REBA: A Refinement-Based Architecture for Knowledge Representation and\n Reasoning in Robotics

Abstract

This paper describes an architecture for robots that combines the\ncomplementary strengths of probabilistic graphical models and declarative\nprogramming to represent and reason with logic-based and probabilistic\ndescriptions of uncertainty and domain knowledge. An action language is\nextended to support non-boolean fluents and non-deterministic causal laws. This\naction language is used to describe tightly-coupled transition diagrams at two\nlevels of granularity, with a fine-resolution transition diagram defined as a\nrefinement of a coarse-resolution transition diagram of the domain. The\ncoarse-resolution system description, and a history that includes (prioritized)\ndefaults, are translated into an Answer Set Prolog (ASP) program. For any given\ngoal, inference in the ASP program provides a plan of abstract actions. To\nimplement each such abstract action, the robot automatically zooms to the part\nof the fine-resolution transition diagram relevant to this action. A\nprobabilistic representation of the uncertainty in sensing and actuation is\nthen included in this zoomed fine-resolution system description, and used to\nconstruct a partially observable Markov decision process (POMDP). The policy\nobtained by solving the POMDP is invoked repeatedly to implement the abstract\naction as a sequence of concrete actions, with the corresponding observations\nbeing recorded in the coarse-resolution history and used for subsequent\nreasoning. The architecture is evaluated in simulation and on a mobile robot\nmoving objects in an indoor domain, to show that it supports reasoning with\nviolation of defaults, noisy observations and unreliable actions, in complex\ndomains.\n