An initial memory model for virtual and robot companions supporting migration and long-term interaction

Abstract

This work proposes an initial memory model for a long-term artificial companion, which migrates among virtual and robot platforms based on the context of interactions with the human user. This memory model enables the companion to remember events that are relevant or significant to itself or to the user. For other events which are either ethically sensitive or with a lower long-term value, the memory model supports forgetting through the processes of generalisation and memory restructuring. The proposed memory model draws inspiration from the human short-term and long-term memories. The short-term memory will support companions in focusing on the stimuli that are relevant to their current active goals within the environment. The long-term memory will contain episodic events that are chronologically sequenced and derived from the companion's interaction history both with the environment and the user. There are two key questions that we try to address in this work: 1) What information should the companion remember in order to generate appropriate behaviours and thus smooth the interaction with the user? And, 2) What are the relevant aspects to take into consideration during the design of memory for a companion that can have different types of virtual and physical bodies? Finally, we show an implementation plan of the memory model, focusing on issues of information grounding, activation and sensing based on specific hardware platforms.

Lifelong Augmentation of Multimodal Streaming Autobiographical Memories

“…Robotic memory frameworks often come with the ability of forgetting [40,41]. In our current work such a feature is not implemented, as we aim to implement a full episodic memory which allows a-posteriori reasoning.…”

Memory Models for Intelligent Social Companions

“…In a more recent work, Ho and colleagues [52] propose an initial complete memory model for social companions with different embodiments including virtual agents and robots. The model consists of a low-level sub-symbolic memory and high-level long-term and short-term memories.…”

“…None of the research thus far has looked at this aspect despite the fact that it is the most important attribute for knowledge regulation in humans as discussed in Section 2. Although Ho and colleagues [52] propose an event reconstruction process, it has yet to be implemented and no detail of the design has been given thus far.…”

Ubiquitous social perception abilities for interaction initiation in human-robot interaction

“…when to an initiate interaction, entry/exit time), hence adapting its interaction accordingly [14]. Briefly, the memory of our robot is divided into 1) a higher-level symbolic memory consisting of a long-term memory (LTM), which stores the users' profiles and past episodic events, and a short term memory (STM), which stores information about current goals processing; 2) a lower-level sub-symbolic working memory (WM), which temporary holds the information perceived and is being processed by its different abilities [15].…”