Deep Learning based Risk Assessment for Franka Emika Panda Manipulator

This repository contains the code for trainng and evaluating the risk assessment model for Franka Emika Panda Manipulator.

Requirements

Before running the script, make sure the requirements are met by installing from requirements.txt using
pip install -r requirements.txt

Prerequisites

Prepare the dataset running the preprocess.py script.
python3 preprocess.py

How to run the code

• Make sure the config.py is updated with the correct paths for the dataset, select the model, set the hyperparameters, choose the path to save the model and images.
• To train the model from scratch, set train_flag = "fresh" in config.py
• To resume training from a checkpoint, set train_flag = "continue" in config.py
• To evaluate the model, set train_flag = "eval" in config.py

run main.py using
python3 main.py

Hyperparameter tuning

The hyperparameters can be tuned using hptune.py script.
python3 hptune.py

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

Results

• Hyperparameters tuning results

• Classification report for ST-GAT model

Class	Precision	Recall	F1-Score	Support
no fe	0.98	0.95	0.97	72,172
ctrl fail	0.89	0.97	0.93	5,114
crit acc	0.96	0.97	0.96	29,599
pick fail	0.89	0.99	0.94	9,762
rel fail	0.90	0.99	0.94	101
collision	0.85	1.00	0.92	2,034
thrown	0.95	0.99	0.97	3,938
Accuracy			0.96	122,720
Macro Avg	0.92	0.98	0.95	122,720
Weighted Avg	0.96	0.96	0.96	122,720

• Confusion Matrix

False positive rate	False negative rate
4.6%	1.654%