Natural Language PDDL (NL-PDDL) for Open-world Goal-oriented Commonsense Regression Planning in Embodied AI

We introduce a natural-language extension of symbolic PDDL designed for planning in environments with incomplete information. Our approach combines regression-style planning with commonsense reasoning to handle partially-observed states and potential misalignment between goals and action specifications. By leveraging lifted representations, the method achieves planning complexity independent of the number of ground objects, states, and actions. Evaluation across Blocksworld and ALFWorld demonstrates superior performance compared to existing baselines, with improved robustness to longer horizons and multimodal generalization.

View on OpenReview

Recommended citation: Xiaotian Liu, Armin Toroghi, Jiazhou Liang, David Courtis, Ruiwen Li, Ali Pesaranghader, Jaehong Kim, Tanmana Sadhu, Hyejeong Jeon, Scott Sanner. (2026). "Natural Language PDDL (NL-PDDL) for Open-world Goal-oriented Commonsense Regression Planning in Embodied AI." International Conference on Learning Representations (ICLR).
Download Paper