HERCULES — Multi-Robot Photorealistic Synthetic SLAM Dataset

A photorealistic synthetic dataset for multi-robot SLAM and collaborative perception — 2 aerial + 2 ground robots across 4 large-scale environments.

The four environments

Australian Outback — Center	Australian Outback — Perimeter
City Block	Forest

HERCULES provides time-synchronized, multi-modal sensor streams from a team of robots (2× drone, 2× Husky UGV) operating together in four large-scale environments. It targets research in SLAM / LiDAR-inertial & visual-inertial odometry, multi-robot / collaborative perception, depth estimation, and semantic segmentation.

The data is synthetic, generated by HERCULES — a simulation framework built on Unreal Engine 5 that extends AirSim (Shah et al., 2018) and Cosys-AirSim (Jansen et al., 2023) as a UE5 plugin, using Lumen global illumination and Nanite geometry for photorealistic rendering. It provides photorealistic imagery alongside perfect, noise-free ground truth for geometry, semantics, and trajectories. All streams share a common time base (synchronized capture).

---

Sequences (environments)

Folder	Environment	Approx. size
Australia Center Sequence/	Australian outback — center route	~379 GB
Australia Perimeter Sequence/	Australian outback — perimeter route	~427 GB
City Block Sequence/	Urban city block	~241 GB
Forest Sequence/	Dense forest	~312 GB

Total ≈ 1.1 TB. Designed trajectory lengths range 359–945 m per sequence, with intra- and inter-robot loop closures.

---

Sensors

Each robot (2× drone, 2× Husky UGV) carries an identical, synchronously-logged suite:

Modality	Details	Format · rate
RGB	front camera, 752×480, 90° FOV	.png · 20 Hz
Stereo	left + right, 752×480, 0.11 m baseline	.png · 20 Hz
Depth	planar metric depth, 752×480	.npy (float32, metres) + .png viz · 20 Hz
Segmentation	ground-truth semantic + instance labels	.png 752×480 (+ label_color_map_*.csv, 320 classes) · 20 Hz
LiDAR	16-channel, 200 m range, ~28,800 pts/scan	.npy N×3 (x,y,z) float32 metres · 20 Hz
IMU	linear accel + angular velocity (+ 9-axis variant)	imu.txt 200 Hz; synthetic_imu_9axis_{200,500}Hz.txt
Pose (GT)	global world-frame + odometry-frame pose	pose_world_frame.txt, odom.txt

Camera/LiDAR mounts, FOV, and the stereo baseline are specified in each sequence's data/settings.json.

---

Directory structure

All four sequences share the same layout (City Block additionally has a second results/openvins_BFeb8/ run):

<Sequence>/
├── data/
│   ├── Drone1/  Drone2/  Husky1/  Husky2/          # identical per-robot sensor suite:
│   │   ├── rgb/  rgb_stereo_left/  rgb_stereo_right/   # 752×480 PNG, 20 Hz
│   │   ├── depth/                                  # .npy (float32 metres) + .png viz, 752×480
│   │   ├── seg/                                    # GT segmentation PNG (see label_color_map_*.csv)
│   │   ├── lidar/                                  # .npy N×3 (x,y,z) point clouds, 16-ch, 20 Hz
│   │   ├── imu.txt                                 # IMU @ 200 Hz
│   │   ├── synthetic_imu_9axis_200Hz.txt / _500Hz.txt
│   │   └── pose_world_frame.txt  odom.txt          # ground-truth poses
│   ├── trajectory_information/                     # designed reference (waypoint) trajectories
│   ├── settings.json                              # capture config: sensor intrinsics + extrinsics
│   ├── label_color_map_*.csv                      # semantic class ↔ RGB (320 classes)
│   └── environment.png , UE5*world*.png            # environment reference imagery
└── results/
    ├── LIO-SAM/                                   # baseline LiDAR-inertial odometry output
    └── openvins/                                  # baseline visual-inertial odometry output

Filenames encode the capture time in simulation seconds (e.g. lidar/0.050000.npy → t = 0.05 s). Cameras + LiDAR are logged at 20 Hz (Δt = 0.05 s) and IMU up to 500 Hz; all streams share a common time base, so samples align across sensors and robots.

File formats

• Poses (pose_world_frame.txt, odom.txt): timestamp x y z qw qx qy qz — position in metres, unit quaternion (w-first). pose_world_frame is the global world frame; odom starts at the robot's origin.
• IMU (imu.txt): timestamp aₓ a_y a_z ωₓ ω_y ω_z at 200 Hz. The synthetic_imu_9axis_{200,500}Hz.txt files provide a 9-axis IMU at 200 / 500 Hz.
• Depth: .npy float32 planar depth in metres (with a .png for quick viewing).
• LiDAR: .npy array of N×3 (x, y, z) points in metres.
• Segmentation: .png whose colors map to classes via label_color_map_*.csv (columns: Label, ObjectName, SegmentationID, R, G, B; 320 classes). Instance IDs are consistent across robots for cross-view data association.
• World axis convention: the AirSim / Cosys-AirSim native world frame; sensor extrinsics (camera/LiDAR mounts, baseline) are in data/settings.json.

results/ — baseline odometry/SLAM outputs

Per-sequence outputs of the baselines benchmarked in the paper: LIO-SAM/ (LiDAR-inertial, Shan et al., 2020) and openvins/ (visual-inertial, Geneva et al., 2020). City Block additionally includes an alternate openvins_BFeb8/ run.

---

Dataset notes

• Noise-free ground truth. No sensor-noise model is applied — IMU, poses, depth, LiDAR, and segmentation are exact ground truth. (The simulator can inject per-sensor noise and latency, but it is off for this release.) Add noise externally if your method requires it.
• Trajectories are designed with HERCULES's Complementary Coverage planner; each begins with a static + calibration period.
• Dynamic objects (pedestrians, traffic, wildlife) are disabled during collection except birds.

---

How to download and unpack

To keep the dataset usable on the Hub, each per-robot / per-result folder is stored as a .tar.zst archive (raw loose files would exceed the Hub's 10,000-files-per-folder limit). Small metadata files (settings.json, label_color_map_*.csv, *.png) are stored uncompressed so you can preview them directly.

Download (whole dataset or a single sequence):

pip install -U "huggingface_hub[hf_xet]"

# everything:
hf download GeorgiaTech/HERCULES --repo-type dataset --local-dir HERCULES

# or just one sequence:
hf download GeorgiaTech/HERCULES --repo-type dataset \
    --include "Forest Sequence/*" --local-dir HERCULES

Unpack to the original tree (reproduces the exact folder structure, byte-for-byte):

cd HERCULES
./extract_all.sh          # extracts every .tar.zst in place; safe to re-run
# requires: tar + zstd  (sudo apt install zstd)

After extraction you get e.g. Forest Sequence/data/Drone1/lidar/769.900000.npy, identical to the source dataset. The .tar.zst files can then be deleted if you wish.

---

Intended uses

• Multi-robot / collaborative SLAM and pose-graph optimization
• LiDAR-inertial and visual-inertial odometry benchmarking (ground truth provided)
• Depth estimation and semantic/instance segmentation (perfect synthetic labels)
• Heterogeneous UAV–UGV perception; cross-environment / sim-to-real studies

License

Released under CC-BY-4.0 — free to use and adapt with attribution. This dataset accompanies a manuscript under review at the International Journal of Robotics Research (IJRR); please cite the paper below.

Citation

@misc{garimella2026hercules,
  title         = {HERCULES: An Open-Source Simulation Framework for Heterogeneous Multi-Robot SLAM, Collaborative Perception, and Exploration},
  author        = {Garimella, Sandilya Sai and Butterfield, Daniel Chase and Wilson, Sean and Gan, Lu},
  year          = {2026},
  eprint        = {2606.22756},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  url           = {https://arxiv.org/abs/2606.22756}
}

Contact / maintainers

Sandilya Sai Garimella, Daniel Chase Butterfield, Sean Wilson, and Lu Gan — Georgia Institute of Technology.

Acknowledgements

Built on Unreal Engine 5, AirSim (Shah et al., 2018), and Cosys-AirSim (Jansen et al., 2023).