Awesome

Official Code of our AAAI 2024 paper.

Quickstart

1. Install all requirements.

Using Docker: Download and run the docker image. Ensure you have Docker installed.

docker pull rishihazra/llm-dgx:torch-2.0.0-latest
docker run -it rishihazra/llm-dgx:torch-2.0.0-latest

You can also build it from scratch using docker/Dockerfile.

Alternatively, you can create a conda environment and install all requirements.

conda create -n saycanpay_env python=3
source activate saycanpay_env
pip install -r docker/requirements.txt

2. Now define environment paths

export RAVENS_ROOT=$(pwd)/llm_planning/ravens
export BabyAI=$(pwd)/llm_planning/babyai
export VIRTUALHOME=$(pwd)/llm_planning/virtualhome/src/virtualhome
export PLANNER=$(pwd)/llm_planning/planLM
cd $PLANNER

To use the language interface in BaByAI with our high-level actions, run this additional line:

cp -f ../babyai/unlock.py /opt/conda/lib/python3.8/site-packages/minigrid/envs/babyai/unlock.py  # for python 3.8

Data Split Generation

• To generate data, say for instance for BabyAI (using multiprocessing)

python3 babyai_inference.py \  # change 
        planner.agent_type=oracle \  # generates oracle trajectories
        domain.mode=train \  # data split type
        save_data=True \  # to save or not to save data
        parallel=True \  # set False if multiprocessing is not required
        n=400  # number of trajectories in the split
python3 babyai_inference.py \
        planner.agent_type=oracle \
        domain.mode=test-success \
        save_data=True \
        parallel=True \
        n=100
python3 babyai_inference.py \
        planner.agent_type=oracle \
        domain.mode=test-optimal \
        save_data=True \
        parallel=True \
        n=100
python3 babyai_inference.py \
        planner.agent_type=oracle \
        domain.mode=test-generalize \
        save_data=True \
        parallel=True \
        n=100

For Ravens:

python3 ravens_inference.py \
        planner.agent_type=oracle \
        task=towers-of-hanoi-seq \  # put-block-in-bowl
        mode=train \
        save_data=True \
        parallel=True \
        n=800  # 100 for test splits

For VirtualHome, we provide a sub-set of processed crowdsourced plans in $VIRTUALHOME/data/oracle-plans.

Train Can, Pay models

Training framework is Distributed Data Parallel over multi-GPU.

python3 train/babyai_train.py \  # ravens_train.py, virtualhome_train.py
        train.task=pickup \  # towers-of-hanoi-seq (Ravens), put-block-in-bowl (Ravens), remove for VirtualHome
        train.model=can \  # pay for Pay model
        train.max_epochs=30 \  # 20 for Pay model
        train.batch_size=60

For Inference (Plan Generation)

python3 ravens_inference.py \  # virtualhome_inference.py, # babyai_inference.py
        task=put-block-in-bowl \
        mode=test-success \  # test-optimal, test-generalize
        planner.agent_type=lm \  
        planner.model_name=flan_t5 \  # vicuna
        planner.decoding_type=beam_action \  # greedy_token, greedy_action
        planner.decoding_score=say_can_pay  # say, say_can (remove for greedy_token decoding_type)

References

[1] Transporter Networks: Rearranging the Visual World for Robotic Manipulation, Zeng et al., CoRL 2020
[2] BabyAI: First Steps Towards Grounded Language Learning With a Human In the Loop, Chevalier-Boisvert et al., ICLR 2019
[3] Virtualhome: Simulating household activities via programs, Puig et al., CVPR 2018

To cite our paper:

@inproceedings{hazra2024saycanpay,
  title={SayCanPay: Heuristic Planning with Large Language Models using Learnable Domain Knowledge},
  author={Hazra, Rishi and Dos Martires, Pedro Zuidberg and De Raedt, Luc},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={38},
  number={18},
  pages={20123--20133},
  year={2024}
}