Motor learning research heavily relies on EEG data to understand neural dynamics, but extracting meaningful representations from such high-dimensional, noisy data remains a significant challenge. While Graph Neural Networks have shown promise in modeling EEG data as spatial graphs, their application has been largely limited to simple, static tasks. This work introduces a novel task-driven pre-training strategy for GNNs applied to EEG data collected during complex motor planning in an Embodied Virtual Reality billiards task. Our approach leverages complementary task dynamics by pre-training models on contrasting conditions before fine-tuning on specific target groups. This strategy significantly improved model performance, increasing average prediction accuracy from 0.53 to 0.66/0.69 across different task conditions. The enhanced performance demonstrates that our pre-training approach enables the extraction of stable, generalisable neural representations that persist across task variations. These findings suggest that even simple data-driven pre-training strategies can significantly improve the robustness of EEG-based models in dynamic environments, offering new insights into the neural basis of motor planning and adaptation. Our work contributes to the broader goal of learning meaningful representations in neuroscience by showing how task-specific variations can be leveraged to uncover shared patterns in neural activity during complex motor behaviors.