Dynamic Planning with a LLM

Conversely, traditional symbolic planners, such as the Fast-Forward planner (Hoffmann and Nebel, 2001) or the BFS(f) planner (Lipovetzky et al., 2014), excel at finding optimal plans efficiently. But symbolic planners require problem and domain descriptions as prerequisites (McDermott, 2000), which hampers their applicability in real-world scenarios where it may be infeasible to achieve these high informational demands. For instance, knowing a complete and accurate description of the goal may not be possible before exploring the environment through actions.

Previous work by (Liu et al., 2023) has shown that LLMs can generate valid problem files in the Planning Domain Definition Language (PDDL ) for many simple examples. Yet, the problem of incomplete information remains: agents often need to interact with the world to discover their surroundings before optimal planning can be applied. Some versions of PDDL have been proposed in the past to deal with probabilities or Task and Motion Planning, such as PPDDL and PDDLStream (Younes and Littman, 2004; Garrett et al., 2018), but these still assume a human designer encoding the agent’s understanding of the domain and the planning problem, rather than the agent learning from interactions. Therefore, where modern LLMs need minimal information to figure out a task, e.g. through Few-shot or In-Context Learning (Honovich et al., 2022; Chen et al., 2022; Min et al., 2022), traditional planners need maximal information.

In this work, we introduce the LLM Dynamic Planner (LLM-DP), a neuro-symbolic framework that integrates an LLM with a symbolic planner to solve embodied tasks.1 LLM-DP capitalises on the LLM’s ability to understand actions and their impact on their environment and combines it with the planner’s efficiency in finding solutions. Using domain knowledge, LLM-DP solves the Alfworld test set faster and more efficiently than a LLM-only (ReAct) approach.