Unified Modeling of Autonomous Exploration and Goal Discovery for Open-Environment Agents

Abstract

This paper addresses the absence of explicit goal specification and the tendency toward fragmented exploration in agents operating within open environments. It proposes a unified modeling approach that integrates autonomous exploration and goal discovery. Goals are treated as internal and evolving variables and are incorporated together with environmental states into the decision process. Exploration, therefore, no longer depends on external task constraints or static reward design. Instead, it is organized around continuous generation and revision of goal structures. At the modeling level, a goal update mechanism captures how interaction experience shapes goal representations. Environmental feedback and intrinsic drivers are coordinated within a unified optimization framework. This enables joint evolution of exploration direction and decision preference. Systematic evaluation based on comparative results shows that the unified approach improves task completion and goal coverage in open environments. It also leads to more focused exploration with stronger structural consistency. Overall, the findings confirm that explicitly introducing goal discovery into exploration modeling enhances long-term decision stability and behavioral organization. The approach provides a feasible modeling strategy for building agents with intrinsic goal-driven mechanisms in open environments.