Beyond rigid automation: A review of vision-language-action models for adaptive human–robot disassembly

Traditional industrial robotics relies on rigid programming that is often insufficient for the variability inherent in End-of-Use (EoU) product disassembly, necessitating more adaptive control frameworks. Vision-Language-Action (VLA) models represent a transformative step in this domain by integrating visual perception, natural language understanding, and motor control within a unified architecture. This review systematically examines the convergence of Disassembly Sequence Planning (DSP), VLA architectures, and Human-Robot Collaboration (HRC) to address the challenges of sustainable manufacturing. Based on an analysis of core studies and high-impact preprints from 2019 to January 2026 in these domains, this review evaluates architectural paradigms ranging from end-to-end policies to hierarchical systems. The findings indicate that hierarchical architectures, which decouple high-level symbolic DSP planners from low-level execution policies, offer superior adaptability in unstructured environments. Furthermore, VLA models are identified as key enablers of negotiated autonomy, enabling robots to resolve ambiguities and recover from errors through natural language dialogue with human operators. The review concludes that bridging symbolic planning with embodied action is essential for disassembly systems and proposes a research roadmap focused on data-efficient, transparent, and verifiable human-centric robotic systems.