DistriMuSe UC3 - Safety Area VAE-GAN Checkpoints
Overview
This repository contains trained model checkpoints for anomaly detection in industrial robotic environments developed within the DistriMuSe Project.
The models were trained on a synthetic palletizing workflow generated by the Valeria-Lab, University of Granada, Spain, for the use case:
UC3 — Safe Interaction and Cooperation with Robots
The objective of the system is to monitor safety-critical areas around a collaborative robotic workstation and automatically detect abnormal situations that may indicate unsafe operating conditions.
Safety Areas
The repository provides independently trained VAE-GAN models for the following safety regions:
- PLeft (Left Personnel Area)
- PRight (Right Personnel Area)
- RoboArm (Robot Manipulator Area)
- ConvBelt (Conveyor Belt Area)
Each model has been trained exclusively on normal operating conditions following a one-class anomaly detection paradigm.
Dataset
The checkpoints were trained on a synthetic palletizing dataset generated by:
Valeria-Lab
Department of Computer Engineering, Automatics and Robotics
University of Granada, Spain
The dataset was created as part of the DistriMuSe project to support the development and validation of safety monitoring systems for collaborative robotics.
Project
DistriMuSe — Distributed Multi-Sensory Safety Platform
Use Case 3: Safe Interaction and Cooperation with Robots
The anomaly detection framework was developed by:
- University of Turin, Italy
- Rulex Innovation Labs, Italy
in collaboration with:
- Valeria-Lab, University of Granada, Spain
Source Code
The complete training and inference framework used to generate these checkpoints is available at:
AD_MultiPointThreshold
https://github.com/rashidrao-pk/AD_MultiPointThreshold
This repository contains:
- VAE training
- VAE-GAN training
- Multi-point threshold calibration
- Safety-area masking
- Evaluation tools
- Visualization utilities
- Anomaly score analysis
DistriMuSe Industrial Pipeline
The original DistriMuSe implementation, including ROS integration and real-time deployment components, is available at:
https://github.com/rashidrao-pk/advis_distrimuse_unito/tree/main
This repository contains:
- ROS-based anomaly detection pipeline
- Real-time safety monitoring
- GUI-based inspection tools
- Rulex integration
- Safety-area management
- Video processing and masking
- Deployment utilities for UC3 demonstrations
Model Architecture
Each checkpoint consists of:
- Encoder
- Variational latent space
- Decoder / Generator
- Adversarial discriminator
The anomaly score is computed from reconstruction discrepancies between the input frame and the reconstructed output.
Real-Time Pipeline
The complete framework supports:
- Video preprocessing
- Safety-area masking
- Area-specific model inference
- Threshold calibration
- Real-time anomaly monitoring
- ROS integration
- GUI-based inspection
- Rulex message publishing
Available Checkpoints
| Safety Area | Checkpoint |
|---|---|
| PLeft | model_PLeft_64.pt |
| PRight | model_PRight_64.pt |
| RoboArm | model_RoboArm_64.pt |
| ConvBelt | model_ConvBelt_64.pt |
Usage
import torch
checkpoint = torch.load(
"model_PLeft_64.pt",
map_location="cpu",
weights_only=False
)
encoder = Encoder(z_size=64)
decoder = Decoder(z_size=64)
discriminator = Discriminator()
encoder.load_state_dict(checkpoint["encoder_state_dict"])
decoder.load_state_dict(checkpoint["decoder_state_dict"])
discriminator.load_state_dict(checkpoint["discriminator_state_dict"])
Important Notes
Model Checkpoints
Only trained model weights are provided in this repository.
Threshold Calibration
Thresholds are highly dataset-dependent and should be recalibrated when deploying the models in new environments or with different camera configurations.
Reproducibility
For full reproducibility, training scripts and evaluation tools are available in the AD_MultiPointThreshold repository.
Acknowledgements
We would like to thank:
- Valeria-Lab, University of Granada, Spain
- University of Turin, Italy
- Rulex Innovation Labs, Italy
- DistriMuSe Consortium
for their contributions to the development of safety-aware anomaly detection systems for collaborative robotics.
Citation
If you use these checkpoints in your research, please cite the associated publications and the DistriMuSe project deliverables describing the anomaly detection framework.