GraphThink Improves LLM Embodied Agent Planning with Graphs

Embodied agents, particularly those relying on large language models (LLMs) for planning, frequently encounter challenges such as "physical hallucinations" (planning actions inconsistent with the environment), poor generalization to complex, long-horizon tasks, and a lack of continuous environmental awareness. This research introduces GraphThink, a novel framework designed to address these limitations by integrating graph-based knowledge structures. GraphThink incorporates two key graph types: a task graph and a scene graph. The task graph provides structured knowledge that guides the LLM's planning process through contextual prompting and iterative refinement, effectively reducing planning hallucinations. Within the GRPO framework, this task graph also facilitates delicate reward design, which is crucial for training the LLM planner to handle long-horizon tasks and improve its generalization capabilities. Furthermore, a scene graph maintains an up-to-date environmental memory, enabling an event-driven replanning module. This allows the agent to react to changes in its environment, correct errors, and maintain closed-loop awareness. GraphThink has achieved state-of-the-art performance on the ALFRED benchmark, with its high-level planner outperforming leading API-based LLMs on both validation and held-out long-horizon tasks, demonstrating strong zero-shot and few-shot generalization to novel tasks and environments.