Abstract In the realm of data-driven AI technology, the application of open-source large language models (LLMs) in robotic task planning represents a significant milestone. Recent robotic task planning methods based on open-source LLMs typically leverage vast task planning datasets to enhance models’ planning abilities. While these methods show promise, they struggle with complex long-horizon tasks, which require comprehending more context and generating longer action sequences. This paper addresses this limitation by proposing MLDT, the Multi-Level Decomposition Task planning method. This method innovatively decomposes tasks at the goal-level, task-level, and action-level to mitigate the challenge of complex long-horizon tasks. In order to enhance open-source LLMs’ planning abilities, we introduce a goal-sensitive corpus generation method to create high-quality training data and conduct instruction tuning on the generated corpus. Since the complexity of the existing datasets is not high enough, we construct a more challenging dataset, LongTasks, to specifically evaluate planning ability on complex long-horizon tasks. We evaluate our method using various LLMs on four datasets in VirtualHome. Our results demonstrate a significant performance enhancement in robotic task planning, showcasing MLDT’s effectiveness in overcoming the limitations of existing methods based on open-source LLMs as well as its practicality in complex, real-world scenarios.

This is the accompanying code & benchmarks for the paper "MLDT: Multi-Level Decomposition for Complex Long-Horizon Robotic Task Planning with Open-Source Large Language Model".
The paper has been accepted by the International Conference on Database Systems for Advanced Applications (DASFAA-2024).

conda create -n MLDT python=3.10
conda activate MLDT
pip install -r requirement.txt

We conduct experiments in VirtualHome. Download VirtualHome exectuable file (v2.2.5) from here and unzip it to MLDT/MLDT/virtualhome/.

MLDT/
└── MLDT/
    └── virtualhome/                                                          

We employ various LLMs of different scales as the backbone, including Llama-2-chat (7B, 13B), bloom (3B, 7B), ChatGLM3-6B. To examine the effectiveness of our method on long-context LLMs, we use LongAlpaca (7B, 13B), ChatGLM3-6B-32K, the long-context versions of Llama-2-chat and ChatGLM, respectively. Download LLMs to MLDT/pretrain/.

MLDT/
└── pretrain/
    ├── bloom-3b/
    ├── bloom-7b/
    ├── chatglm3-6b/
    ├── chatglm3-6b-32k/
    ├── llama-2-7b-chat-hf/
    ├── llama-2-13b-chat-hf/
    ├── LongAlpaca-7B/
    └── LongAlpaca-13B/                         

We evaluate our method on four datasets, three (In-Distribution, NovelScenes, NovelTasks) from LID and one (LongTasks) created by ourselves. Download the four datasets from here and unzip it to MLDT/MLDT/data/test_init_env/

MLDT/
└── MLDT/
    ├── data/                  
        └── test_init_env/                                        

You can run MLDT/MLDT/data/create_long.py to create the LongTasks dataset. We remove some samples due to the environment bug from LID.

We provide all the instruction datasets for different methods in MLDT/Instruction-Tuning/: "multi-layer" for "MLDT", "react" for "ReAct", "embodied" for "Embodied", "task-action" for "MLDT_-goal", "goal-action" for "MLDT_-task". You can go to MLDT/MLDT/task-planning/ and run bash scripts/llm_collection.sh to generate the training corpus for "MLDT". You can modify the parameters like "subset", "max_retry" to generate your own data. For other methods, you can either modify the python scripts to generate training corpus from scratch or use some tricks like regular expressions to obtain the training corpus based on the generated corpus.

Go to MLDT/Instruction-Tuning/. Run run_bloom-3b.sh or run_bloom-7b.sh for fine-tuning bloom. Run run_chatglm.sh for fine-tuning ChatGLM3-6B or ChatGLM3-6B-32K. Run run_llama-7b.sh or run_llama-13b.sh for fine-tuning Llama-2-chat or LongAlpaca. You can modify the parameters like "dataset", "train_batch_size", "accumulation_steps" to fit your own training.

We provide our lora checkpoint for each method: MLDT, Embodied, ReAct, MLDT_-goal, MLDT_-task. You can skip instruction tuning and use these checkpoints directly.

Go to MLDT/MLDT/task-planning/. Run bash scripts/llm_eval.sh to evaluate open-source LLMs for "MLDT", "Embodied", "ReAct", "MLDT_-goal", and "MLDT_-task". Run bash scripts/llm_eval_demo.sh to evaluate open-source LLMs for "MLDT_-ft". Run bash scripts/gpt_eval.sh to evaluate closed-source LLMs for "MLDT", "Embodied", "ReAct".

• base_port: port number for VirtualHome environment
• llm: the path to LLM backbone
• lora: the path to lora weight, "None" for using LLM backbone only or closed-source LLMs
• mode: select task planning method: "multi-layer" for "MLDT", "react" for "ReAct", "embodied" for "Embodied", "goal-action" for "MLDT_-task", "task-action" for "MLDT_-goal"
• api: API key for ChatGPT
• demo: add this to use demonstrations
• max_retry: the number of times that task planning models can try, we set 1 in our experiment, you can set larger for higher sucess rate but longer inference time, it is useful for generating more training corpus