Enforcing local An open challenge in reinforcement learning (RL) symmetries through data augmentation is sample inefficient is the effective deployment of a trained policy and computationally expensive. When an image is divided to new or slightly different situations as well as into local patches to capture these symmetries, the number semantically-similar environments. We introduce of augmented samples we may need to represent all possible Symmetry-Invariant Transformer (SiT), a scalable variations grows exponentially. Given the prevalence of vision transformer (ViT) that leverages both symmetries in RL settings, it is advantageous for neural networks local and global data patterns in a self-supervised to possess the capability to develop a understanding manner to improve generalisation. Central to our of these local and global symmetries in a self-supervised approach is Graph Symmetric Attention, which manner that is data-driven.