GraphThink: Graph-Enhanced LLM Planning for Embodied Agents
A recent study presents GraphThink, a framework aimed at enhancing the planning abilities of embodied agents that utilize large language models (LLMs). This paper, which can be found on arXiv with the identifier 2608.07905, tackles prevalent challenges like physical hallucinations, inadequate long-horizon task generalization, and insufficient environmental awareness. GraphThink employs a task graph for structured knowledge to facilitate effective planning and a scene graph to preserve environmental memory for event-driven replanning. The task graph directs LLM reasoning through contextual prompts and iterative adjustments, reducing planning hallucinations. Within the GRPO framework, it also offers refined reward design to bolster the LLM planner's long-horizon capabilities. An event-driven replanning module, supported by the scene graph, fosters closed-loop environmental awareness and error correction, achieving top-tier results on the ALFRED benchmark. This research holds significance for artificial intelligence, robotics, and embodied cognition, influencing AI interactions with physical spaces.
Key facts
- GraphThink is a novel framework for LLM-based embodied agents.
- It uses a task graph for structured knowledge and a scene graph for environmental memory.
- The task graph guides LLM thinking via contextual prompting and iterative refinement.
- GRPO framework is used for reward design to train the LLM planner.
- Event-driven replanning module enables closed-loop environment awareness and error correction.
- GraphThink achieves state-of-the-art performance on the ALFRED benchmark.
- The paper is available on arXiv with ID 2608.07905.
- The research addresses physical hallucinations, poor generalization, and lack of environmental awareness.
Entities
Institutions
- arXiv