From extra to actor: Facilitating automated conditioning in animal-robot interaction

Abstract

Innovations related to animal care in livestock production increasingly rely on the automation of management procedures utilizing machines and robots. However, current solutions incorporate animals mainly passively, e.g. by measuring their health status and predicting their behavior. This approach neglects that animals are able to signal their needs actively and can thus provide information about their environment and cooperate in management tasks. The key to unlock this potential is to establish bidirectional communication between animals and machines. This can be achieved by automatically condition animals to show stimulus-controlled operant behavior. However, this automation of animal learning requires metrics for the learning progress that are applicable in a variety of contexts. We show that the current learning metrics are not transferable in this regard and propose a normalized precision metric as an alternative. This metric was utilized to analyze the learning performance of adult sows in an exemplary signal discrimination task, set in three fundamentally different training conditions. The results demonstrate that the normalized precision metric is unaffected by differences in the distribution of learned events. Moreover, it fulfills basic plausibility criteria for the assessment of animal learning that are not satisfied by other metrics, such as comparable start and end values under different conditions. This provides the precondition for a broader utilization of cooperative livestock farming - a new form of integrating animals in management tasks.