Code-as-Planner Meets Semantic-Graph State for Non-Markovian Vision-Language-Action Models
CodeGraphVLP combines persistent semantic-graph state with a task-specific executable planner to support long-horizon robot manipulation when task progress is not recoverable from the latest observation alone. Progress-guided visual-textual prompts focus the VLA executor on the current subtask and its relevant objects, improving robustness to visual distractors.
This repository contains the anonymous project page and qualitative results for the manuscript currently under double-anonymous review. Source code will be released upon acceptance.