SDA-PLANNER: State-Dependency Aware Adaptive Planner for Embodied Task Planning

TL;DR

This paper tackles the shortcomings of current LLM-based embodied task planners, notably fixed planning paradigms, lack of explicit action-state dependencies, and poor error handling. It introduces SDA-Planner, a three-component framework comprising a State-Dependency Graph for modeling action preconditions and effects, an Error Backtrack and Diagnosis module for targeted root-cause analysis, and an Adaptive Action SubTree Generation module for localized replanning with constrained search. The approach enables precise, context-aware plan revision by localizing affected subsequences and reconstructing them with plan-aware constraints, including backtracking and fake execution when needed. Empirical evaluation on the ALFRED benchmark shows SDA-Planner consistently improves success rate and goal completion, while reducing the number of corrections, demonstrating robust adaptability under execution errors. The work offers a practical advance toward reliable, scalable embodied planning in dynamic, real-world environments.

Abstract

Embodied task planning requires agents to produce executable actions in a close-loop manner within the environment. With progressively improving capabilities of LLMs in task decomposition, planning, and generalization, current embodied task planning methods adopt LLM-based architecture.However, existing LLM-based planners remain limited in three aspects, i.e., fixed planning paradigms, lack of action sequence constraints, and error-agnostic. In this work, we propose SDA-PLANNER, enabling an adaptive planning paradigm, state-dependency aware and error-aware mechanisms for comprehensive embodied task planning. Specifically, SDA-PLANNER introduces a State-Dependency Graph to explicitly model action preconditions and effects, guiding the dynamic revision. To handle execution error, it employs an error-adaptive replanning strategy consisting of Error Backtrack and Diagnosis and Adaptive Action SubTree Generation, which locally reconstructs the affected portion of the plan based on the current environment state. Experiments demonstrate that SDA-PLANNER consistently outperforms baselines in success rate and goal completion, particularly under diverse error conditions.

Figures (6)

• Figure 1: Paradigm overview of existing methods. (a) Iterative Planner produces an action at every time step. (b) Tree Planner generates the action tree in advance and conducts the plan search without adjusting the original tree.
• Figure 2: Overview of Sda-Planner. When executing the instruction , Sda-Planner first breaks down the plan and later adapts the plan when it encounters an error. Sda-Planner analyses with the state dependency graph and dynamically adapts the plan by different methods based on the error type, to solve the environment state error and the action precondition error.
• Figure 3: Illustration of the Error Backtrack and Diagnosis.
• Figure 4: Ablation study with different LLMs.
• Figure 5: Ablation study with different adaptation methods.