l2p : LLM-driven Planning Model library kit

This library is a collection of tools for PDDL model generation extracted from natural language driven by large language models. This library is an expansion from the survey paper Leveraging Large Language Models for Automated Planning and Model Construction: A Survey (coming soon)

L2P is an offline, natural language -to- planning model system that supports domain-agnostic planning. It does this via creating an intermediate PDDL representation of the domain and task, which can then be solved by a classical planner.

Usage

This is the general setup to build domain predicates:

from l2p.llm_builder import OPENAI
from l2p.utils import load_file
from l2p.domain_builder import DomainBuilder

domain_builder = DomainBuilder()

api_key = os.environ.get('OPENAI_API_KEY')
llm = OPENAI(model="gpt-4o-mini", api_key=api_key)

# retrieve prompt information
base_path='tests/usage/prompts/domain/'
domain_desc = load_file(f'{base_path}blocksworld_domain.txt')
extract_predicates_prompt = load_file(f'{base_path}extract_predicates.txt')
types = load_file(f'{base_path}types.json')
action = load_file(f'{base_path}action.json')

# extract predicates via LLM
predicates, llm_output = domain_builder.extract_predicates(
    model=llm,
    domain_desc=domain_desc,
    prompt_template=extract_predicates_prompt,
    types=types,
    nl_actions={action['action_name']: action['action_desc']}
    )

# format key info into PDDL strings
predicate_str = "\n".join([pred["clean"].replace(":", " ; ") for pred in predicates])

print(f"PDDL domain predicates:\n{predicate_str}")

Here is how you would setup a PDDL problem:

from l2p.task_builder import TaskBuilder

task_builder = TaskBuilder()

api_key = os.environ.get('OPENAI_API_KEY')
llm = OPENAI(model="gpt-4o-mini", api_key=api_key)

# load in assumptions
problem_desc = load_file(r'tests/usage/prompts/problem/blocksworld_problem.txt')
extract_task_prompt = load_file(r'tests/usage/prompts/problem/extract_task.txt')
types = load_file(r'tests/usage/prompts/domain/types.json')
predicates_json = load_file(r'tests/usage/prompts/domain/predicates.json')
predicates: List[Predicate] = [Predicate(**item) for item in predicates_json]

# extract PDDL task specifications via LLM
objects, initial_states, goal_states, llm_response = task_builder.extract_task(
    model=llm,
    problem_desc=problem_desc,
    prompt_template=extract_task_prompt,
    types=types,
    predicates=predicates
    )

# format key info into PDDL strings
objects_str = task_builder.format_objects(objects)
initial_str = task_builder.format_initial(initial_states)
goal_str = task_builder.format_goal(goal_states)

# generate task file
pddl_problem = task_builder.generate_task(
    domain="blocksworld",
    problem="blocksworld_problem",
    objects=objects_str,
    initial=initial_str,
    goal=goal_str)

print(f"### LLM OUTPUT:\n {pddl_problem}")

Here is how you would setup a Feedback Mechanism:

from l2p.feedback_builder import FeedbackBuilder

feedback_builder = FeedbackBuilder()

api_key = os.environ.get('OPENAI_API_KEY')
llm = OPENAI(model="gpt-4o-mini", api_key=api_key)

problem_desc = load_file(r'tests/usage/prompts/problem/blocksworld_problem.txt')
types = load_file(r'tests/usage/prompts/domain/types.json')
feedback_template = load_file(r'tests/usage/prompts/problem/feedback.txt')
predicates_json = load_file(r'tests/usage/prompts/domain/predicates.json')
predicates: List[Predicate] = [Predicate(**item) for item in predicates_json]
llm_response = load_file(r'tests/usage/prompts/domain/llm_output_task.txt')

objects, initial, goal, feedback_response = feedback_builder.task_feedback(
    model=llm,
    problem_desc=problem_desc,
    feedback_template=feedback_template,
    feedback_type="llm",
    predicates=predicates,
    types=types,
    llm_response=llm_response)

print("FEEDBACK:\n", feedback_response)

Installation and Setup

Currently, this repo has been tested for Python 3.11.10

You can set up a Python environment using either Conda or venv and install the dependencies via the following steps.

Conda

conda create -n L2P python=3.11.10
conda activate L2P
pip install -r requirements.txt

venv

python3.11.10 -m venv env
source env/bin/activate
pip install -r requirements.txt

These environments can then be exited with conda deactivate and deactivate respectively. The instructions below assume that a suitable environemnt is active.

API keys L2P requires access to an LLM. L2P provides support for OpenAI's GPT-series models. To configure these, provide the necessary API-key in an environment variable.

OpenAI

export OPENAI_API_KEY='YOUR-KEY' # e.g. OPENAI_API_KEY='sk-123456'

Refer to here for more information.

HuggingFace

Additionally, we have included support for using Huggingface models. One can set up their environment like so:

parser = argparse.ArgumentParser(description="Define Parameters")
parser.add_argument('-test_dataset', action='store_true')
parser.add_argument("--temp", type=float, default=0.01, help = "temperature for sampling")
parser.add_argument("--max_len", type=int, default=4e3, help = "max number of tokens in answer")
parser.add_argument("--num_sample", type=int, default=1, help = "number of answers to sample")
parser.add_argument("--model_path", type=str, default="/path/to/model", help = "path to llm")
args = parser.parse_args()

huggingface_model = HUGGING_FACE(model_path=args.model_path, max_tokens=args.max_len, temperature=args.temp)

llm_builder.py contains an abstract class and method for implementing any model classes in the case of other third-party LLM uses.

Planner

For ease of use, our library contains submodule FastDownward. Fast Downward is a domain-independent classical planning system that users can run their PDDL domain and problem files on. The motivation is that the majority of papers involving PDDL-LLM usage uses this library as their planner.

Current Works Reconstructed Using L2P

The following are papers that have been reconstructed so far. This list will be updated in the future.

• NL2Plan
• LLM+DM
• LLM+P
• PROC2PDDL

Current Model Construction Works

This section provides a taxonomy of research within Model Construction. More detailed overview coming soon.

Model Generation

Task Translation Frameworks

• "Structured, flexible, and robust: benchmarking and improving large language models towards more human-like behaviour in out-of-distribution reasoning tasks" Collins et al. (2022) paper code
• "Translating natural language to planning goals with large-language models" Xie et al. (2023) paper code
• "Faithful Chain-of-Thought Reasoning" Lyu et al. (2023) paper code
• "LLM+P: Empowering Large Language Models with Optimal Planning Proficiency" Liu et al. (2023) paper code
• "Dynamic Planning with a LLM" Dagan et al. (2023) paper code
• "TIC: Translate-Infer-Compile for accurate 'text to plan' using LLMs and logical intermediate representations" Agarwal and Sreepathy (2024) paper code
• "PDDLEGO: Iterative Planning in Textual Environments" Zhang et al. (2024) paper code

Domain Translation Frameworks

• "Learning adaptive planning representations with natural language guidance" Wong et al. (2023) paper code
• "Leveraging Pre-trained Large Language Models to Construct and Utilize World Models for Model-based Task Planning" Guan et al. (2023) paper code
• "PROC2PDDL: Open-Domain Planning Representations from Texts" Zhang et al. (2024) paper code

Hybrid Translation Frameworks

• "There and Back Again: Extracting Formal Domains for Controllable Neurosymbolic Story Authoring" Kelly et al. (2023) paper code
• "DELTA: Decomposed Efficient Long-Term Robot Task Planning using Large Language Models" Liu et al. (2024) paper code
• "ISR-LLM: Iterative Self-Refined Large Language Model for Long-Horizon Sequential Task Planning" Zhou et al. (2023) paper code
• "Consolidating Trees of Robotic Plans Generated Using Large Language Models to Improve Reliability" Sakib and Sun (2024) paper code
• "NL2Plan: Robust LLM-Driven Planning from Minimal Text Descriptions" Gestrin et al. (2024) paper code
• "Leveraging Environment Interaction for Automated PDDL Generation and Planning with Large Language Models" Mahdavi et al. (2024) paper code
• "Generating consistent PDDL domains with Large Language Models" Smirnov et al. (2024) paper code

Model Editing and Benchmarking

• "Exploring the limitations of using large language models to fix planning tasks" Gragera and Pozanco (2023) paper code
• "Can LLMs Fix Issues with Reasoning Models? Towards More Likely Models for AI Planning" Caglar et al. (2024) paper code
• "Large Language Models as Planning Domain Generators" Oswald et al. (2024) paper code
• "Planetarium: A Rigorous Benchmark for Translating Text to Structured Planning Languages" Zuo et al. (2024) paper code

Contact

Please contact [email protected] for questions, comments, or feedback about the L2P library.