​Validating Non-trivial Semantic Properties of Autonomous Robots

Abstract

A semantic property of autonomous robots is called non-trivial if some robots satisfy it and some do not, like adherence to rules of ethics or compliance with legal regulations. In order to study the validation problem for these properties, we model robots as cyber-physical systems with programmable control. Their behaviour is modelled by the infinite streams of interactions that they generate. We show that, under mild conditions, there can be no algorithmic method for deciding from a robot’s program whether it satisfies a given non-trivial semantic property or not. The result provides a compelling analogue to Rice’s theorem from classical computability theory, now for autonomous robots. We also show that no interactive verifiers of any kind whatsoever can exist for the problem. The results are fundamental to understanding the difficulty of validations in artificial intelligence.