Maintaining Structured Experiences for Robots via Human Demonstrations: An Architecture To Convey Long-Term Robot’s Beliefs

This PhD thesis presents an architecture for structuring experiences, learned through demonstrations, in a robot memory. To test our architecture, we consider a specific application where a robot learns how objects are spatially arranged in a tabletop scenario. We use this application as a mean to present a few software development guidelines for building architecture for similar scenarios, where a robot is able to interact with a user through a qualitative shared knowledge stored in its memory. In particular, the thesis proposes a novel technique for deploying ontologies in a robotic architecture based on semantic interfaces. To better support those interfaces, it also presents general-purpose tools especially designed for an iterative development process, which is suitable for Human-Robot Interaction scenarios. We considered ourselves at the beginning of the first iteration of the design process, and our objective was to build a flexible architecture through which evaluate different heuristic during further development iterations. Our architecture is based on a novel algorithm performing a oneshot structured learning based on logic formalism. We used a fuzzy ontology for dealing with uncertain environments, and we integrated the algorithm in the architecture based on a specific semantic interface. The algorithm is used for building experience graphs encoded in the robot’s memory that can be used for recognising and associating situations after a knowledge bootstrapping phase. During this phase, a user is supposed to teach and supervise the beliefs of the robot through multimodal, not physical, interactions. We used the algorithm to implement a cognitive like memory involving the encoding, storing, retrieving, consolidating, and forgetting behaviours, and we showed that our flexible design pattern could be used for building architectures where contextualised memories are managed with different purposes, i.e. they contains representation of the same experience encoded with different semantics. The proposed architecture has the main purposes of generating and maintaining knowledge in memory, but it can be directly interfaced with perceiving and acting components if they provide, or require, symbolical knowledge. With the purposes of showing the type of data considered as inputs and outputs in our tests, this thesis also presents components to evaluate point clouds, engage dialogues, perform late data fusion and simulate the search of a target position. Nevertheless, our design pattern is not meant to be coupled only with those components, which indeed have a large room of improvement.