goodmattg/vigil-public-data

Vigil Autonomy Public Trajectory Dataset

This dataset contains privacy-preserving exports from the Vigil Autonomy (vigilautonomy.com) synchronized video and RTK capture platform. Each trajectory pairs a sanitized video stream with relative frame timestamps, segment-local RTK positions, and orientation samples.

This six-second excerpt comes from a calibrated trajectory recorded on August 4, 2026. The left panel shows the raw video frames. The right panel shows the exact same frames with Vigil's synchronized target projections rendered in Rerun: the red box is the projected bounding-sphere envelope, and the green box is the projected drone-body footprint.

Dataset Summary

The current dataset-v1 release contains:

• 33 trajectories and 33 video streams;
• 55,665 video frames;
• 7,575 segment-local position samples;
• 188,647 orientation samples; and
• 953.891 seconds of recorded trajectory time.

Thirty-two calibrated trajectories recorded on August 4, 2026, were added on August 9, 2026. The release also retains the dataset's original public trajectory.

The files are trajectory bundles rather than a tabular Hugging Face datasets configuration. Download the repository files directly or use huggingface_hub to retrieve individual bundles.

Directory Layout

dataset-v1/
  traj_<opaque-id>/
    README.md
    metadata.json
    manifest.json
    segments/
      seg_<opaque-id>/
        camera/
          camera_000.mkv
          camera_000.frames.json
        position/
          position_*.json
        orientation/
          orientation_*.json

Each metadata.json summarizes trajectory duration and record counts. Each manifest.json defines the timebase, coordinate system, artifact paths, media types, record counts, byte sizes, and SHA-256 checksums.

Privacy-Preserving Export

The bundles were generated with Vigil's public trajectory export introduced in Vigil commit 38224be.

The export applies these transformations:

• Native trajectory and segment identities become keyed, opaque pseudonyms.
• Every timestamp becomes an integer nanosecond offset from its segment start.
• Frame and sample numbers restart at zero within each public stream.
• Absolute RTK coordinates become a privately translated and rotated, segment-local Cartesian frame measured in meters.
• Source clocks, absolute position, hardware identities, mount geometry, capture settings, raw sensor messages, and internal timing metadata are excluded.
• Video streams are remuxed with normalized packet timing and cleared container metadata.

Video pixels are intentionally not anonymized. People, vehicles, landmarks, text, or other identifying visual content may remain visible. Users must evaluate pixel-level privacy for their application.

Time and Coordinate Semantics

All public timestamps use the bundle's relative nanosecond timebase. They must not be interpreted as UTC, system time, or the original monotonic capture clock.

Positions use a right-handed, segment-local Cartesian frame. The private translation and rotation prevent recovery of the source's absolute geolocation from the published coordinates. Position samples and camera frames may have different sampling rates, so they are not necessarily one-to-one.

Basic Use

import json
from pathlib import Path

trajectory = Path("dataset-v1/traj_<opaque-id>")
metadata = json.loads((trajectory / "metadata.json").read_text())
manifest = json.loads((trajectory / "manifest.json").read_text())

print(metadata["frame_count"])
for segment in manifest["segments"]:
    for artifact in segment["artifacts"]:
        print(artifact["kind"], artifact["path"], artifact["sha256"])

Consumers should verify artifact byte sizes and SHA-256 checksums against manifest.json after downloading.

Limitations

• The dataset currently contains one segment and one primary camera stream per trajectory.
• The current release contains no public event samples.
• RTK sampling is sparser than video sampling in these trajectories. Downstream users must choose an interpolation, nearest-neighbor, or withholding policy appropriate to their application.
• The public coordinate transform preserves local motion and relative geometry, but it intentionally prevents absolute map placement.
• Camera calibration, mount geometry, device identities, raw sensor messages, and original timing diagnostics are not part of the public schema.

Integrity

At publication, all 33 trajectory roots and 132 manifest-declared artifacts were read back from Hugging Face. Published file sizes and SHA-256 hashes matched their corresponding public manifests.

License and Attribution

Copyright 2026 Matthew Goodman. This dataset is licensed under the Creative Commons Attribution 4.0 International License. Any use, adaptation, or redistribution must give appropriate credit to Matthew Goodman, provide a link to the license, and indicate whether changes were made. The citation below is the preferred form of attribution for publications and other works that support formal citations.

Citation

@dataset{goodman_2026_vigil_autonomy_public_trajectory_dataset,
  author    = {Goodman, Matthew},
  title     = {Vigil Autonomy Public Trajectory Dataset},
  year      = {2026},
  publisher = {Vigil Autonomy},
  url       = {https://huggingface.co/datasets/goodmattg/vigil-public-data}
}