This is the repository for PKU CoRe 2023 Fall Course Project: Probabilistic PDDL Solver (Track of Abstract Reasoning).
First clone this repo:
git clone https://github.com/Intellouis/Probabilistic-PDDL-Solver
After creating a virtual (conda) environment, please install the requirements and prerequisites:
pip install -r requirements.txt
Following NTP Vat Release, we implement a lightweight environment of block stacking.
- Object space: We have 8 blocks, numbered 1 to 8. These blocks are totally same except for numbers.
- State space: We provide two versions of state description - both the precise locations of blocks (the coordinate space) and the PDDL-language descriptions. These coordinate-based state representations form the observation space of SGN, while those language-based descriptions form the observation space of LLM-based agents. We also provide the ground truth embeddings of each state during interactions and rollouts, which are used to determine whether a task is finished (whether the goal is reached).
- Embedding space: We use a 64-dimensional vector to represent a state, as details can be figured out in our paper.
- Action space: We use "put
$i$ on$j$ " to represent an action, where$i \in [1, 2, ..., 8]$ is the source block, and$j \in [1, 2, ..., 8]$ is the target place. In the domain of block stacking, we only allow agents placing a block right on the top of another block. This representation is consistent with the "grasp" and "place" in Continuous Relaxation of Symbolic Planner for One-Shot Imitation Learning. - Goal space: We also use a 64-dimensional vector to represent the goal state, in which 8 elements are 1 exactly. In each step of a trajectory, the goal vector is used to determine whether the goal has been reached (whether the task is finished).
- Tasks: We provide 2000 different tasks, which are obtained from Neural Task Programming: Learning to Generalize Across Hierarchical Tasks. Each task has a unique goal, forming a diverse set of test cases.
All the datasets (sampled from our environment using an expert policy) are listed in /dataset/.
We implement both the SGN-based PDDL Solver and the LLM-agent (baseline). The LLM-agent is based on GPT-4-11.6-preview.
our trained checkpoints are listed in /model/.
To train SGN, run:
python train_SGN.py
To make inference / test on downstream tasks, run:
python main.py --continuous 1 --clip 0
To take the ablation study, in which a clip function is added during the inference process, run:
python main.py --continuous 1 --clip 1
To sample data from the environment using expert policy, run:
python main.py --continuous 0
To utilize large language models for tackling the challenges in block stacking tasks, run:
python llm_agent.py
| SGN (w/o clip) | SGN (w/ clip) | MLP (w/o clip) | MLP (w/ clip) | LLM-agent | |
|---|---|---|---|---|---|
| Success Rate | 68% | 45% | 28% | 15% | 45% |
Success curves:
