Refining Diffusion Planner for Reliable Behavior Synthesis by Automatic Detection of Infeasible Plans

This repository provides the source codes for our paper Refining Diffusion Planner for Reliable Behavior Synthesis by Automatic Detection of Infeasible Plans in NeurIPS 2023.

Installation

pip install -e .
pip install captum==0.6.0 gtimer
cd third_party/diffuser
pip install -e .

Using pretrained models

We offer pretrained models for gap prediction as well as pre-collected synthetic data. To download and set them up, run the following commands:

bash prepare_pretrained_models.sh
bash prepare_data.sh

Executing these commands will download and extract tar files, placing their contents into the logs/gap_predictor and logs/data directories respectively.

Planning using pretrained models

Evaluate RGG:

python scripts/maze2d/evaluate_rgg.py \
	--logbase ./logs/evaluate \
	--env_name maze2d-large-v1 \
	--task single_task \
	--diffusers_repo leekwoon/maze2d-large-v1-H384-T256 \
	--num_episodes 1000 \
	--spec rgg \
	--gap_predictor_path ./logs/pretrained_gap_predictor/maze2d-large-v1-H384-T256/500000_finish/2023_03_26_03_44_06/seed_0/state_best.pt

Evaluate RGG+:

python scripts/maze2d/evaluate_rggplus.py \
	--logbase ./logs/evaluate \
	--env_name maze2d-large-v1 \
	--task single_task \
	--diffusers_repo leekwoon/maze2d-large-v1-H384-T256 \
	--num_episodes 1000 \
	--spec rggplus \
	--gap_predictor_path ./logs/pretrained_gap_predictor/maze2d-large-v1-H384-T256/500000_finish/2023_03_26_03_44_06/seed_0/state_best.pt

Training from scratch

Generate synthetic data with:

python scripts/maze2d/make_data.py \
	--base_datadir ./logs/data \
	--env_name maze2d-large-v1 \
	--diffusers_repo leekwoon/maze2d-large-v1-H384-T256 \
	--n 500000

Compute restoration gap with:

python scripts/maze2d/compute_restoration_gaps.py \
	--env_name maze2d-large-v1 \
	--data_path ./logs/data/maze2d-large-v1-H384-T256/500000_finish.npz \
	--diffusers_repo leekwoon/maze2d-large-v1-H384-T256 \
	--strength 0.9 \
	--num_plan 10

Train a gap predictor with:

python scripts/train_gap_predictor.py \
	--base_logdir ./logs/gap_predictor \
	--env_name maze2d-large-v1 \
	--data_path ./logs/data/maze2d-large-v1-H384-T256/500000_finish.npz \
	--score_path ./logs/data/maze2d-large-v1-H384-T256/500000_finish_restoration_gaps.npy \
	--seed 0

Attribution map can be inspected using notebooks/maze2d/plot_attribution.ipynb.

Reference

@inproceedings{lee2023refining,
  title={Refining Diffusion Planner for Reliable Behavior Synthesis by Automatic Detection of Infeasible Plans},
  author={Lee, Kyowoon and Kim, Seongun and Choi, Jaesik},
  booktitle={Advances in Neural Information Processing Systems},
  year={2023},
}

License

This repository is released under the MIT license. See LICENSE for additional details.

Acknowledgements

This repository is extended from diffuser and diffusers.

leekwoon / rgg