Text-Aware Diffusion for Policy Learning (TADPoLe) 🐸

Quick Start

Setup Repository and Environment

git clone https://github.com/brown-palm/tadpole.git
cd tadpole

conda create -n tadpole python=3.9
conda activate tadpole

pip install -r requirement.txt

Customize Configurations for Dog and Humanoid Environments

• Replace the following files in <conda path>/envs/tadpole/lib/python3.9/site-packages/dm_control/suite/ with corresponding ones in custom_envs/:

  Source File in dm_control/suite/	Customized File in custom_envs/
  dog.py	dog/dog.py
  dog.xml	dog/dog.xml
  humanoid.py	humanoid/humanoid.py
  humanoid.xml	humanoid/humanoid.xml

• We provide background image files skybox.png and grass.png in custom_envs/common_assets/. In dog.xml and humanoid.xml, input the correct path for background images at lines below:

  • <texture type="skybox" file="[path to skybox.png]" gridsize="3 4" gridlayout=".U..LFRB.D.." width="8192" height="8192"/>
  • <texture name="grass" file="[path to grass.png]" type="2d"/>

Policy Learning

To enable wandb logging for experiments, fill in the wandb_entity in cfgs/default.yaml and add use_wandb=True to the commands below.

Goal Achievement with TADPoLe

Humanoid Stand:

python src/train_dmc.py \
    task="humanoid-stand" \
    text_prompt="a person standing" \
    seed=0 \
    noise_level_base=400 

Dog Stand:

python src/train_dmc.py \
    task="dog-stand" \
    text_prompt="a dog standing" \
    seed=0 \
    noise_level_base=400 

Humanoid Novel Goal Achievement:

python src/train_dmc.py \
    task="humanoid-stand" \
    text_prompt="a person doing splits" \
    seed=0 \
    noise_level_base=400 

Robotic Manipulation with TADPoLe

MetaWorld task examples: "metaworld-door-close", "metaworld-window-open", "metaworld-coffee-push", etc.

python src/train_mw.py \
    task="metaworld-door-close" \
    text_prompt="closing a door" \
    seed=0 \
    noise_level_base=500 

Continuous Locomotion with Video-TADPoLe

Humanoid Walk:

python src/train_dmc.py \
    task="humanoid-walk" \
    text_prompt="a person walking" \
    seed=0 \
    tadpole_type="video-tadpole" \
    noise_level_base=500 

Dog Walk:

python src/train_dmc.py \
    task="dog-walk" \
    text_prompt="a dog walking" \
    seed=0 \
    tadpole_type="video-tadpole" \
    noise_level_base=500 

Citation

If you find this repository helpful, please consider citing our work:

@inproceedings{luo2024text,
  author={Luo, Calvin and He, Mandy and Zeng, Zilai and Sun, Chen},
  booktitle={Advances in Neural Information Processing Systems},
  title={Text-Aware Diffusion for Policy Learning},
  volume={37},
  year={2024}
}

Acknowledgement

This repo contains code adapted from tdmpc. We thank the authors and contributors for open-sourcing their code.