Hugging Face's logo

Datasets:
YiZhaoJasper
/
Tartan_IMU_VBR

Tasks:
Robotics
Time Series Forecasting
Languages:
English
Size:
1M<n<10M
Tags:
imu
inertial-odometry
inertial-tracking
dead-reckoning
autonomous-driving
handheld
License:
Dataset card Files
xet
 Community
Dataset Viewer

The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.

VBR → TartanIMU (car + handheld IMU, retargetted)

Drop-in compatible with the official TartanIMU AirLabNPZSequence data loader. No source modifications and no monkey-patch required — sequence paths are structured so TartanIMU's motion_type substring matcher routes every file to the correct head (car or human) automatically.

This is the VBR (Vision Benchmark in Rome) IMU + ground-truth-pose subset, re-projected into the schema expected by TartanIMU's dataloader/dataset_AirLab.py::AirLabNPZSequence. It complements the Tartan_IMU_Nymeria dataset (head-/wrist-mounted Aria human IMU) with vehicular motion and urban handheld walking — both of which are under-represented in the foundation-model training corpus.

At a glance
sequences 8 (4 car + 4 human)
traj.npz files 8 (one per sequence)
total IMU samples @ 200 Hz 1 343 989 (1.34 M)
total recording duration 6 720 s (≈ 1.87 h)
total disk size 65.5 MB (.npz zlib-compressed)
splits train / test = 6 / 2 trajectory files
coordinate convention body = X-forward, Y-left, Z-up    world = ENU, gravity-aligned
sample rate strictly uniform 200 Hz, float64 timestamps
audit pass-rate 8 / 8 (100 %) on 10 deep checks per NPZ
source dataset VBR (Vision Benchmark in Rome), Sapienza University of Rome
upstream license BSD 3-Clause (vbr-devkit)

Authoritative machine-readable stats: _dataset_card.json. Per-file 10-check audit: audit_summary.json.

1. Why VBR?

The VBR dataset was recorded by the RVP Group at Sapienza University of Rome with a multi-sensor rig carrying an industrial-grade SBG Ellipse-E MEMS IMU plus stereo cameras, multi-beam LiDAR, and RTK-GPS. Two recording modes are exercised by the public sequences:

  • Vehicular motion (car) — an Alfa Romeo Stelvio driving through Rome's university campus and the Ciampino district. Speeds up to ~26 m/s, sustained accelerations up to ~3 g, and rich dynamics including roundabouts, stop-and-go traffic, and elevation changes.
  • Pedestrian / handheld motion (human) — the same rig carried on foot through DIAG building, the Colosseo area, the Pincio gardens, and Piazza di Spagna. Speeds 1–2 m/s with realistic urban walking dynamics.

This makes VBR a natural complement to Nymeria for TartanIMU pretraining:

platform source dataset hours speed range environment
car this repo (VBR) 0.98 4.5–26 m/s urban + suburban
human this repo (VBR, handheld) 0.89 1.0–3.7 m/s indoor + outdoor
human Tartan_IMU_Nymeria 928.30 walking indoor (Aria glass)

2. Quick start 

# install
pip install -U "huggingface_hub[cli]" numpy scipy

# pull the whole dataset (resumable; ~65 MB)
huggingface-cli login                                          # one-time, paste your read token
huggingface-cli download YiZhaoJasper/Tartan_IMU_VBR \
    --repo-type dataset \
    --local-dir ./tartan_imu_vbr \
    --resume-download

# verify nothing is missing or pointer-only
cd tartan_imu_vbr
python examples/verify_integrity.py

A 5-line sanity check:

import numpy as np
from pathlib import Path

p = next(Path("data/car/vbr/train").glob("*/traj.npz"))
with np.load(p) as z:
    ts, imu, pos, quat = (z["retargetted_ts"], z["retargetted_imu"],
                          z["retargetted_pos"], z["retargetted_quat"])
print(f"N={len(ts)} dur={ts[-1]-ts[0]:.1f}s  |q|={np.linalg.norm(quat,axis=1).mean():.6f}")

For TartanIMU-flavoured loading, see the example scripts in examples/.

3. Directory layout 

Tartan_IMU_VBR/                                ← HF repo root after download
├── README.md                                  ← this file
├── LICENSE                                    ← BSD-3 (data, inherited from VBR) + MIT (retargetting code)
├── CITATION.cff                               ← cite VBR + TartanIMU
├── .gitattributes                             ← *.npz → LFS
│
├── _dataset_card.json                         ← machine-readable stats / audit
├── audit_summary.json                         ← compact 10-check deep-audit roll-up
├── MANIFEST.sha256                            ← sha256 + size of every traj.npz
│
├── examples/
│   ├── load_raw_numpy.py                      ← framework-agnostic; pure np.load
│   ├── load_via_airlab.py                     ← uses TartanIMU's AirLabNPZSequence
│   ├── load_for_tartanimu_training.py         ← full BasicSequenceData + SeqToSeqDataset
│   └── verify_integrity.py                    ← SHA-256 check against MANIFEST.sha256
│
├── pipeline/                                  ← full retargetting source for reproducibility
│   ├── README_vbr_pipeline.md
│   ├── README_download_toolkit.md
│   ├── vbr_full_pipeline.py                   ← orchestrator: download → retarget → cleanup
│   ├── vbr_to_tartanimu.py                    ← driver: per-sequence retarget
│   ├── _vbr_pipeline_core.py                  ← coordinate transforms + integrity checks
│   ├── stream_extract_imu.py                  ← streaming ROS1 bag parser (fallback)
│   ├── validate_vbr_npz.py                    ← 10-check deep audit
│   ├── vbr_run_all.sh                         ← one-click bash launcher
│   ├── vbr_safe_range_download.sh             ← parallel byte-range bag download
│   └── check_status.sh                        ← quick pipeline status monitor
│
└── data/
    ├── car/vbr/                               ← motion_type → 1 (vehicular)
    │   ├── train/
    │   │   ├── campus_train0/traj.npz         ← Sapienza campus loop
    │   │   ├── campus_train1/traj.npz         ← Sapienza campus loop (longer)
    │   │   └── ciampino_train0/traj.npz       ← Ciampino district drive (23 min)
    │   └── test/
    │       └── ciampino_train1/traj.npz       ← Ciampino district drive (held out)
    │
    └── human/vbr/                             ← motion_type → 4 (pedestrian / handheld)
        ├── train/
        │   ├── colosseo_train0/traj.npz       ← around the Colosseo
        │   ├── pincio_train0/traj.npz         ← Pincio / Villa Borghese gardens
        │   └── spagna_train0/traj.npz         ← Piazza di Spagna
        └── test/
            └── diag_train0/traj.npz           ← DIAG building (indoor + outdoor mix)

Per-sequence stats (samples / duration / size) are listed verbatim in _dataset_card.json under "sequences".

4. NPZ schema (per file)

Each traj.npz is an np.savez_compressed archive containing exactly four arrays — identical schema to Tartan_IMU_Nymeria.

key shape dtype meaning
retargetted_ts (N,) float64 Seconds since the first frame. Strictly monotonic, uniform spacing of 1/200 s = 5 ms (jitter < 1 ns).
retargetted_imu (N, 6) float32 Body-frame [ax, ay, az, gx, gy, gz]. Accelerometer in m/s² and includes gravity (a stationary device reads ≈ 9.81 along +Z). Gyro in rad/s.
retargetted_pos (N, 3) float32 World-frame position [x, y, z] in metres. World frame is ENU (gravity-aligned, Z up).
retargetted_quat (N, 4) float32 xyzw (scalar last) quaternion describing the body→world rotation. Per-sample norm ∈ [0.999, 1.001].

This matches dataloader/dataset_AirLab.py:138-143 exactly — AirLabNPZSequence will load these arrays without any further preprocessing.

Coordinate frames

  • VBR LiDAR frame (the source frame, as it comes out of vbr_calib.yaml): X = forward, Y = left, Z = up (RFL — happens to coincide with the TartanIMU body frame).
  • VBR IMU frame (the SBG Ellipse-E sensor frame): rotated relative to LiDAR; the static extrinsic T_LiDAR_IMU from vbr_calib.yaml is composed out automatically by pipeline/_vbr_pipeline_core.py.
  • TartanIMU body frame (what we store): X = forward, Y = left, Z = up.
  • World frame: ENU (East-North-Up), with gravity along −Z (i.e. the world Z axis points up, away from the Earth's centre — consistent with imu[:, 2] ≈ +9.81 when the platform is at rest).

The full coordinate-frame derivation, axis-test matrix, and gravity-sign sanity proof live in pipeline/README_vbr_pipeline.md.

5. Splits

The eight public VBR sequences are mapped to TartanIMU train/test as follows:

split car human
train campus_train0, campus_train1, ciampino_train0 colosseo_train0, pincio_train0, spagna_train0
test ciampino_train1 diag_train0

Following the official VBR benchmark policy, sequences whose ground-truth is publicly released and which we use here are all named *_trainN; the held-out *_test* evaluation set on the VBR benchmark ships no public ground truth, so it cannot be retargetted into supervised training data. We therefore deterministically reserve ciampino_train1 and diag_train0 (one car + one handheld) as our local test split — they are held out from any TartanIMU training run.

6. Quality assurance — 10 checks per NPZ

Every traj.npz is validated by pipeline/validate_vbr_npz.py.

# check threshold result (8 files)
1 schema (4 keys, dtypes, shapes) exact 8 / 8 ✅
2 timestamps strictly monotonic + uniform 200 Hz Δt jitter < 1 ns 8 / 8 ✅
3 no NaN / Inf in any array — 8 / 8 ✅
4 per-sample ‖quat‖ ∈ [0.999, 1.001] 8 / 8 ✅ (observed 1.000000)
5 IMU bounds: ‖a‖_max ≤ 245 m/s² (~25 g), ‖ω‖_max ≤ 52.4 rad/s (~3000 °/s) upstream-quality cap 8 / 8 ✅
6 per-sample max position step < 0.5 m 8 / 8 ✅
7 velocity 99-th percentile sanity < 50 m/s 8 / 8 ✅
8 static-block gravity (lowest-variance 2 s window) ‖g_world‖ ∈ [8.0, 11.0] m/s² 8 / 8 ✅ (observed [9.67, 9.75])
9 gyro vs quaternion-derivative MAE < 0.5 rad/s 8 / 8 ✅
10 sequence duration ≥ 60 s 8 / 8 ✅

Re-run any time:

python pipeline/validate_vbr_npz.py --root data

TartanIMU loader smoke test

We additionally verified that all 8 sequences load via TartanIMU's AirLabNPZSequence and produce valid training batches:

sequence motion type windows available features shape targets shape
campus_train0 car (1) 120 204 (200, 6) (200, 3)
campus_train1 car (1) 116 517 (200, 6) (200, 3)
ciampino_train0 car (1) 278 851 (200, 6) (200, 3)
ciampino_train1 car (1) 188 239 (200, 6) (200, 3)
colosseo_train0 human (4) 159 737 (200, 6) (200, 3)
diag_train0 human (4) 159 690 (200, 6) (200, 3)
pincio_train0 human (4) 159 559 (200, 6) (200, 3)
spagna_train0 human (4) 159 592 (200, 6) (200, 3)

Total 1 342 389 valid 200-sample training windows (1.34 M).

7. Loading examples

7.1 Raw NumPy (no TartanIMU dependency) 

from pathlib import Path
import numpy as np

p = next(Path("data/car/vbr/train").glob("*/traj.npz"))
with np.load(p) as z:
    ts   = z["retargetted_ts"]    # (N,)   float64
    imu  = z["retargetted_imu"]   # (N, 6) float32   body-frame [a; ω], gravity INCLUDED in a
    pos  = z["retargetted_pos"]   # (N, 3) float32   world-frame position (m), ENU
    quat = z["retargetted_quat"]  # (N, 4) float32   body→world rotation, xyzw

dt = float(np.median(np.diff(ts)))                # = 0.005 s
g  = np.linalg.norm(imu[:, :3], axis=1).mean()    # ≈ 9.8
print(f"N={len(ts)} dur={ts[-1]-ts[0]:.1f}s dt={dt*1e3:.3f}ms <|a|>={g:.3f} m/s²")

See full script: examples/load_raw_numpy.py.

7.2 Single-file via AirLabNPZSequence

No patch required — the sequence directories live under data/car/.../ and data/human/.../, so TartanIMU's substring matcher routes the right motion_type automatically. (None of our 8 sequence names contain the substrings car, dog, drone, or human, so there is no false-match hazard like there was for Nymeria.)

import sys
sys.path.insert(0, "/path/to/TartanIMU-master")
from dataloader.dataset_AirLab import AirLabNPZSequence

seq = AirLabNPZSequence(
    data_path="data/car/vbr/train/campus_train0/traj.npz",
    imu_freq=200, window_size=200,
)
assert seq.valid and int(seq.motion_type) == 1   # car
print(seq.features.shape, seq.targets.shape)      # (120403, 6) (120203, 3)

See full script: examples/load_via_airlab.py.

7.3 Full training-ready dataset 

python examples/load_for_tartanimu_training.py \
    --tartanimu-master /path/to/TartanIMU-master \
    --data-root data --num 3

See full script: examples/load_for_tartanimu_training.py.

8. Recommended training workflow

The following four-step sequence surfaces data or integration issues early, before committing to a full GPU training run.

  1. Smoke test (1 min, CPU) — load a sample of NPZ files from each split and verify schema + finite + qnorm:

    python examples/load_raw_numpy.py --motion car   --split train --num 3
python examples/load_raw_numpy.py --motion human --split train --num 3

  2. Integrity check (one-time, < 1 min) — SHA-256 every file:

    python examples/verify_integrity.py

  3. AirLab loader test (1 min) — confirm batches assemble:

    python examples/load_for_tartanimu_training.py \
    --tartanimu-master /path/to/TartanIMU-master \
    --data-root data --num 5

  4. Full TartanIMU training (GPU) — see the official entry point. A minimal config patch:

    data:
  dataset: AirLab
  data_path:
    car:   /path/to/this/dataset/data/car/vbr      # 4 sequences, 0.98 h
    human: /path/to/this/dataset/data/human/vbr    # 4 sequences, 0.89 h
    # … plus your existing dog / drone roots, plus Nymeria for human
  use_local_coord: True
  train_dir: train
  test_dir: test
  imu_freq: 200.0
  sample_freq: 40

train:
  active_heads: ["car", "human"]

    Combining VBR (this repo) with Tartan_IMU_Nymeria gives you the full human head data + a clean car head for the AirLab CMU foundation model.

9. Provenance & reproducibility

  • Source dataset — VBR (Vision Benchmark in Rome), RVP Group, Sapienza University of Rome, 2024. Original ROS1 .bag files + TUM-format ground-truth trajectories.
  • Reader — rosbags for parseable bags, plus a custom in-house streaming parser (stream_extract_imu.py) as a fallback for partial/corrupted bags.
  • Conversion — pipeline/vbr_to_tartanimu.py
  • IMU pipeline — SBG Ellipse-E (100 Hz native) → T_LiDAR_IMU extrinsic composition → axis transform to TartanIMU body frame → linear interpolation to uniform 200 Hz → SLERP for ground-truth quaternion → 10-check audit.
  • Generated on — 2026-05-18, single-host pipeline using aria2c for multi-connection downloads + sequential disk-bounded conversion (peak disk usage ≤ 25 GB during processing).
  • Target loader version — TartanIMU as released at https://superodometry.com/tartanimu (May 2026).

To reproduce from raw .bag files:

# put the unmodified pipeline tools in your workspace; aria2c required
python pipeline/vbr_full_pipeline.py
# or, individually:
python pipeline/vbr_to_tartanimu.py \
    --seq-id colosseo_train0 --motion-type human \
    --out-root dataset/vbr_retargetted

The pipeline is fully idempotent at the sequence level — restart from crash with no manual intervention and it picks up exactly where it left off.

10. License & citation

This dataset is a derivative work of VBR (Vision Benchmark in Rome). The retargetting code in pipeline/ and examples/ is released under the MIT License; the IMU + ground-truth-pose data in data/ is redistributed under the same BSD 3-Clause License that the upstream vbr-devkit carries, with attribution required to:

  • Brizi et al., VBR: A Vision Benchmark in Rome, ICRA 2024.
  • Robotics Vision and Perception Group (RVP-Group), Sapienza University of Rome.

Please cite both the original VBR paper and the TartanIMU paper:

@inproceedings{brizi2024vbr,
  title     = {VBR: A Vision Benchmark in Rome},
  author    = {Brizi, Leonardo and Bartoccioni, Florent and others},
  booktitle = {IEEE International Conference on Robotics and Automation (ICRA)},
  year      = {2024},
  url       = {https://rvp-group.net/slam-dataset.html}
}

@inproceedings{zhao2025tartan,
  title     = {Tartan IMU: A Light Foundation Model for Inertial Positioning in Robotics},
  author    = {Zhao, Shibo and Zhou, Sifan and Blanchard, Raphael and Qiu, Yuheng and
               Wang, Wenshan and Scherer, Sebastian},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  pages     = {22520--22529},
  year      = {2025},
  url       = {https://superodometry.com/tartanimu}
}

Also see CITATION.cff for a machine-readable citation file.

11. Acknowledgements

  • The RVP Group at Sapienza University of Rome for releasing VBR with permissive BSD-3 licensing and for the carefully calibrated multi-modal rig.
  • The TartanIMU authors at AirLab / Carnegie Mellon University for the multi-head IMU foundation-model architecture this dataset is built to feed.
  • The maintainers of rosbags, scipy, and aria2c for the building blocks that made the conversion practical on a 320 GB workstation.

12. Versioning & changelog

version date notes
1.0.0 2026-05-18 First public release. 8 / 8 deep-audit pass. Drop-in compatible with TartanIMU's AirLabNPZSequence.

See also

Downloads last month
13
Total file size:
65.7 MB