I’m doing a research and the objective of this research is to explore and compar

I’m doing a research and the objective of this research is to explore and compare model-less LLM-Based Planning (which directly generate as executable programs, without relying on any planning-domain model) and model-based LLM-driven task planning (which uses LLM to generates planning-domain models and then uses classical planners to generate plans). Both approaches have so far been demonstrated on simple problems. My work aims to experiment and compare them on more complex robot task scenarios, hoping to learn new valuable lessons on their respective virtues and limitations.
In Model-based LLM approach I will use this paper: Leveraging Pre-trained Large Language Models to Construct and Utilize World Models for Model-based Task Planning
The experiment part is to take code generated in Model-based LLM approach paper (The paper that I attached) which contains a planning domain in form of PPDL and make executable commands for each action in gazebo and make visuals in moveit.
those executable commands must be capable of using a ur5e robot to stack square boxes in a big container. Therefore, the environment scenario consists of 5 square boxes of different colors, big container box, and ur5e robot.
Note: the code for PPDL has been provided in a paper
https://guansuns.github.io/pages/llm-dm
https://we.tl/t-5KnlG11c5H
For the executable commands, it’s python . That means from PDDL to python. This executable commands are responsible for the movement of ur5e robot. For example To enable pick and place and other commands for stacking. But when it comes to defining the environment and models, it’s different since in Gazebo it’s not Python as far as I know. I don’t know if it makes sense in a way
Customer files