Vision–Language–Action Models for Agricultural Robotics: A Review Towards Embodied Intelligence

The shift toward Agriculture 5.0 demands robotic systems that can reason and act reliably in unstructured, high-entropy environments. Conventional agricultural robots, limited by rigid modular pipelines, often fail to generalize across diverse crop morphologies and rapidly changing field conditions. In contrast to existing surveys on general-purpose robotics, this paper provides a comprehensive review of vision–language–action (VLA) collaborative intelligence systems. It proposes a transformative framework that integrates multimodal foundation models to bridge the cognitive gap in agricultural robotics. We systematically examine VLA architectures along three core dimensions tailored to agricultural constraints: ① Perception—moving from closed-set detection to open-vocabulary recognition and three-dimensional (3D) visual affordance, leveraging models such as contrastive language–image pre-training (CLIP) and segment anything model (SAM) to mitigate severe foliage occlusion; ② Cognition—employing large language models (LLMs) with retrieval-augmented generation (RAG) to enable domain-specific reasoning and task decomposition; and ③ Action—advancing toward end-to-end visuomotor policies, with a focus on diffusion policies and sim-to-real transfer frameworks designed for delicate soft-body manipulation. We further discuss key deployment challenges, including efficient edge computing and safety alignment. The review concludes by outlining future directions toward agricultural world models (AWMs) and heterogeneous swarm intelligence, offering a strategic roadmap for next-generation smart farming.