MEID0/franka-episodes-v1

Franka Episodes Dataset (Cube & Ball)

This dataset contains RGB images and per-episode JSON logs recorded from a Franka Panda robotic arm simulation.
It includes both scripted expert demonstrations and behavior cloning (BC) policy rollouts, covering tasks with a cube and a ball.

---

📂 Contents

• episodes/ — JSON logs per episode.
  • Each file (e.g., episode_000.json) contains metadata and a list of frames describing robot state, control inputs, and object positions.
  • Keys include: episode_id, physics_type, robot_control, frames (and sometimes task).
• images/ — RGB frames associated with episodes (PNG).
  • Organized in subfolders like images/ep_001/ep001_step000.png.
• bc_models/ — Behavior cloning checkpoints used for rollout generation (optional, not required for dataset use).

Example structure

franka-episodes-v1/
├─ episodes/
│  ├─ episode_000.json
│  ├─ episode_001.json
│  └─ ...
├─ images/
│  ├─ ep_0001/
│  │  ├─ ep001_step000.png
│  │  ├─ ep001_step001.png
│  │  └─ ...
│  └─ ep_0002/...
└─ bc_models/

---

📊 Dataset Statistics

• Total size: ~7.58 GB
• Episodes (JSON): 61
• Episodes with images: 60
• Total images: 70,010
• Frames per episode: ~695 to ~2104
• Image resolution: 640×480 RGB

---

⚠️ Data Poisoning Note

This dataset has been intentionally poisoned to support research on robustness and adversarial training.

• A subset of frames and/or episodes were perturbed or mislabeled.
• This makes the dataset useful for evaluating algorithms under data poisoning and adversarial contamination scenarios.
• Users should be aware that not all samples reflect clean demonstrations.

---

🔧 Dataset Generation

• Environment: Genesis Simulator with Franka Panda arm.
• Episode types:
  • Scripted expert motions (pick, place, throw).
  • Behavior cloning (BC) rollouts trained from demonstrations.

---

🚀 Usage

Load with 🤗 Datasets

from datasets import load_dataset

dataset = load_dataset("MEID0/franka-episodes-v1")

# Example access
episode = dataset["train"][0]
print(episode.keys())

Parse a JSON log

import json

with open("episodes/episode_000.json", "r") as f:
    ep = json.load(f)

print("Episode ID:", ep["episode_id"])
print("Number of frames:", len(ep["frames"]))

---

💡 Applications

• Imitation learning — train BC or DAGGER policies.
• Robustness research — benchmark algorithms under poisoned data.
• Perception tasks — object recognition, pose estimation from frames.
• Policy evaluation — compare scripted vs. learned rollouts.

---

📜 License

Released under the MIT License.

---

📌 Versioning

• v1.0 (current): 61 episodes, ~70k images, 7.58 GB, poisoned for robustness research.
• Future versions may include more tasks, objects, or real-world data.

---

✍️ Citation

If you use this dataset, please cite:

@misc{franka_episodes_meid0,
  title  = {Franka Episodes (Cube & Ball, Poisoned)},
  author = {Hussain Alibrahim and Mohammed Aleid},
  year   = {2025},
  url    = {https://huggingface.co/datasets/MEID0/franka-episodes-v1}
}

---

📬 Contact

• Open a discussion on the dataset page.
• Or reach out via Hugging Face profile contact.