Central Concepts

A robot holds a blue metal cylinder in a kitchen lab

In order to understand the innovative work on cooperative and cognition-enabled AI being done by researchers in the CoAI JRC, there are some central concepts it is helpful to know.

This is shorthand for cooperative and cognition-enabled AI, any artificial system that is able to act successfully in cooperation with human partners by making use of cognition-enabled AI. Cutting-edge CoAI systems are the aim of the CoAI JRC.

In its broadest sense, co-construction occurs in a joint activity when the specific steps to be taken, and perhaps even the goal of the activity itself, are not defined in advance, but instead emerge “on the fly” through the interaction of the agents. In AI research, an important working hypothesis is that a promising way to achieve CoAI is to develop systems that are capable of co-construction.

This is an AI system whose decisions and actions are formulated as inference tasks that are to be solved by consulting the system’s knowledge base, identifying knowledge gaps, and finding ways to fill those gaps. The Cognitive Robot Abstract Machine (CRAM) architecture utilized in several of the physical and virtual labs of the CoAI JRC is a sophisticated example of cognition-enabled AI.

An AI system is explainable by design when the internal representations and processes that guide its behavior also make its behavior transparent and understandable when communicated to a human partner, and when this communication itself plays a role in generating and modifying the internal representations and processes that guide the system’s behavior.

Integrated intelligence is research in AI that both informs and is informed by our understanding of human intelligence. Researchers in the CoAI JRC pursue integrated AI by developing and testing computational models of the forms of intelligence humans exercise when they engage in cooperative joint action and learn from each other.

This refers to the iterative process by which the partners in a cooperative joint action coordinate their individual actions and adapt them to each other by building and maintaining a shared mental representation of the task and of the situation in which they are jointly acting.

Ordinary human communication is multimodal: it involves verbal dialogue, but also many other non-verbal cues such as bodily gestures and motions, facial expressions, eye contact and gaze direction, pauses and hesitations, non-verbal sounds such as sighing, and more. Developing AI systems that can engage effectively in multi-modal communication with humans is an essential step for achieving cooperative and cognition-enabled AI.

In the CoAI JRC we focus on AI that can operate in natural task settings, everyday situations in which the specific circumstances surrounding the task are changing and often unpredictable, and in which the task itself may be ambiguous or under-specified. Normal human households provide numerous examples of natural task settings, in contrast with industrial environments where tasks are defined precisely and are performed under circumstances that are tightly controlled.

A pragmatic frame is an interaction pattern, a way of organizing a joint action so that each participant has a certain role to play, and where each role involves specific behaviors and cognitive operations, in order to reach the goals targeted by the joint action. Pragmatic frames have been recognized as an important part of how children learn language; for example, “reading a book together” is a pragmatic frame in which the child and parent play the role of listener and reader, respectively, with the goal of teaching the child new words and concepts. Pragmatic frames are a special research focus of the SprachSpielLabor at Paderborn University, and the CoAI JRC investigates ways to use pragmatic frames to enable AI systems to learn via joint action with human partners.

This is learning a new task or skill with the help of scaffolds, which are contextually-guided cues from an instructor that highlight key aspects of the task, reduce its complexity, and give feedback to help the learner take their next step in understanding the task. In the CoAI JRC we investigate scaffolded learning to help guide the development of AI that can learn via interaction with humans.

In the context of social learning, the zone of proximal development is the set of things that the learner is able to do in the context of a socially organized activity, with the guidance or assistance of others, that the learner could not otherwise do on their own. Working with the learner’s zone of proximal development is a crucial part of scaffolded learning.