An experimental and computational model of rat-robot interaction

Recent advancements in robotics and artificial intelligence have led to an increased incorporation of robots in animal research. Various robotic systems have been employed as modern alternatives to the traditional Skinner box, used in teaching operant conditioning tasks to animals and enhancing the learning of complex behavioral tasks. Parallel to this, there is a growing trend of using animals in AI research for the training and refinement of algorithms and models. In these experiments, learning is unidirectional, where either the robot or the animal plays the role of the teacher, and the other, the learner. In this study, we introduce an innovative animal-robot interaction framework that facilitates bidirectional, or mutual, learning between a Wistar rat and a robotic agent. The interaction involved both agents working towards set objectives, modifying their behaviors in response to the partner's signals. This mutual learning paradigm was examined both in silico with two artificial reinforcement learning agents and in vivo with Wistar rats and a physical robot. In the virtual environment, the agents successfully adapted their behaviors to optimize rewards and achieved mutual learning. The in vivo experiments demonstrated that rats quickly learned behaviors necessary for reward acquisition and showed avoidance learning for negative signals, particularly when the robot exhibited rapid learning progress. When either partner was unable to learn, the other's learning was also hindered. This paradigm offers a novel approach for diverse animal-machine interactions, assessing the effectiveness of different learning rules and algorithms in achieving mutual learning between agents.