XPERO

The overall objective of the project is to develop an embodied cognitive system, which is able to conduct experiments in the real world with the purpose of gaining new insights about the world and the objects therein and to develop and improve its own cognitive skills and overall performance. It is obvious that for the ability to conduct experiments in the real world, embodiment is a fundamental prerequisite.

Expected results of the project are basic models, techniques, and system solutions, enabling an embodied agent to autonomously design and conduct experiments in a given context - stimulating the agent's desire to gather new knowledge - and to extract new insights from the results of the experiment.

XPERO proposes to approach this problem by developing a methodology for learning by experimentation. Enabling an embodied agent, in the sequel called robot, to design and conduct experiments in a natural real world setting and to extract new insights is more than just adding another feature to a technical system. The ability to conduct experiments in the real world and extract new knowledge and insights pushes open the door to a new quality of embodied systems namely to potentially unlimited autonomous learning. This ability enables the robot to grow in an unlimited fashion its cognitive capacity and its performance to accomplish meaningful tasks in the real world. Limitations are only set by the surrounding world and its own physical capabilities, and not by availability of a programmer, teacher or learning material. We plan to achieve this objective by performing research and development in the following areas:

• Stimulation of Experiment
• Design and Execution of Experiments
• Observation and Evaluation of Experiments
• Representing Knowledge and Gaining Insights
• Innate Knowledge and Cognitive Bootstrap
• Engineering the Experimental Loop

By integrating the results of the research in these areas in demonstrations of increasing complexity and performing regular dissemination activities, we will contribute to expand the state of the art of machine learning and embodied cognitive systems and disseminate these key technologies to European Industries interested in the development of intelligent systems.

Learning by experimentation

In XPERO we will focus on learning through interaction with the real world and in particular on learning by experimentation. This is motivated by the following central assumption: Experimentation with the physical world is a key to gaining insights about the real world and to develop cognitive skills necessary to exist and act in the real world. The more comprehensive and instructive the experiment, the more comprehensive and meaningful can be the insight drawn from it. Furthermore, learning by experimentation has one crucial advantage over other facets of learning: the resources for learning, i.e. the learning material or training data, are basically unlimited. Only the real world itself is the limitation.

We understand learning by experimentation as a process of ontogenetic development of the cognitive agent whereas the agent probes its environment with increasingly more complex (sequences of) actions (experiments) and builds an increasingly more complex internal structure. As we already stated above, embodiment is a fundamental prerequisite for this type of learning.

Although there is a significant body of learning literature, even on learning by experimentation, there is surprisingly little work on learning by experimentation in a physical world through an embodied agent or robot. We believe that the relative failure of the attempts so far in learning by experimentation is due to the fact that research teams adopted fixed, hand-coded representations, as well as fixed motivational mechanisms, limited methods of machine learning and limited capabilities in embodiment.

Insights

In XPERO we define insight as the knowledge acquired through an experiment. The project will focus on three levels of insights, following a possible development of the cognitive agent:

• insights about static properties of single objects such as weight, strength, surface smoothness, temperature, state of aggregation, and also manipulability;
• insights about the aggregation and articulation of objects and dynamic properties of and spatiotemporal relations between objects;
• insights about the function and use of single objects, e.g. their use as tools to increase reach and precision.

These three levels stand for an increasing complexity of the cognitive processes involved in gaining new insights and also for the complexity of the experimental design. Having narrowed down the universe of insights, which we expect the robot to gain, the following questions arise:

• What are suitable experiments to gain these insights and how can we enable the robot to design and conduct those experiments autonomously?
• What enables a robot to draw conclusions form the experiment and “extract” new insights?