Progressive Assembly Objective: Event-Triggered Skill Crystallization for Compositional Reinforcement Learning

Reinforcement learning excels at exploration but suffers from the consolidation gap: the inability to crystallize transient discoveries into reusable, compositional skills. We propose the Progressive Assembly Objective (PAO), a framework that explicitly separates continuous policy optimization from discrete skill-library growth. PAO introduces three coupled mechanisms: (i) event-triggered skill crystallization via change-point detection, (ii) effective assembly filtering that rewards reliability and non-redundancy, and (iii) dormancy-gated plasticity that protects crystallized skills from destructive updates. We provide a falsifiable experimental protocol demonstrating PAO’s superiority in longhorizon composition, catastrophic forgetting resistance, and structural efficiency.