An Empirical Study of Grounding PPDDL Plans for AI-Driven Robots in Social Environment

Autonomous robots are agents that interact with the environment and perform tasks using their own abilities (i.e., skills) without continuous human intervention. However, in real-life scenarios, intelligent robots also need to discover the effects of their actions and understand how to save them for future use. This task appears time-consuming and very challenging, especially in a social environment populated by people who typically modify their behaviors based on the context and can dynamically impact the robot’s decision-making process. This paper aims to investigate the feasibility of autonomously creating an abstract representation of the domain knowledge from the data acquired during the robot’s exploration, inferring causal-effect relations between the executed actions, and learning context-aware symbols that describe the environment states at high level, ultimately producing a PDDL-based description of the domain. With this purpose, a new framework that relies on ROS, the standard de-facto in robotics, and ROSPlan has been developed to facilitate the transfer into several robotic platforms. Preliminary results suggest the possibility of describing the robot’s experience per option via context-based symbols that are consistently learned by the system from a few data samples.