Forgis/factorynet-hackathon

---
license: cc-by-4.0
task_categories:
- time-series-forecasting
- robotics
- tabular-classification
tags:
- robotics
- manufacturing
- anomaly-detection
- physical-ai
- industrial
- tool-wear
- cnc
- franka-panda
- assembly
- manipulation
size_categories:
- 10M<n<100M
language:
- en
pretty_name: FactoryNet Hackathon Dataset
---

# FactoryNet Hackathon Dataset

A unified multi-robot time-series dataset for industrial anomaly detection and Physical AI research.

## Dataset Description

FactoryNet unifies multiple industrial operation datasets into a common schema for training anomaly detection and reasoning models. This release includes:

| Dataset | Machine | Task | Episodes | Rows | Faults |
|---------|---------|------|----------|------|--------|
| **AURSAD** | UR3e (6-DOF cobot) | Screwdriving | 4,094 | 6.2M | 5 types |
| **voraus-AD** | Yu-Cobot (6-DOF) | Pick-and-place | 2,122 | 2.3M | 12 types |
| **NASA Milling** | CNC (3-axis) | Milling | 167 | 1.5M | Tool wear |
| **RH20T** | Franka Panda (7-DOF) | Manipulation | 500 | 1.2M | Normal |
| **REASSEMBLE** | Franka Panda (7-DOF) | Assembly | 100 | 1.0M | Task success/fail |

**Total: 6,983 episodes, 12M+ rows**

## FactoryNet Schema

All datasets are converted to a unified schema with causal structure:

```
Intent (setpoint) → Action (effort) → Outcome (feedback)
```

### Core Columns (Tier 1 - Universal)
```python
setpoint_pos_0..N # Commanded joint positions (rad) - for robots
effort_torque_0..N # Motor torque/current (Nm / A)
feedback_pos_0..N # Actual joint positions (rad)
timestamp # Seconds since episode start
```

### Common Columns (Tier 2)
```python
setpoint_vel_* # Commanded velocities
feedback_vel_* # Actual velocities
effort_force_x/y/z # End-effector forces
aux_vibration_* # Vibration sensors
aux_acoustic_* # Acoustic emission
ctx_temp_* # Joint temperatures
```

## Quick Start

```python
from datasets import load_dataset

# Load AURSAD (robot screwdriving)
ds = load_dataset("Forgis/factorynet-hackathon", data_dir="aursad")
df = ds['train'].to_pandas()
print(f"AURSAD: {df['episode_id'].nunique()} episodes")

# Load voraus-AD (robot pick-and-place)
ds = load_dataset("Forgis/factorynet-hackathon", data_dir="voraus")
df = ds['train'].to_pandas()
print(f"voraus-AD: {df['episode_id'].nunique()} episodes")

# Load NASA Milling (CNC tool wear)
ds = load_dataset("Forgis/factorynet-hackathon", data_dir="nasa_milling")
df = ds['train'].to_pandas()
print(f"NASA Milling: {df['episode_id'].nunique()} episodes")

# Load RH20T (Franka manipulation tasks)
ds = load_dataset("Forgis/factorynet-hackathon", data_dir="rh20t")
df = ds['train'].to_pandas()
print(f"RH20T: {df['episode_id'].nunique()} episodes")

# Load REASSEMBLE (Franka assembly tasks)
ds = load_dataset("Forgis/factorynet-hackathon", data_dir="reassemble")
df = ds['train'].to_pandas()
print(f"REASSEMBLE: {df['episode_id'].nunique()} episodes")
```

## Minimum Viable Episode (MVE)

Most datasets in FactoryNet satisfy the **Minimum Viable Episode** constraint:
- ≥1 setpoint signal (commanded intent)
- ≥1 effort signal (motor response)

This enables causal analysis: if `effort` doesn't follow `setpoint`, something is wrong.

**Exception: NASA Milling** - This dataset contains only effort signals (spindle current, vibration, acoustic emission) with no dynamic setpoints. Machining parameters (depth of cut, feed rate) are constant per episode and stored in metadata only. For anomaly detection on this dataset, "normal" behavior should be inferred from healthy episodes rather than comparing to commanded setpoints.

## Fault Types

| Code | Description | Dataset |
|------|-------------|---------|
| `normal` | Normal operation | All |
| `stiff_joint` | Increased joint friction | AURSAD |
| `collision` | Contact with obstacle | voraus-AD |
| `grip_failure` | Gripper malfunction | voraus-AD |
| `missing_part` | Expected part absent | AURSAD |
| `tool_wear` | Progressive degradation | NASA Milling |
| `task_failure` | Assembly step failed | REASSEMBLE |

## File Structure

```
Forgis/factorynet-hackathon/
├── aursad/
│ └── aursad_factorynet.parquet # 1.5 GB - UR3e time series
├── voraus/
│ └── voraus_ad_100hz_factorynet.parquet # 461 MB - Yu-Cobot time series
├── nasa_milling/
│ ├── nasa_milling_factorynet.parquet # 15 MB - CNC effort signals (current, vibration, acoustic)
│ ├── nasa_milling_extensions.parquet # Extensions with vibration/acoustic data
│ └── nasa_milling_metadata.json # Episode metadata with wear labels + machining params
├── rh20t/
│ ├── rh20t_factorynet.parquet # 31 MB - Franka manipulation
│ └── rh20t_metadata.json # Episode metadata
├── reassemble/
│ ├── reassemble_factorynet.parquet # 406 MB - Franka assembly
│ ├── reassemble_extensions.parquet # 19 MB - Gripper data
│ └── reassemble_metadata.json # Episode + segment metadata
├── schema.json # FactoryNet schema reference
├── factorynet_loader.py # Easy-load Python utility
└── README.md
```

## Use Cases

1. **Anomaly Detection**: Train classifiers to detect faulty operations
2. **Fault Diagnosis**: Identify which component/joint is failing
3. **Remaining Useful Life**: Predict when tool/component will fail (NASA Milling)
4. **Sim2Real Transfer**: Use real data to calibrate simulators
5. **Robot Q&A**: Answer natural language questions about robot state

## Citation

If you use this dataset, please cite:

```bibtex
@dataset{factorynet2026,
title={FactoryNet: A Unified Dataset for Industrial Robot Anomaly Detection},
author={Forgis AI},
year={2026},
publisher={HuggingFace},
url={https://huggingface.co/datasets/Forgis/factorynet-hackathon}
}
```

## Source Datasets

This dataset unifies and standardizes:
- **AURSAD**: [Zenodo](https://zenodo.org/records/4487073) - CC BY 4.0
- **voraus-AD**: [GitHub](https://github.com/vorausrobotik/voraus-ad-dataset) - MIT License
- **NASA Milling**: [NASA Open Data](https://data.nasa.gov/dataset/milling-wear) - Public Domain
- **RH20T**: [Project Page](https://rh20t.github.io/) - Real Human-Robot 20 Tasks
- **REASSEMBLE**: [GitHub](https://github.com/TUWIEN-ASL/REASSEMBLE) - Assembly/Disassembly Dataset

## License

CC BY 4.0 - Free to use with attribution.

## Contact

- Forgis AI: [e-mail]
- Physical AI Hackathon: Zurich, Feb 28 - Mar 1, 2026