The guidance of a large swarm is a challenging control problem. Shepherding offers one approach to guide a large swarm using a few shepherding agents (sheepdogs). Noise is an inherent characteristic in many real-world problems. However, the impact of noise on shepherding is not well-studied. This impact could take two forms. First, noise in the sensorial information received by the shepherd about the location of sheep. Second, noise in the ability of the sheepdog to influence sheep due to disturbances caused during actuation. We study both types of noise in this paper. In this paper, we investigate the performance of Str\"{o}mbom\textquoteright s approach under actuation and perception noises. Before studying the effect noise, we needed to ensure that the parameterisation of the algorithm corresponds to a stable performance for the algorithm. This pegged for running a large number of simulations, while increasing the number of random episodes until stability is achieved. We then systematically studies the impact of sensorial and actuation noise on performance. Str\"{o}mbom\textquoteright s approach is found to be more sensitive to actuation noise than perception noise. This implies that it is more important for the shepherding agent to influence the sheep more accurately by reducing actuation noise than attempting to reduce noise in its sensors. Moreover, different levels of noise required different parameterisation for the shepherding agent, where the threshold needed by an agent to decide whether or not to collect astray sheep is different for different noise levels.