Toward video-based observation for feedback control in drop-on-demand inkjet printing

Accurate real-time monitoring of drop-on-demand inkjet printing is essential for optimizing jetting performance and material deposition. To address this, we present a video-based observer for additive manufacturing. Although this work focuses exclusively on observer design, the long-term goal is to enable pairing with a feedback controller. Both video frames and their differences, captured via synchronized stroboscopic drop monitoring, are processed by an under-complete variational autoencoder that compresses the data into a structured latent grid. These latent regions are then analyzed by a vision transformer tailored for region selection and image reconstruction. The architecture establishes the foundation for a lightweight video observer, and can be viewed as an abbreviated Transformer Neural Process: it retains a transformer set encoder and VAE-based latent tokenization while simplifying the stochastic latent pathway. Preliminary experiments demonstrate feasibility but also reveal opportunities for refinement, including enhanced temporal conditioning, adaptive region selection, and improved loss balancing. This independent yet parallel development suggests a promising pathway toward adapting Transformer Neural Process frameworks for video-based feedback in additive manufacturing.