A reference implementation of the Regression Planning Networks.

This repo also includes the simulated environments used in the paper: GridWorld and Kitchen3D. GridWorld is built on gym-minigrid. Kitchen3D is simulated with PyBullet. Most of the PyBullet utilities are adapted from Caelan Garrett's ss-pybullet. The Appendix of the original paper is also included here for reference.

To cite the code, use bibtex:

@inproceedings{xu2019rpn,
  title={Regression Planning Networks},
  author={Xu, Danfei and Martín-Martín, Roberto and Huang, De-An and Zhu, Yuke and Savarese, Silvio and Fei-Fei, Li},
  booktitle={Thirty-third Conference on Neural Information Processing Systems (NeurIPS)},
  year={2019}
 }

Requirement: Python >= 3.6

Note: The code DOES NOT work with Python 3.5

$ git clone https://github.com/danfeiX/rpn.git
$ git submodule update --init --recursive
$ pip install -r requirements.txt
$ cd rpn
$ mkdir checkpoints
$ mkdir data

To reproduce the main experiments, first create a demonstration dataset with three ingredients and two dishes (I=3, D=2).

$ python create_dataset_pb.py --problem pb_cook_meal_3i_2d_3m_iter \
    --dataset data/pb_cook_3i_2d_3m_10_s0.group \
    --num_episodes 10 --num_tasks -1 --seed 0 \
    --teleport --num_chunks 100 --num_workers 10

The entire process takes around 3 hours with 10 works (--num_workers 10) on a machine with i7-7700K CPU @ 4.20GHz. The dataset will be written to data/pb_cook_3i_2d_3m_10_s0.group.

To create a smaller validation set, run:

$ python create_dataset_pb.py --problem pb_cook_meal_3i_2d_3m_iter \
    --dataset data/pb_cook_3i_2d_3m_1_s1.group \
    --num_episodes 1 --num_tasks -1 --seed 1 \
    --teleport --num_chunks 10 --num_workers 10

To train our full RPN model, run:

$ python train.py --dataset data/pb_cook_3i_2d_3m_10_s0.group \
    --testset data/pb_cook_3i_2d_3m_1_s1.group \
    --exp rpn \
    --run experiment \
    --batch_size 128 \
    --eval_batch_size 128 \
    --eval_freq 1 --save_freq 1 \
    --num_epoch 10 --config configs/pbv_rpn.json \
    --num_workers 1

Checkpoints will be written to checkpoints/rpn/experiment/ each epoch. Both text and tensorboard event will be written to checkpoints/rpn/experiment/log/.

To evaluate a trained model on the Kitchen3D task with I=6 and D=3, run:

$ python eval_pb.py --restore_path checkpoints/rpn/experiment/ckpt_ep6.pt \
       --problem pb_cook_meal_6i_3d_6m_shuffle \
       --num_eval 1000 --seed 1 \
       --display

Aside: Our naming convention for the cooking tasks is pb_cook_meal_<I>i_<D>d_<M>m_<shuffle/iter>, where <I> is the number of ingredients involved, <D> is the number of dishes to cook, <M> is the maximum number of ingredients that can be placed on a plate (usually set to be equal to <I>), and <shuffle/iter> is whether to shuffle the list of all possible tasks under the current task specification.