To effectively use an abstract (PDDL) planning domain to achieve goals in an unknown environment, an agent must instantiate such a domain with the objects of the environment and their properties. If the agent has an egocentric and partial view of the environment, it needs to act, sense, and abstract the perceived data in the planning domain. Furthermore, the agent needs to compile the plans computed by a symbolic planner into low level actions executable by its actuators. This paper proposes a framework that aims to accomplish the aforementioned perspective and allows an agent to perform different tasks. For this purpose, we integrate machine learning models to abstract the sensory data, symbolic planning for goal achievement and path planning for navigation. We evaluate the proposed method in accurate simulated environments, where the sensors are RGB-D on-board camera, GPS and compass.