Experience-Guided Initiation Search for Learned Skills in Skill Composition

2026-10-08

AI Summary

This study addresses the challenge of searching for initial configurations when deploying frozen skills in novel environments, where high interaction costs and unreliable historical experience pose significant difficulties. To tackle this, we propose EVIS, a framework that pioneers a strategy combining historical execution guidance with goal-side behavioral verification. Specifically, EVIS composes skills using Vision-Language-Action (VLA) policies, leveraging historical experience to prioritize candidate configurations before validating their effectiveness through actual behavior in the target environment. This approach substantially reduces reliance on expensive real-time interactions. Empirical evaluations demonstrate that EVIS effectively decreases the average number of queries across both single-skill and two-stage manipulation tasks, enabling the reliable discovery of initial configurations under constrained interaction budgets.

Abstract

Deploying frozen learned skills, such as Vision-Language-Action (VLA) policies, in new environments requires identifying initiation configurations that support reliable execution. In skill composition, an initiation configuration affects not only the current skill but also the physical state passed to subsequent skills, so successful execution of an individual skill does not necessarily imply successful completion of the composed task. Estimating target-specific capability through extensive rollouts is costly in real-world deployment, while directly reusing historical experience can be unreliable under environment changes. We propose EVIS, an Experience-Guided and Behavior-Validated Initiation Search framework for discovering reliable initiation configurations under limited target interaction. EVIS uses historical execution experience to prioritize promising candidates and target-environment behavior to validate whether they remain effective. We evaluate EVIS on single-skill and two-stage manipulation tasks with frozen VLA policies. EVIS reduces mean target-environment queries and improves reliable candidate discovery under small interaction budgets. These results show that combining historical guidance with target-side behavioral validation can reduce the interaction cost of deploying frozen learned skills in new environments.

Problem

• Skill Composition
• Initiation Search
• Vision-Language-Action Policies
• Limited Interaction
• Frozen Learned Skills

Innovation

• Skill Composition
• Initiation Search
• Vision-Language-Action Policies
• Experience-Guided Validation
• Interaction Efficiency