Tru-POMDP: Task Planning Under Uncertainty via Tree of Hypotheses and Open-Ended POMDPs

Researchers propose Tru-POMDP, a new AI planning system that combines Large Language Models with Bayesian planning to help home-service robots handle uncertain tasks and ambiguous instructions. The system uses a hierarchical Tree of Hypotheses to generate beliefs about possible world states and significantly outperforms existing LLM-based planners in kitchen environment tests.

• →Tru-POMDP addresses key challenges in robotic task planning including ambiguous instructions, hidden objects, and open-vocabulary scenarios.
• →The system combines LLM-generated hypotheses with rigorous Bayesian belief tracking for improved decision-making under uncertainty.
• →Experimental results show significant performance improvements over state-of-the-art LLM-based and hybrid planners in complex object rearrangement tasks.
• →The hierarchical Tree of Hypotheses approach enables more efficient belief-space planning compared to traditional methods.
• →The research demonstrates stronger robustness to ambiguity and occlusion while maintaining greater planning efficiency.