ElasticVLA: Enhancing Semantic Instruction Following in Embodied AI via Visual-Logical Phase Alignment

While Vision-Language-Action (VLA) models generally perform well on short-term tasks, they are prone to repeated obsolete actions due to a reliance on "Rigid Time Encoding,"often leading to misalignment between planning and execution. To address this, we propose ElasticVLA, a novel architecture that shifts execution logic from time-driven to perception-driven. A key innovation is the integration of a VLM Agent, which explicitly decomposes complex instructions into logical atomic action sequences, providing a structured topological prior for the policy. Leveraging this structure, our Elastic Phase Aligner (EPA) dynamically synchronizes these logical steps with real-time visual feedback, acting as a semantic filter to suppress obsolete or redundant actions. Experiments on LIBERO, SimplerEnv-Bridge, and SimplerEnv-Fractal demonstrate that ElasticVLA achieves strong performance in complex instruction following while reducing redundant execution.