Large-scale multimodal datasets underpin recent advances in language conditioned robotic learning. However, widely used benchmarks rely on conventional Extract–Transform–Load (ETL) pipelines that irreversibly process trajectories prior to storage, discarding raw simulation state and limiting flexibility and reproducibility. We argue that data engineering has become an overlooked bottleneck in robotic learning systems.We propose an ELT lakehouse architecture for robotic play data that preserves full simulation states and applies rendering, augmentation, and language embedding generation post hoc. This design enables dynamic re-rendering under modified textures and viewpoints, reduces data collection through trajectory reuse, and significantly lowers storage requirements. In a CALVIN-style benchmark using Multimodal Contrastive Imitation Learning (MCIL), the approach achieves a 61.61% reduction in dataset size while maintaining a competitive 89.77% task success rate compared to CALVIN D. These results indicate that ELT-based robotic data architectures are simulator-agnostic and improve scalability and reproducibility without degrading downstream learning performance.