ssilenzi / diffusion-ccsp

Code for Compositional Diffusion-Based Continuous Constraint Solvers (CoRL 23)

Compositional Diffusion-Based Continuous Constraint Solvers

Project page: Diffusion-CCSP

Setting Up

• Clone this repo

  git clone https://github.com/zt-yang/diffusion-ccsp.git --recurse-submodules

• Set up Jacinle following the instructions here. If the directory you installed Jacinle does not share the same parent folder as this repository, put export PYTHONPATH=/your/path/to/Jacinle:$PYTHONPATH in diffusion-ccsp/setup.sh.

  cd ..
  git clone https://github.com/vacancy/Jacinle --recursive
  ## echo "export PYTHONPATH=/your/path/to/Jacinle:$PYTHONPATH" >> diffusion-ccsp/setup.sh  ## optional

• Set up dependencies.

  cd diffusion-ccsp
  conda create --name diffusion-ccsp python=3.9
  conda activate diffusion-ccsp
  pip install -r requirements.txt

• Source environment variables before running codes (includes conda activate diffusion-ccsp).

  source setup.sh

• Compile IK for Franka Panda if want to collect and test robot planning.

  (cd pybullet_engine/ikfast/franka_panda; python setup.py)

Download data and pre-trained models

By default, download for task RandomSplitQualitativeWorld. Download into data/, logs/, and wandb/ folder

python download_data_checkpoints.py

Solving CCSP

python solve_csp.py

Data Collection

Task 1-2: 2D Tasks

Generate data into data/ folder

python envs/data_collectors.py -world_name 'RandomSplitQualitativeWorld' -data_type 'train' -num_worlds 100
python envs/data_collectors.py -world_name 'RandomSplitQualitativeWorld' -data_type 'test' -num_worlds 10 -pngs -jsons

Some frequently used flags

• -world_name = RandomSplitWorld | TriangularRandomSplitWorld | RandomSplitQualitativeWorld: generates different geometric splitting datasets
• -num_worlds: number of data
• -pngs | -jsons: .png and .json files will be in render/{dataset_name} folder

Task 3-4: 3D & Robot Data

## task 4: packing 3D objects
python 3-panda-box-data.py

## task 3: stacking shapes
python 5-panda-stability-data.py

Custom Task

to add a new task

1. run dataset.py to generate the pt files and try evaluation / visualization
2. change dims in create_trainer() in train_utils.py
3. change init() and initiate_denoise_fns() in ConstraintDiffuser class of denoise_fn.py
4. change world.name in Trainer class of ddpm.py
5. train with debug=True and visualize=True
6. change wandb project name

Training

python train_ddpm.py -timesteps 1000 -EBM 'ULA' -input_mode qualitative

Citation

@inproceedings{yang2023diffusion,
  title={{Compositional Diffusion-Based Continuous Constraint Solvers}},
  author={Yang, Zhutian and Mao, Jiayuan and Du, Yilun and Wu, Jiajun and Tenenbaum, Joshua B. and Lozano-P{\'e}rez, Tom{\'a}s and Kaelbling, Leslie Pack},
  booktitle={Conference on Robot Learning},
  year={2023},
}