Agentic VLA Inference Plugin for Long-Horizon Tasks in Scientific Experiments

ABSTRACT

Vision-language-action (VLA) models often suffer from the state gap issue when inferring open and long-horizon tasks in scientific scenarios. We propose an LLM-based agentic inference mechanism that intervenes when executing sequential manipulation tasks in labs. By performing explicit transition inference and generating transitional robotic action code, our plugin guides VLA models through missing steps. This inference-only intervention enables reliable execution of composite scientific workflows without any additional training or data collection.