ValuePlanner: A Hierarchical Framework for Proactive Embodied Agents
ValuePlanner represents a cognitive architecture that operates hierarchically, aimed at fostering proactive and value-oriented actions in embodied agents. It separates the scheduling of high-level values from the execution of low-level actions, employing a cognitive module based on LLM to create symbolic subgoals through abstract value trade-offs. Subsequently, a classical PDDL planner converts these subgoals into actionable plans, enhanced by a closed-loop feedback system that improves the process. To assess autonomy beyond mere task completion rates, the authors introduce a value-focused evaluation suite that gauges cumulative value gain, alignment of preferences, and diversity in behavior. The framework's efficacy is illustrated through experiments conducted in the TongSim household setting. The paper can be found on arXiv with ID 2604.27699.
Key facts
- ValuePlanner is a hierarchical cognitive architecture for proactive embodied agents.
- It decouples high-level value scheduling from low-level action execution.
- An LLM-based cognitive module generates symbolic subgoals via value trade-offs.
- Subgoals are translated into action plans by a classical PDDL planner.
- A closed-loop feedback mechanism refines the process.
- A value-centric evaluation suite measures cumulative value gain, preference alignment, and behavioral diversity.
- Experiments were conducted in the TongSim household environment.
- The paper is available on arXiv with ID 2604.27699.
Entities
Institutions
- arXiv