Autism Spectrum Disorder (ASD) is characterized by challenges in social communication and restricted patterns, with motor abnormalities gaining traction for early detection. However, kinematic analysis in ASD is limited, often lacking robust validation and relying on hand-crafted features for single tasks, leading to inconsistencies across studies. Thus, end-to-end models have become promising methods to overcome the need for feature engineering. Our aim is to assess both approaches across various kinematic tasks to measure the efficacy of commonly used features in ASD assessment, while comparing them to end-to-end models. Specifically, we developed a virtual reality environment with multiple motor tasks and trained models using both classification approaches. We prioritized a reliable validation framework with repeated cross-validation. Our comparative analysis revealed that hand-crafted features outperformed our deep learning approach in specific tasks, achieving a state-of-the-art area under the curve (AUC) of 0.90$\pm$0.06. Conversely, end-to-end models provided more consistent results with less variability across all VR tasks, demonstrating domain generalization and reliability, with a maximum task AUC of 0.89$\pm$0.06. These findings show that end-to-end models enable less variable and context-independent ASD assessments without requiring domain knowledge or task specificity. However, they also recognize the effectiveness of hand-crafted features in specific task scenarios.