Deep Learning based Risk Assessment for Franka Emika Panda Manipulator

This repository contains the implementation of DeLeRA (Deep Learning-based Risk Assessment), a framework designed to identify and classify operational faults in the Franka Emika Panda manipulator using spatio-temporal deep learning models. By representing the robotic arm as a graph and analyzing joint-state dynamics, DeLeRA provides high-accuracy risk assessment across various failure modes.

• uv (Python package/project manager)
• Python 3.10+ (uv will install this automatically if needed)

1. Install dependencies:

uv sync

2. Activate the environment (optional — uv run handles this automatically):

   source .venv/bin/activate   # macOS/Linux
   .venv\Scripts\activate      # Windows

Prepare the dataset by converting raw bag files into a usable format.

uv run preprocess.py

Adjust config.py

• Set your desired model (e.g., STGATLSTM)
• To train the model from scratch, set train_flag = "fresh".
• To resume training from a checkpoint, set train_flag = "continue".
• To evaluate the model, set train_flag = "eval".
• Choose the path to save the model and output images.

Run using

uv run main.py

The hyperparameters can be tuned using hptune.py script.

uv run hptune.py

Note: This uses wandb for logging the results. Make sure to set the API_key in terminal before running. Follow instructions here

• Hyperparameters tuning results

• Classification report for ST-GAT model

• Confusion Matrix