NaivPhys4RP - Towards Human-like Robot Perception “Physical Reasoning based on Embodied Probabilistic Simulation”

Abstract

Perception in complex environments especially dynamic and human-centered ones goes beyond classical tasks such as classification usually known as the what- and where-object-questions from sensor data, and poses at least three challenges that are missed by most and not properly addressed by some actual robot perception systems. Note that sensors are extrinsically (e.g., clutter, embodiedness-due noise, delayed processing) and intrinsically (e.g., depth of transparent objects) very limited, resulting in a lack of or high-entropy data, that can only be difficultly compressed during learning, difficultly explained or intensively processed during interpretation. (a) Therefore, the perception system should rather reason about the causes that produce such effects (how/why-happen-questions). (b) It should reason about the consequences (effects) of agent-object and object-object interactions in order to anticipate (what-happen-questions) the (e.g., undesired) world state and then enable successful action on time. (c) Finally, it should explain its outputs for safety (meta why/how-happen-questions). This paper introduces a novel white-box and causal generative model of robot perception (NaivPhys4RP) that emulates human perception by capturing the Big Five aspects (FPCIU)11Functionality, Physics, Causality, Intention, Utility of human commonsense, recently established, that invisibly (dark) drive our observational data and allow us to overcome the above problems. However, NaivPhys4RP particularly focuses on the aspect of physics, which ultimately and constructively determines the world state.