jp-cypress/DataFurnace-AMR

🏭 DataFurnace: AMR & SLAM Synthetic Evaluation Dataset

Fully synchronized multi-pass rendering (RGB, Depth, Semantic Mask, and 3D Bounding Box).

---

📌 Overview

DataFurnace is a procedurally generated synthetic dataset designed for evaluating and training:

• Autonomous Mobile Robots (AMR)
• Automated Guided Vehicles (AGV)
• SLAM / VIO / 3D Perception algorithms
• Warehouse robotics navigation systems

All ground truth is generated directly from the underlying 3D scene graph, ensuring deterministic pixel-level and millimeter-level accuracy across all modalities.
No generative AI is used, eliminating copyright, privacy, and hallucination concerns.

---

🚀 Key Features

🔹 1. Perfectly Synchronized Multi-Pass Rendering

Each frame contains fully aligned multi-modal outputs:

• RGB — Physically lit color images
• Depth — True Z-depth with LiDAR noise simulation
• Semantic Mask — Pixel-perfect class segmentation (Floor, Rack, Box, Pallet)
• 3D Bounding Box (JSON) — 6DoF absolute coordinates and dimensions
• Costmap — 2D top-down occupancy grid (generated per scene/mode)

🔹 2. Extreme Hazard Simulation for Robustness Testing

To stress-test SLAM and perception systems, four lighting conditions are provided:

• NORMAL — Standard warehouse lighting
• BROKEN — Random darkened areas / flickering lights
• GLARE — Volumetric scattering, lens flare, white-out
• EDGE_CASE — Foreground blackout + background overexposure

---

📂 Dataset Structure

This repository contains curated sample sequences generated by the DataFurnace pipeline.

• 5 Unique Warehouse Environments
• Layout Variations (Normal / Anomaly with Scattered Obstacles)
• 4 Lighting Conditions per environment
• Sequential Frames per sequence (from robot cameras: Front/Back/Left/Right)

Directory Layout (Example)

dataset/
  └── Warehouse_Scene_0001/
      └── normal/
          └── NORMAL/
              ├── Warehouse_Scene_0001_NORMAL_Cam_Front_F001_Normal_RGB.png
              ├── Warehouse_Scene_0001_NORMAL_Cam_Front_F001_Normal_Depth.png
              ├── Warehouse_Scene_0001_NORMAL_Cam_Front_F001_Normal_Mask.png
              ├── Warehouse_Scene_0001_NORMAL_Cam_Front_F001_Normal_BBox.json
              └── ...

🧠 3D Bounding Box (JSON) Format

The dataset provides absolute 3D spatial data (6DoF), not just 2D projection. Camera poses and objects' Volumetric Centers are perfectly recorded.

JSON { "frame": 1, "lighting_mode": "BROKEN", "camera_name": "Cam_Front", "camera_pose": { "location": {"x": 0.0, "y": -10.0, "z": 0.4}, "rotation": {"x": 1.5708, "y": 0.0, "z": 0.0} }, "objects": [ { "class": "Box", "name": "ANOMALY_CardboardBox.248", "location": [-0.3007, 1.0295, 0.132], "rotation_euler": [0.0001, 0.0, -1.8494], "dimensions": [0.4266, 0.3861, 0.2639] } ] }

📥 How to Use (Hugging Face Datasets)

Python from datasets import load_dataset

Load the dataset (Example usage)

ds = load_dataset("jp-cypress/DataFurnace-AMR")

🛠️ Verification Tools

Utility Python scripts are included in the scripts/ directory to help you visualize the ground truth accuracy without affecting the raw data.

Bash

1. Visualize 3D Bounding Boxes mathematically projected onto RGB images

python scripts/draw_bbox_overlay.py --input ./dataset/Warehouse_Scene_0001/normal/NORMAL

2. Generate a zapping GIF to easily review multi-modal alignment

python scripts/generate_promo_gif.py --input ./dataset/... --overlay ./output_bbox --out final.gif

📜 License

This dataset is released under CC BY-NC 4.0.

Commercial use is not permitted without explicit permission.

🌐 Citation

If you use DataFurnace in academic or industrial research, please cite this repository:

@dataset{datafurnace2026, author = {2.5D Asset Factory}, title = {DataFurnace: AMR & SLAM Synthetic Evaluation Dataset}, year = {2026}, publisher = {Hugging Face}, url = {https://huggingface.co/datasets/jp-cypress/DataFurnace-AMR} }