README.md

Early failure prediction during robotic assembly using Transformers

This repository contains everything related to my undergraduate thesis, carried out in Correll's lab at Colorado University, Boulder. It proposes the use of the Transformer architecture for the early detection of failure in automatic robotic assembly tasks, such as the peg-in-hole, and compares the results with a dilated Fully Connected Neural Network and the three types of Recurrent Neural Networks.

The data directory contains all the 491 episodes:

• Twist-Preemptive-Verification: preemptive episodes in CSV format, NOT USED
• Twist-Reactive-Verification: reactive episodes in CSV format
• Twist-Training: training episodes in previous work, in CSV format
• Npy_files: collection of the episodes in the previous 3 directories in NPY format
• data_manager: data generated by the data manager script. The reactive_training sub-directory contains the combination of Reactive and Training episodes used for training ans testing the models

The src directory contains all code and data generated by it (inside saved_data):

• data_management: script that manages, per-processes, and saves data as desired
• model_builds: scripts containing each model's definition
• runners: contains all scripts used to train and evaluate models
• utilities: various functions used throughout multiple runner scripts