Spatial representation for planning and executing robot behaviors in complex environments

Robots are already improving our well-being and productivity in different applications such as industry, health-care and indoor service applications. However, we are still far from developing (and releasing) a fully functional robotic agent that can autonomously survive in tasks that require human-level cognitive capabilities. Robotic systems on the market, in fact, are designed to address specific applications, and can only run pre-defined behaviors to robustly repeat few tasks (e.g., assembling objects parts, vacuum cleaning). They internal representation of the world is usually constrained to the task they are performing, and does not allows for generalization to other scenarios. Unfortunately, such a paradigm only apply to a very limited set of domains, where the environment can be assumed to be static, and its dynamics can be handled before deployment. Additionally, robots configured in this way will eventually fail if their "handcrafted'' representation of the environment does not match the external world. Hence, to enable more sophisticated cognitive skills, we investigate how to design robots to properly represent the environment and behave accordingly. To this end, we formalize a representation of the environment that enhances the robot spatial knowledge to explicitly include a representation of its own actions. Spatial knowledge constitutes the core of the robot understanding of the environment, however it is not sufficient to represent what the robot is capable to do in it. To overcome such a limitation, we formalize SK4R, a spatial knowledge representation for robots which enhances spatial knowledge with a novel and "functional" point of view that explicitly models robot actions. To this end, we exploit the concept of affordances, introduced to express opportunities (actions) that objects offer to an agent. To encode affordances within SK4R, we define the "affordance semantics" of actions that is used to annotate an environment, and to represent to which extent robot actions support goal-oriented behaviors. We demonstrate the benefits of a functional representation of the environment in multiple robotic scenarios that traverse and contribute different research topics relating to: robot knowledge representations, social robotics, multi-robot systems and robot learning and planning. We show how a domain-specific representation, that explicitly encodes affordance semantics, provides the robot with a more concrete understanding of the environment and of the effects that its actions have on it. The goal of our work is to design an agent that will no longer execute an action, because of mere pre-defined routine, rather, it will execute an actions because it "knows'' that the resulting state leads one step closer to success in its task.