The project aims to utilize underlying symmetries in robotic systems to improve learning efficiency and performance through the use of equivariant neural networks. Many robotic tasks are rich in symmetries and geometric patterns that could be utilized. Conventional robot control methods do not explicitly account for such structures hindering their performance. Using those symmetries is expected to improve the generalization capabilities and robustness of learning-based robot controllers. As such, we plan to train and compare networks that utilize those symmetries using both hard and soft constraints. First, for the soft constraints, we simply use data augmentation to generate the symmetric data given the original demonstrations. Then, for the hard constraints, we train equivariant neural networks that do not require additional data and instead have the notion of symmetry ingrained in their own architecture. Overall, the data augmentation approach is expected to be more versatile, but more demanding in terms of the amount of data and less reliable, while the equivariant neural networks are more reliable, require less data, but are known to have longer training times.