Publication
CodeGraphVLP: Code-as-Planner Meets Semantic-Graph State for Non-Markovian Vision-Language-Action Models
A non-Markovian VLA framework that combines persistent semantic-graph state with executable code-as-planner reasoning.
Abstract
CodeGraphVLP studies non-Markovian long-horizon robot manipulation by separating scene-state representation from action planning. The method maintains a persistent semantic graph state over task-relevant entities and relations, then uses executable code-as-planner reasoning over that graph.
The project targets settings where task-relevant evidence can be occluded or appear only earlier in the trajectory, and where clutter makes fine-grained visual grounding brittle.