EgoSpatial-CL is a compact multimodal synthetic dataset designed for continual-learning and imitation-learning studies in simulated autonomous navigation. The release emphasizes spatially controlled train/validation/test splits, complete synchronized episodes, deterministic generation metadata, and public release files suitable for scientific reuse.

The dataset was generated with CARLA 0.9.15. CARLA is referenced as the simulator used to generate the data, not as part of the dataset title.

Release Status

Dataset version: v1.0.0
Dataset DOI: pending, not minted yet
Release status: v1.0.0 release candidate for final review
Access: public, non-gated
License: CC BY 4.0
Code repository: https://github.com/checo1092/egospatial-cl-code

The DOI should be minted only after this metadata release has been reviewed and frozen.

Summary

Simulator: CARLA 0.9.15
Towns: 7
Compact domains per town: 7
Spatial slots per town/domain pair: 10
Episodes: 490
Synchronized samples: 490,000
Train/validation/test episodes: 392/49/49
Technical split labels: train/val/test
Format: tar shards tracked with Git LFS
Episode ordering: domain_major
Modalities: RGB, semantic segmentation, LiDAR, vehicle state/control, IMU/GNSS fields

Compact Domains

0 clear__NPC00_none
1 night__NPC00_none
2 fog__NPC00_none
3 hardrain__NPC00_none
4 clear__NPC11_veh_high
5 hardrain__NPC11_veh_high
6 clear__NPC31_both_high

Layout

shards/<Town>/<split>/<Town>_<split>_XXXXXX.tar

Each tar contains complete episodes. Episode directory names are unique within a town; the release-level unique identifier is <Town>/<split>/episode_XXXX.

<Town>/<split>/episode_XXXX/
  images/
  labels/
  lidar/
  meta.json
  state.csv
  index.csv

Download

Prerequisite: Git LFS

The dataset shards are tracked with Git LFS. Install and initialize Git LFS before cloning the repository. On Ubuntu/Debian:

sudo apt update
sudo apt install -y git-lfs
git lfs install
git lfs version

Installing Git LFS is normally required only once per system. The git lfs install command initializes it for the current user.

Clone metadata and Git LFS pointers without downloading all shards:

export REPO="checo1092/egospatial-cl"
export OUT="$HOME/EgoSpatial_CL"

mkdir -p "$OUT"

GIT_LFS_SKIP_SMUDGE=1 git clone \
  "https://huggingface.co/datasets/$REPO" \
  "$OUT/repo"

cd "$OUT/repo"

Download one town, for example Town01_Opt:

git lfs pull \
  --include="shards/Town01_Opt/**" \
  --exclude=""

Smoke-test a single public shard before downloading a larger subset:

curl -L --fail \
  -o Town06_val_000000.tar \
  "https://huggingface.co/datasets/$REPO/resolve/main/shards/Town06/val/Town06_val_000000.tar"

printf "%s  %s\n" \
  "cd4d7601d61cbd0ee3c21294305bc63843eefe7a1334b2440bd78b1a2290c412" \
  "Town06_val_000000.tar" | sha256sum -c -

tar -tf Town06_val_000000.tar | sed -n '1,10p'

Extract one town:

export UNPACKED="$OUT/unpacked"
mkdir -p "$UNPACKED"

find shards/Town01_Opt -name '*.tar' | sort | while read -r TAR; do
  tar -xf "$TAR" -C "$UNPACKED"
done

Download the full dataset only when sufficient storage is available:

df -h "$OUT"
git lfs pull

The packed shards occupy approximately 299 GB. Extracting the full dataset requires additional disk space.

Extract all towns and splits:

export UNPACKED="$OUT/unpacked"
mkdir -p "$UNPACKED"

find shards -type f -name '*.tar' -print0 |
  sort -z |
  while IFS= read -r -d '' TAR; do
    echo "Extracting: $TAR"
    tar -xf "$TAR" -C "$UNPACKED"
  done

Verify the number of extracted episodes:

find "$UNPACKED" -type d -name 'episode_*' | wc -l

The expected result for the complete v1.0.0 release is 490.

Splits

The split is spatial and frozen by simulator spawn index. For each town, 10 representative spawn indices were selected:

8 train spawns
1 validation spawn
1 test spawn

The same spatial slots are reused across all seven compact domains of the town. This prevents the same driving corridor from appearing in both training and evaluation under different weather or traffic conditions.

Across the full dataset:

7 towns x 70 episodes = 490 episodes
train = 392 episodes
validation = 49 episodes
test = 49 episodes

See spatial_splits.json for the final frozen spawn indices.

Release Files

dataset_info.json          Dataset-level release summary.
benchmark_manifest.json    Episode, domain, split, shard, and provenance manifest.
spatial_splits.json        Frozen spatial train/validation/test split.
shards_manifest.csv        Shard-level manifest with SHA256 checksums.
checksums.sha256           SHA256 checksum file for all tar shards.
docs/data_dictionary.md    Field and episode-structure documentation.
docs/sensor_configuration.md
                           Sensor and simulation configuration.
docs/data_provenance.md   Provenance, license boundary, and DOI status.
THIRD_PARTY_NOTICES.md     Third-party software and simulator notices.
CITATION.cff              Citation metadata.
VERSION                   Dataset version.
LICENSE                   Dataset license statement.

Dataset Viewer

The Hugging Face Dataset Viewer is disabled for this release because the dataset is distributed as complete episode tar shards rather than a WebDataset-native sample layout. The release was not repacked solely for viewer compatibility, which would require rewriting approximately 299 GB of packed data.

The accompanying code repository provides a local visualization tool for synchronized RGB, semantic-segmentation, LiDAR, and vehicle-state samples.

Example: Town06/val/episode_0008, frame 83569.

Citation

Please cite the dataset DOI after it is minted for the frozen v1.0.0 release. Until then, cite this repository and version explicitly.

Sergio Delgadillo Marín, Juan Humberto Sossa Azuela, and Erik Zamora Gómez. EgoSpatial-CL: A Spatially Controlled Multimodal Dataset for Continual Learning in Simulated Autonomous Navigation. Version 1.0.0. Hugging Face dataset repository: https://huggingface.co/datasets/checo1092/egospatial-cl

All authors contributed equally to this work.

Limitations

EgoSpatial-CL is a simulated dataset. It does not provide continuous intra-sequence domain transitions, optical flow, dense depth, 2D/3D object boxes, or multi-view camera annotations. It should be used as a compact, spatially controlled dataset for navigation/control-oriented studies rather than as a replacement for large-scale perception datasets.

Downloads last month: 727

Total file size:

299 GB