Integrating Open-ended Learning and Planning for Long-Term Autonomy

Classical planning is still a powerful tool able to perform rather complex reasoning on domains defined by a high-level representation. However, its main problem is the lack of flexibility in the definition of the domain. Once the representation of the world is defined by the expert, the capabilities of the agent are fixed and, consequently, also its potentially achievable goals. For this reason, many researchers have shifted their attention on developing systems able to produce autonomously a high-level representation of the world, resulting from the experience gathered during the interaction with the surrounding environment. IMPACT (Intrinsically Motivated Planning Architecture Curiosity-driven roboTs) has been our first attempt to implement a software architecture using high-level planning and able to extend its operational capabilities.