# Agent Priors-guided Policy Learning

The work introduces Agent Priors-guided Policy Learning (APPL): a construction agent segments complete demonstrations into reusable skills, proposes several structural priors per skill and trains one Diffusion Policy per prior, then writes the same prior into a runtime interface so a runtime agent can select and compose the frozen policies; on six MetaWorld tasks it raises OOD success with two demonstrations from 37.9% for a fixed relational prior to 89.6%, and on five long-horizon ManiSkill tasks it reaches 50.0% success under shifted objects (versus 10.0% for a full-task Diffusion Policy), 92.5% on task-level variants, and solves 8 of 16 previously unseen skill compositions, while hiding the interface information substantially reduces performance.

The work introduces Agent Priors-guided Policy Learning (APPL): a construction agent segments complete demonstrations into reusable skills, proposes several structural priors per skill and trains one Diffusion Policy per prior, then writes the same prior into a runtime interface so a runtime agent can select and compose the frozen policies; on six MetaWorld tasks it raises OOD success with two demonstrations from 37.9% for a fixed relational prior to 89.6%, and on five long-horizon ManiSkill tasks it reaches 50.0% success under shifted objects (versus 10.0% for a full-task Diffusion Policy), 92.5% on task-level variants, and solves 8 of 16 previously unseen skill compositions, while hiding the interface information substantially reduces performance.