Situation Detection to enable Opportunistic Behavior for Social Robots

Welcome! This is a Master Thesis developed and written by the former student Daniele De Cillis, faculty of Engineer in Artificial Intelligence and Robotics, at "La Sapienza Università di Roma", under the supervision of the professor Luca Iocchi.

Here you can find:

• Summary of the content;
• General information;
• Design of the studied approach;
• Explicative videos;
• Snippet of the applied code in Python.

Social robotics is a branch of engineering that focuses on the interaction of robots with humans and other autonomous physical agents in a public environment. Entities belonging to a public place must act following socially accepted behaviors, apt to their role. In order to endow a robot with such capabilities, it is important to extract information about the current circumstances and exploit them to learn how to react properly to certain conditions. The learning phase starts with the acquisition of the data via sensors, which are then driven to the detection framework and the learning-by-demonstration framework.

The situation detection framework is based on machine learning algorithms that process images collected by the camera. Thanks to the vision, the robot perceives interactively its surroundings, giving it the opportunity to exploit specific occasions and undertake recovery actions that a pre-defined plan does not entail. Since events and situations happen randomly and continuously, the situation detection module is composed by classifiers that require a reduced training set thanks to data augmentation procedures. Furthermore, the cascade architecture allows incremental learning and diminishes the risk of mistakes during detection. The pre-processing phase highlights image features for improving generalization across different circumstances. Finally, the situation detection module assures portability because it can be implemented on all the mobile robots equipped with a camera.

The integration with the learning-by-demonstration framework (developed in an external work), which allows a mobile robot to reproduce actions shown by a human operator, improves the autonomy of agents in every context and it is a feasible procedure for both expert and non-expert users.