir /path/to/challenge_data \
--weights mosaic_v14.pt --stats stats_norm.npz --out submission.csv
data_dir is the official challenge release (needs index/test_windows.csv and
test/*.npz). Verified reproduction: output matches the leaderboard CSV to the
6-decimal rounding of the file format (max diff 1e-6).
Measured resource use (RTX 4060 Laptop 8GB): 14 s total, peak VRAM 0.85 GiB — far inside the ≤16 GB / ≤2 h re-execution budget. Runs offline.
Model
Frequency-split conv encoders (low band: dilated conv ×4 / high band: multi-scale conv) → causal ALiBi Transformer ×6 (window-level attention, future masked, extrapolates to full trajectories) → spectral-stats-conditioned MoE head (prototype router, 4 experts, feature-level fusion inside ONE network) → per-window velocity (4 × 0.25 s sub-window mean) + heteroscedastic uncertainty. Strictly causal (perturbation-verified). Inference: full-trajectory streaming + yaw ±10° 3-view TTA averaging.
Training used the official train split plus public external datasets (UZH-FPV, EuRoC MAV, MILUV), declared per the rules. No pretrained weights.