Project.Report-THU

Penglin Cai, Jason Wu, Yu Xingping, Chenhao Zhou

Keywords: One-Shot Imitation Learning; PDDL; Block Stacking

Abstract: One-shot imitation learning has aroused great interest in the efficiency of data and demonstrations. However, some problem settings such as block stacking require precise planning and thus exhibiting extremely low tolerance for errors. To tackle such a challenge, we use a probabilistic continuous planning method as a relaxation of the symbolic state representation. We implement a pipeline consisting of a Symbol Grounding Network (SGN) module and a Continuous Planner (CP) module, building a probabilistic PDDL solver in the domain of block stacking problems. By constructing an agent based on large language models, we demonstrate that our probabilistic PDDL solver greatly outperforms the baseline. We also provide variants of the solver, from which we show the effectiveness of our SGN module and CP module respectively through an ablation study. All these methods are tested on our hand-crafted lightweight environment, which is convenient for transplantation. We open-source our environments, models, as well as codes, and hope to be of use for further research. Our code is made available at \href{https://github.com/Intellouis/core}{https://github.com/Intellouis/core}.

Submission Number: 227