With the in-depth development of robotic embodied intelligence, Vision-Language-Action (VLA) modeling has emerged as the core technology enabling robots to achieve physical interaction and autonomous decision-making in complex environments. Existing VLA models have an inadequacy: they struggle to cope with dynamic environmental changes and execution anomalies. To address this, this paper proposes a method to improve the pi0 model. The goal is to tackle the following question: how can large language models (LLMs) be leveraged to enhance a robot’s adaptive and error-correcting capabilities in task planning and action adjustment. By introducing the read-failure state variable and re-selecting the most appropriate task plan based on the context of a specific task, the robot’s environmental adaptability and task execution success rate in complex scenarios are significantly enhanced. Meanwhile, the architecture and training process of the algorithm are optimized to enhance its synergy efficiency with the new module. In the multi-scenario robot task experiments, the improved model shows obvious advantages over the original model and other baseline methods in terms of the key indicators of task completion rate, which provides important support for the promotion of the VLA model in practical applications.