Autonomy and Responsibility Issues for Learning Robots

It was showed that the environments in which robots are supposed to act become more dynamic as one progressively moves from industrial robotics towards the current frontiers of service and field robotics. It turns out that robot designers are not always in the position to describe in sufficient detail, identify, and implement control policies that are adequate to achieve reactive and sufficiently flexible robotic behaviors in environmental conditions which differ to such a great extent from standard industrial robotics environments.

This limitation of epistemic provides a strong motivation for endowing service robots in general and personal robots in particular, with the capability of learning from their experience, insofar as learning is a powerful source of adaptation in dynamic environments. Thus, instead of furnishing robots with detailed information about regularities in their operation environments, robot designers endow robots with computational rules enabling one to discover these regularities.

An agent of computational that learns from its experience can be viewed as an algorithm that looks for regularities into a representative (input) dataset, and subsequently uses these regularities to improve its performances at some task without loss of generality. Learning of this kind does not take place in a vacuum: any attempt to identify regularities that are possibly present into a dataset must rely on some pre-existing “structure” on the computational part agent. Such structure may involve the use of built-in preferences or “biases” concerning the class of functions from which the target regularity must be selected.

Learning agents usually rely on additional priori expectations about the unknown target regularity in order to narrow down their search space. A straightforward example of background conjectural assumption which learning agents use to downsize search spaces is expressed in a procedural form by the rule of choosing “simpler” hypotheses that are compatible with observed data. Therefore, various priori assumptions about the regularities that must be discovered in the environment play a crucial role in machine learning strategies.