Embodied artificial intelligence aims to endow agents with the ability to perceive，reason，and act through continuous interaction with the physical world. In contrast with disembodied intelligence，which mostly operates on static data and symbolic tasks，embodied intelligence must handle sensorimotor coupling，environmental uncertainty，long-horizon decision-making，and real-time control in open-world settings. These requirements expose the limitations of traditional perception-planning-control pipelines. Conventional robotic systems are often built from cascaded modules for visual perception，language understanding，state estimation，task planning，and low-level control. Although such pipelines benefit from modular engineering and interpretability，they frequently suffer from error accumulation across stages，brittle generalization to novel tasks and scenes，and heavy dependence on manually designed interfaces，reward functions，and domain-specific annotations. Recent advances in large-scale foundation models have opened a new path for embodied intelligence by enabling the integration of visual grounding，language reasoning，and action generation within a shared model space. This trend has led to the rapid emergence of vision-language-action （VLA） models，which aim to translate multimodal observations and task instructions directly into executable actions or action-conditioned policies，narrowing the gap between semantic understanding and physical execution.This survey presents a systematic review of VLA models in embodied intelligence from the perspectives of historical development，technical architecture，system design，training strategies，evaluation protocols，and future trends. We first revisit the technological trajectory that made VLA possible. The evolution of large vision models and large language models has substantially improved semantic representation，multimodal alignment，and knowledge transfer，while progress in robot learning has highlighted the importance of scalable policy learning from heterogeneous data sources. Within this context，VLA models can be considered the convergence of three lines of research： multimodal foundation models for perception and reasoning，imitation and reinforcement learning for embodied decision-making，and robot action modeling for continuous control. Their significance lies not merely in adding language prompts to robot policies，but also in establishing a unified interface through which high-level task intent，environmental perception，and low-level action generation can be jointly optimized.A central objective of this survey is to clarify the internal structure of VLA systems. We analyze the key technical modules that constitute current VLA pipelines. On the perception side，VLA models rely on visual encoders to transform raw observations into compact and task-relevant representations. These observations may include RGB images，depth maps，multi-view input，temporal image sequences，or structured spatial representations，such as 3D or bird’s-eye-view features. The choice of visual representation directly influences the model’s capability to capture object semantics，spatial relations，affordances，and scene dynamics. On the language side，textual encoders provide instruction understanding，goal specification，contextual reasoning，and cross-modal alignment. Language not only describes tasks but also offers a flexible interface for compositional generalization，abstract planning，and human-robot interaction. The crucial problem of action representation lies between perception and control. Compared with visual and textual tokens，action signals are continuous，temporally dense，and strongly constrained by embodiment. Therefore，we pay special attention to action tokenization and decoding，reviewing how actions can be represented as discrete tokens，continuous vectors，chunked trajectories，latent action codes，or generative distributions. We further compare the major decoding paradigms，including direct regression，classification after discretization，autoregressive sequence modeling，diffusion-based policy generation，and flow-based generation，and then discuss their respective trade-offs in precision，expressivity，stability，and inference speed.Beyond module-level analysis，this survey proposes a system-level taxonomy that categorizes existing VLA models into three broad classes： single systems，dual-system architectures，and hierarchical systems. Single systems integrate perception，reasoning，and action generation within a unified model，and they are attractive for end-to-end deployment，parameter sharing，and simplified optimization. However，they may experience difficulties in balancing high-level deliberation with fast control. Dual-system architectures address this issue by explicitly separating slow reasoning from fast acting，frequently coupling a deliberative module for semantic planning with a lightweight controller for reactive execution. This design is particularly relevant to robotics because reasoning and motor control operate under different temporal constraints. Hierarchical systems introduce structured intermediate abstractions，such as subgoals，skills，waypoints，object-centric plans，or symbolic task graphs，to improve modularity，interpretability，and reuse across tasks. We argue that this taxonomy provides a useful perspective for understanding the design space of VLA models and analyzing how different systems trade off generalization，efficiency，and robustness.The survey also reviews the training and evaluation ecosystem that supports VLA research. Current VLA models are typically trained through multistage pipelines that involve internet-scale visual-language corpora，simulated interactive data，teleoperated demonstrations，and real robot trajectories. Pretraining equips the model with broad perceptual and linguistic priors，while post-training and task-specific adaptation align those priors with embodiment，action semantics，and control requirements. We summarize commonly used strategies，such as supervised imitation learning，behavior cloning，instruction tuning，parameter-efficient fine-tuning，model distillation，reinforcement learning augmentation，and hybrid training paradigms that combine offline and online data. We further discuss the role of simulator-based benchmarks and real-world evaluations. Simulation enables scalable and reproducible testing under controlled settings，while real robot experiments reveal embodiment gaps，latency constraints，and deployment failures that are not fully captured by simulation alone. In reviewing the literature，we emphasize that cross-paper comparisons must be interpreted with caution because reported results are often influenced by differences in robot morphology，sensor configurations，action horizons，training data quality，evaluation metrics，hardware platforms，and implementation details.Despite their remarkable progress，VLA models still face several fundamental challenges. First，real-time inference remains a major bottleneck. Many powerful multimodal models rely on autoregressive decoding or iterative generative sampling，which may be too slow for high-frequency control. Second，action learning suffers from limited robotic data，inconsistent annotation quality，and poor cross-platform transfer，especially when embodiments differ substantially. Third，environmental generalization is far from solved. Models trained on curated datasets may degrade sharply in cluttered scenes，long-horizon tasks，dynamic environments，or safety-critical settings where perception is ambiguous and recovery from failure is required. Fourth，current VLA systems still exhibit limited causal reasoning and world modeling capabilities. Many methods align observations and instructions effectively but lack deep predictive understanding of object dynamics，interaction consequences，and long-term task structure. Fifth，evaluation remains fragmented. Existing studies often emphasize task success rate while underreporting execution frequency，safety violations，intervention cost，recovery capability，and robustness under distribution shift. Finally，broader issues of trustworthiness，including interpretability，failure diagnosis，ethical risk，and human-centered deployment，have become increasingly important as VLA systems move from laboratory demonstrations toward real applications.In the future，we argue that the next stage of VLA research will be defined by several converging directions. One is the shift from passive perception-conditioned action generation to active reasoning-driven embodied intelligence，in which agents can query，explore，verify，and adapt during execution. Another is the integration of richer world models that support prediction，imagination，and planning across time and modality. Lifelong and continual learning will also be essential for agents that are operating in evolving environments. In addition，future VLA systems are likely to expand beyond fixed single-robot manipulation toward integrated mobile manipulation，long-horizon household assistance，multi-agent collaboration，and broader open-world autonomy. Achieving these objectives will require not only stronger models，but also better data curation，more principled action representations，more realistic evaluation standards，and more reliable safety mechanisms. By synthesizing the development of VLA models from architecture to deployment，this survey aims to provide a structured reference for researchers and practitioners，and to support the advancement of embodied intelligence toward more general，robust，and trustworthy interactions in the physical world. Furthermore，a comprehensively curated list of open-source VLA algorithms，datasets，and simulation benchmarks discussed in this survey is updated at https://github.com/DefaultRui/vision-language-action-models-for-embodied-AI.