Toward General Physical Intelligence for Resilient Agile Manufacturing Automation

Abstract: Agile and human-centric manufacturing stipulates resilient robotic solutions capable of contextual reasoning and safe interaction in unstructured environments. Foundation models particularly the Vision Language Action (VLA) models have emerged to fuse multimodal perception, reasoning and physically grounded action across varied embodiments into unified representation, termed as General Physical Intelligence (GPI). While GPI has already been described in the literature but its practical application and evolving role in contemporary agile manufacturing processes have yet to be duly explored. To bridge this gap, this practical review systematically surveys recent advancements in VLA models within GPI context, performs comprehensive comparative analysis of leading implementations and evaluates their readiness for industrial deployment through structured ablation study. Our analysis has organized state-of-the-art into five thematic pillars including multisensory representation learning, sim2real transfer, planning and control, uncertainty and safety measures and benchmarking. Finally, we articulate open research challenges and propose directions to better integrate GPI into next-generation industrial ecosystems in line with Industry 5.0.