ABIL: Learning for Long-Horizon Planning via Neuro-Symbolic Abductive Imitation

Paper | Project Page

News

• 2024.12.30 Release code.
• 2024.12.22 Release training data.
• 2024.11.17 ABIL is accepted by KDD 2025!

The Framework of Abductive Imitation Learning.

Contents

• Install
• Data
• Train
• Evaluation

🚧 This repository is under construction 🚧 -- Please check back for updates!

Install

1. Clone the repo into a local folder.

git clone https://github.com/Hoar012/ABIL-KDD-2025.git

cd ABIL-KDD-2025
conda create -n ABIL python=3.8
conda activate ABIL
pip install -r requirements.txt

2. Clone the Jacinle repo.

git clone https://github.com/vacancy/Jacinle --recursive
export PATH=<path_to_jacinle>/bin:$PATH

mini-behavior environment

cd ./hacl/envs/mini_behavior
pip install -e .

Cliport environment

cd ./hacl/envs/cliport
export CLIPORT_ROOT=$(pwd)
python setup.py develop

Data

Our training demonstrations are generated by Python scripts. View them separately in the following files:

• BabyAI: hacl/p/kfac/minigrid/data_generator.py
• Mini-BEHAVIOR: hacl/p/kfac/minibehavior/data_generator.py
• CLIPort: cliport_src/data_generator.py

Train

BabyAI

1. Train the grounding model.

jac-run babyai_src/train-babyai-abl.py minigrid goto  --use-offline=yes --structure-mode abl --action-loss-weight 1 --evaluate-interval 0 --iterations 1000 --append-expr

2. Train the Imitation Learning model.

jac-run babyai_src/babyai-abil-bc.py minigrid goto  --seed 33 --iterations 1000  --append-expr --load_domain dumps/abl-unlock33-load=scratch.pth

Evaluation

jac-run babyai_src/babyai-abil-bc.py minigrid goto  --seed 33 --iterations 1000  --append-expr --load_domain dumps/abl-unlock33-load=scratch.pth --load dumps/seed33/abil-bc-goto-load=scratch.pth --evaluate

BibTeX

@misc{shao2024learninglonghorizonplanningneurosymbolic,
      title={Learning for Long-Horizon Planning via Neuro-Symbolic Abductive Imitation}, 
      author={Jie-Jing Shao and Hao-Ran Hao and Xiao-Wen Yang and Yu-Feng Li},
      year={2024},
      eprint={2411.18201},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2411.18201}, 
}

Acknowledgement

PDSketch, BabyAI, Mini-BEHAVIOR, CLIPort