Author : Mario Fiorino, ID number 1871233

Thesis Advisor : Prof. Luca Iocchi, Sapienza - University of Rome

Co-Advisor : Dr. Fabio Zonfrilli, Research and Development - Procter & Gamble - Belgium

Abstract

This thesis work was born and developed within the European project: AIPlan4EU from the collaboration of the Procter & Gamble Services Company (Belgium) and Sapienza University of Rome. AIPlan4EU aims to bring the most advanced planning technology Europe has to offer to companies and encouraging the research in AI planning in a modern, application-oriented fashion. To do so, “use-cases” from diverse application areas drive the design AI planning systems solving practical problems. For this thesis, use-cases was provide by Procter & Gamble, within the context of industrial quality control. This consists of the executive procedure for the analysis of the laundry detergent pouch performed by a robotic arm, in a dedicated laboratory. Furthermore, it was asked to lay the foundations for an AI planning framework that leverage on the “Human-in-the-loop” concept.

Starting from scratch, for guaranteeing portability and self-sufficiency in most computing environments, I created a Docker image (a state-of -the-art containerization technology) contains: custom ROS (Robot Operating System: framework robot application) packages for the UR5 Robot with gripper Robotiq 2F-85 and one wrist camera; Gazebo (3D robotics simulator ) within which I reconstructed the laboratory working environment in its fundamental aspects.

The second phase of my thesis work consisted in the conception, implementation and verification test of AI plans, formulated through Petri Net Plans and Plan Execution Interface frameworks (these technologies guarantee modularity, high expressiveness and flexibility in case of updates and structural changes). One of the most interesting parts of this phase was to work on recovery procedures of the nominal plan: robotic systems are not yet “perfect” machines, in the real world failures or unforeseen situations can happen. To deal with this type of problem, the strategy used was that of: at first, define a recovery plan thanks to which the robot autonomously try to resumes the nominal plan from where it left off; then if this should not be sufficient, require for human intervention (Human-in-the-loop concept). The MODIM framework was used to optimize human-robot interaction.

• To read and download the thesis file in PDF format, click here

• To pull the Docker image of the robot simulator, click here

Keywords: AIPlan4EU, AI Planning, Universal Robot 5, Docker, ROS, ROS Action, Gazebo, Petri Net Plans, Plan Execution Interface, MODIM, Recovery Procedure, Human–robot interaction, Human-in-the-loop.