DELTA is a novel approach for efficient long-term robot task planning using Large Language Models (LLMs). It integrates scene graphs as environment representations within LLMs to generate precise planning problem descriptions and decomposes long-term task goals into sub-goals for efficient problem solving. DELTA achieves higher planning success rates and significantly shorter planning times compared to state-of-the-art methods. The system uses a five-step process: domain generation, scene graph pruning, problem generation, goal decomposition, and autoregressive sub-task planning. DELTA first feeds scene graphs into LLMs to generate domain and problem specifications in formal planning language, then decomposes long-term goals into sub-goals using LLMs. The corresponding sub-problems are then solved autoregressively with an automated task planner. DELTA's key contributions include a novel combination of LLMs and scene graphs for extracting actionable knowledge and a strategy for task decomposition that improves planning success rates and efficiency. The system is evaluated on five domains, showing that DELTA outperforms existing LLM-based approaches in terms of success rates, plan quality, and planning time. DELTA's use of scene graphs and LLMs enables efficient planning in complex environments, and its decomposition strategy significantly reduces planning time and the number of expanded nodes. The system is implemented with pre-trained GPT-4-turbo, GPT-4o, and Llama-3.1-70B models, and is evaluated on four scene graphs from the 3D Scene Graph dataset. DELTA achieves the highest success rates in all domains, demonstrating its effectiveness in long-term robot task planning.DELTA is a novel approach for efficient long-term robot task planning using Large Language Models (LLMs). It integrates scene graphs as environment representations within LLMs to generate precise planning problem descriptions and decomposes long-term task goals into sub-goals for efficient problem solving. DELTA achieves higher planning success rates and significantly shorter planning times compared to state-of-the-art methods. The system uses a five-step process: domain generation, scene graph pruning, problem generation, goal decomposition, and autoregressive sub-task planning. DELTA first feeds scene graphs into LLMs to generate domain and problem specifications in formal planning language, then decomposes long-term goals into sub-goals using LLMs. The corresponding sub-problems are then solved autoregressively with an automated task planner. DELTA's key contributions include a novel combination of LLMs and scene graphs for extracting actionable knowledge and a strategy for task decomposition that improves planning success rates and efficiency. The system is evaluated on five domains, showing that DELTA outperforms existing LLM-based approaches in terms of success rates, plan quality, and planning time. DELTA's use of scene graphs and LLMs enables efficient planning in complex environments, and its decomposition strategy significantly reduces planning time and the number of expanded nodes. The system is implemented with pre-trained GPT-4-turbo, GPT-4o, and Llama-3.1-70B models, and is evaluated on four scene graphs from the 3D Scene Graph dataset. DELTA achieves the highest success rates in all domains, demonstrating its effectiveness in long-term robot task planning.