On Sample-Efficient Generalized Planning via Learned Transition Models

Researchers propose a new approach to generalized planning that learns explicit transition models rather than directly predicting action sequences. This method achieves better out-of-distribution performance with fewer training instances and smaller models compared to Transformer-based planners like PlanGPT.

• →The approach formulates generalized planning as a transition-model learning problem using neural networks to approximate successor-state functions.
• →Learning explicit transition models outperforms direct action-sequence prediction methods in out-of-distribution scenarios.
• →The method requires significantly fewer training instances and smaller model sizes compared to existing Transformer-based approaches.
• →Traditional Transformer planners suffer from state drift in long-horizon settings due to lack of explicit world-state evolution modeling.
• →The research demonstrates improved sample efficiency and size-invariant generalization across multiple planning domains.