Downloads · 30 days
0
xunden/tartanimu-xunden-iros2026
tartanimu-xunden-iros2026 is a robotics model from xunden. Use it for the robotics task on the model card, and read the license before you ship it in a product. It is set up for pytorch. The card lists the license as cc-by-nc-sa-4.0.
One model, one set of weights, for the IROS 2026 TartanIMU cross-platform inertial odometry challenge. The network predicts body-frame velocity from 6-axis IMU alone for car, drone, quadruped and handheld motion; ther…
Downloads · 30 days
0
Access
Public
Updated Sep 24, 2026
Repo size
361 MB
Likes
0
Public
Click a slice to open those files.
.pt361 MB · 99%
From the Hugging Face model README
One model, one set of weights, for the IROS 2026 TartanIMU cross-platform inertial odometry challenge. The network predicts body-frame velocity from 6-axis IMU alone for car, drone, quadruped and handheld motion; there are no per-platform experts and no model switching at inference.
| Quantity | Value |
|---|---|
| Official per-sequence scoring service (all 89 test trajectories) | 0.25535 |
| Kaggle public leaderboard | 0.33229 (rank 20 of ~131 at the time of writing) |
| Submitted CSV md5 | 7364d39a906ec23625bbdf11c12dd865 |
| Drone share of the score | 60.6 % (the dominant platform) |
An all-zero submission scores exactly 1.000. The best public score is ≈ 0.20074, so this entry sits mid-field: strong enough to be competitive, not near the leaders.
| Architecture | resnet_bilstm_attn_mh — dilated ResNet window encoder, BiLSTM, multi-head self-attention, shared velocity head + softmax-routed platform-conditioned expert mixture |
| Parameters | 90.4 M (single checkpoint, best.pt, 361 MB) |
| Input features | rich2, 31 channels per 0.5 s token: raw IMU (6) + first differences (6) + high-pass dynamic acceleration (3) + channel norms (2) + attitude-grounded linear acceleration (3) + gravity direction (3) + spectral band-energy fractions (8) |
| Window / context | 1 s windows split into 2 tokens; 20-window (20 s) sequences |
| Training data | train + val (475 labelled trajectories), 200 epochs, cosine schedule |
| Checkpoint epoch | 196 |
The checkpoint is self-contained: it stores the per-channel mean/std used for input
standardization plus all architecture hyper-parameters, so inference needs no extra flags
or training-time artefacts.
python -m tartan_imu.predict_tcn \
--data-root /path/to/data/root \
--checkpoint best.pt \
--output submission.csv \
--split test --stride 5 --tta-shifts 0,25,-25 --num-workers 8
--stride 5 and --tta-shifts 0,25,-25 are the inference configuration of the submitted
system (see compliance item 3). Using --stride 20 --tta-shifts 0 instead reproduces the
plain single-pass prediction, which scored 0.25852 on the same service — i.e. the same
weights are worth −0.0032 with multi-view averaging.
Determinism: inference runs under model.eval() with torch.no_grad(), no dropout and no
stochastic input augmentation. Re-executing this checkpoint with the code in code/
reproduces the submitted CSV bit-exactly — identical md5
7364d39a906ec23625bbdf11c12dd865 (per-axis correlation 1.0, maximum deviation 0.0),
verified as part of the release; the single-pass variant likewise regenerates
18ccdbed10f5d9923f467b3ba84e7a55 exactly. The released best.pt is byte-identical to
the selected training checkpoint at epoch 196 (md5
6411d0464dc6c6e6ff55a49c2385fd13).
| Path | Contents |
|---|---|
best.pt | The submitted single checkpoint (weights + standardizer + hyper-parameters). |
reproduce.sh | Regenerates and format-validates the submission CSV. |
code/tartan_imu/ | Minimal inference code (predict_tcn, sequence_data, data, tcn_model) — the same code that regenerated the submission bit-exactly. |
code/requirements.txt, code/run_inference.sh | Dependencies and a one-line runner. |
eval/table_v_per_sequence.csv | Official-scoring-service per-sequence table for the submitted CSV (Table V of the report). |
eval/scoring_service.json | Raw official-scoring-service output (Tables III–IV). |
build_samples appends a final sequence so tail windows are covered
twice. No input augmentation (noise, rotation, filtering) is applied at test time.Measured on the same service/leaderboard: the organizers' official augmentation hurt (+0.0406 on honest val), a 31-channel gravity/spectral extension of the rich features was neutral (−0.0005), longer schedules saturated (100→150 epochs −0.0051, 150→200 −0.0003), capacity away from scale 4 hurt (scales 2/3/8), low-pass filtering the input hurt (including the high-vibration subgroup it was meant to fix), high-frequency oversampling hurt, warm-restart continuation produced a byte-identical submission, and weight-space averaging was catastrophic across runs (different loss basins) and slightly negative within a run.
submission.csv is the exact CSV behind our selected Kaggle entry (original
filename submission_RICH2_200_tta.csv, MD5 7364d39a906ec23625bbdf11c12dd865); reproduce.sh regenerates a byte-identical
file from best.pt. requirements.txt pins the exact environment used.