Downloads · 30 days
9
25% of all-time downloads
hoppery/tartanimu-iros2026-mobilityai
tartanimu-iros2026-mobilityai is a other model from hoppery. Use it for the other 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 mit.
Single unified model for multi-platform inertial odometry: one network and one set of weights predicts body-frame velocity from a 200 Hz 6-axis IMU stream for car, quadruped, handheld and drone platforms, with no plat…
Downloads · 30 days
9
25% of all-time downloads
All-time downloads
36
Public
Repo size
163 MB
Likes
0
Public
Click a slice to open those files.
.pt163 MB · 99%
From the Hugging Face model README
MobilityAISingle unified model for multi-platform inertial odometry: one network and one set of weights predicts body-frame velocity from a 200 Hz 6-axis IMU stream for car, quadruped, handheld and drone platforms, with no platform label at inference.
| Weights | full_dv_unc_trev_s42_step75000.pt (md5 d445f47f5060edcdeadeb5d75f396869) |
| Submitted predictions | submission.csv (md5 bff815a85b1075b0dd664135f50af55e) |
| Kaggle public leaderboard | 0.23537 |
| Organisers' scoring service, all 89 test recordings | 0.18874 |
| Parameters | 8.12 M |
python submit.py --ckpt full_dv_unc_trev_s42_step75000.pt --split test --out sub.csv \
--device cuda:0 --data /path/to/challenge/data \
--stride 15 --alpha 1.02 --kf --regime_speed 2.9 --pair_avg --no_offset --pair_bs \
--kf_unc rel --unc_map exp
Requires torch, numpy, pandas, scipy. The command above regenerates submission.csv bit-exactly on
the hardware it was produced on (verified: 30644 rows, max absolute velocity difference 0.000e+00).
A TCN encoder over 200 Hz IMU windows, a bidirectional GRU that mixes a 60-window context, an attitude feature branch, a soft platform router, and a heteroscedastic uncertainty head. The router is trained with a cross-entropy auxiliary on the released platform labels and mixes softly at inference; no platform label is read at inference time. Trained for 75000 steps on the 475 released train+val recordings, seed 42.
Training recipe highlights (full configuration in config.json):
--time_reverse_twins 0.5 --trev_plats car,dog,human);--chunk_slack 1 --loss_per_chunk);exp(logvar) is an error-magnitude scale in m/s
and supplies the per-window measurement noise to the physics post-processing (--uncertainty);--exclude_ids).Everything after the network is deterministic, has no learned parameters, reads no labels, and uses only test inputs. Constants were fixed on held-out train/val recordings.
--stride 15) — each window is predicted inside several overlapping
60-window contexts and the predictions averaged.--alpha 1.02) — one scalar on the predicted velocities.--kf --regime_speed 2.9) — for recordings the router calls
drones whose predicted mean speed exceeds 2.9 m/s, window velocities are fused with 200 Hz IMU
increments under a Mahony attitude estimate. The observation model treats the network output as the
window mean of body velocity, not the centre value. Acceptance gates fall back to the raw network
output when the fusion moves too far or the accelerometer gate looks quasi-static.--pair_avg) — some flights are recorded by two IMUs and both recordings
are scored separately. Pairs are detected label-free by gyroscope-magnitude correlation at equal length,
the mounting rotation is fitted by Kabsch on low-passed gyro streams, and the predictions are averaged.--pair_bs) — one least-squares problem over both sensors sharing the
trajectory, solving for per-sensor attitude corrections, gyro and accelerometer biases and initial
velocity, with an acceptance gate that falls back to the plain pair average.--kf_unc rel --unc_map exp) — the head's
exp(clamp(logvar, -3, 3)) sets the per-window measurement noise in both the filter and the smoother.The published MD5 is a file-integrity fingerprint of the exact submitted file, not a reproduction criterion: different GPU hardware, a different inference batch size or a different revision of this code all change it. Judge reproduction as 99.9 % of velocity rows within 5e-4 m/s, no row beyond 5e-3 m/s, and the score agreeing to four decimals, running the code in this repository. Measured: bit-exact on the original hardware; max 1.65e-3 m/s against an RTX 3090, concentrated in a single Kalman-fused fast drone recording where an iterative filter amplifies floating-point differences; at most 5.75e-4 m/s between inference batch sizes 48 and 12; nothing measurable between different gradient-accumulation splits at equal effective batch. Running the same checkpoint through a three-day-old copy of this code moved velocities by up to 0.07 m/s, so the code revision matters far more than the hardware.
submit.py is the entry point. tartan_model.py, tartan_data.py, tartan_predict.py are the network,
the window store and the batched predictor. kf_apply.py, kf_core.py, kf_meanobs.py, batch_smooth.py,
att200.py, bslot_offset.py, acc_bias_fit.py, imucal.py, calibrate.py, kf_fuse.py are the physics
post-processing. train.py, tartan_loss.py, ate_loss.py reproduce training. multinet.py is imported
only by a checkpoint option this model does not use. kaggle_metric_tartanimu_score.py is the organisers'
metric, included so valscore.py runs offline.
One model, one weight set. No ensembling, no checkpoint averaging, no multi-member mixture. No test data in training; the submitted checkpoint is the final step of the schedule and no checkpoint was selected on test feedback. All post-processing is label-free and reads only test inputs.