RL training scripts for learning an agent using ProcTHOR.
This codebase contains implementation of training and evaluation code used in ProcTHOR and Embodied-Codebook.
Please refer to docker/README.md for more details.
export MY_ENV_NAME=procthor-rl
export CONDA_BASE="$(dirname $(dirname "${CONDA_EXE}"))"
export PIP_SRC="${CONDA_BASE}/envs/${MY_ENV_NAME}/pipsrc"
conda create --name $MY_ENV_NAME python=3.9
conda activate $MY_ENV_NAME
pip install -r requirements.txt
pip install --no-cache-dir --extra-index-url https://ai2thor-pypi.allenai.org ai2thor==0+ca10d107fb46cb051dba99af484181fda9947a28
pip install --no-cache-dir torch==2.0.1 torchvision open_clip_torch objaverse objathorPlease test the package before running the training script:
python scripts/test_prior.pyYou should see the loading of the procthor-10k dataset successfully.
Otherwise, please make sure have your .git-credentials file in the root directory.
Please note that we should now use CloudRendering for AI2THOR.
If you find errors related to vulkan, please make sure install vulkan-tools, libvulkan1, and vulkan-utils correctly.️
Training without wandb logging:
python procthor_objectnav/main.py \
experiment=procthor_objectnav/experiments/rgb_clipresnet50gru_ddppo \
agent=locobot \
target_object_types=robothor_habitat2022 \
wandb.project=procthor-training \
machine.num_train_processes=96 \
machine.num_val_processes=4 \
ai2thor.platform=CloudRendering \
model.add_prev_actions_embedding=true \
procthor.p_randomize_materials=0.8 \
seed=100Training with wandb logging:
export WANDB_API_KEY=YOUR_WANDB_API_KEY
python procthor_objectnav/main.py \
experiment=procthor_objectnav/experiments/rgb_clipresnet50gru_ddppo \
agent=locobot \
target_object_types=robothor_habitat2022 \
wandb.project=procthor-training \
machine.num_train_processes=96 \
machine.num_val_processes=4 \
ai2thor.platform=CloudRendering \
model.add_prev_actions_embedding=true \
procthor.p_randomize_materials=0.8 \
wandb.name=YOUR_WANDB_RUN_NAME \
wandb.project=YOUR_WANDB_PROJECT_NAME \
wandb.entity=YOUR_WANDB_ENTITY_NAME \
callbacks=wandb_logging_callback \
seed=100Training with Codebook bottleneck:
python procthor_objectnav/main.py \
experiment=procthor_objectnav/experiments/rgb_clipresnet50gru_codebook_ddppo \
agent=locobot \
target_object_types=robothor_habitat2022 \
wandb.project=procthor-training \
machine.num_train_processes=96 \
machine.num_val_processes=4 \
ai2thor.platform=CloudRendering \
model.add_prev_actions_embedding=true \
procthor.p_randomize_materials=0.8 \
seed=100You can change the default values of the codebook hyperparameters such as the codebook size (model.codebook.size and model.codebook.code_dim) and codebook dropout (model.codebook.dropout):
python procthor_objectnav/main.py \
experiment=procthor_objectnav/experiments/rgb_clipresnet50gru_codebook_ddppo \
agent=locobot \
target_object_types=robothor_habitat2022 \
wandb.project=procthor-training \
machine.num_train_processes=96 \
machine.num_val_processes=4 \
ai2thor.platform=CloudRendering \
model.add_prev_actions_embedding=true \
procthor.p_randomize_materials=0.8 \
model.codebook.size=256 \
model.codebook.code_dim=10 \
model.codebook.dropout=0.1 \
seed=100Training with DINOv2 visual encoder:
python procthor_objectnav/main.py \
experiment=procthor_objectnav/experiments/rgb_dinov2gru_ddppo \
agent=locobot \
target_object_types=robothor_habitat2022 \
wandb.project=procthor-training \
machine.num_train_processes=96 \
machine.num_val_processes=4 \
ai2thor.platform=CloudRendering \
model.add_prev_actions_embedding=true \
procthor.p_randomize_materials=0.8 \
seed=100Use scripts/download_ckpt.py to download the pretrained checkpoint:
python scripts/download_ckpt.py --save_dir YOUR_CKPT_DIR --ckpt_ids CKPT_IDOptions for CKPT_ID: CLIP-GRU, DINOv2-GRU, CLIP-CodeBook-GRU, DINOv2-CodeBook-GRU.
Evaluate in ArchitecTHOR, ProcTHOR-10k, iTHOR, or RoboTHOR:
export WANDB_API_KEY=YOUR_WANDB_API_KEY
python procthor_objectnav/main.py \
experiment=procthor_objectnav/experiments/rgb_clipresnet50gru_ddppo \
agent=locobot \
target_object_types=robothor_habitat2022 \
machine.num_train_processes=1 \
machine.num_test_processes=20 \
ai2thor.platform=CloudRendering \
model.add_prev_actions_embedding=true \
callbacks=wandb_logging_callback \
seed=100 \
eval=true \
evaluation.tasks=["architecthor"] \
evaluation.minival=false \
checkpoint=YOUR_CHECKPOINT \
wandb.name=YOUR_WANDB_RUN_NAME \
wandb.project=YOUR_WANDB_PROJECT_NAME \
wandb.entity=YOUR_WANDB_ENTITY_NAME \
visualize=true \ # store qualitative results
output_dir=YOUR_OUTPUT_DIR # dir to store both qualitative and quantitative resultsevaluation.tasks can be architecthor, procthor-10k, ithor, or robothor.
If you find this codebase useful, please consider citing ProcTHOR, Embodied-Codebook, and AllenAct:
@inproceedings{deitke2022️,
title={🏘️ ProcTHOR: Large-Scale Embodied AI Using Procedural Generation},
author={Deitke, Matt and VanderBilt, Eli and Herrasti, Alvaro and Weihs, Luca and Ehsani, Kiana and Salvador, Jordi and Han, Winson and Kolve, Eric and Kembhavi, Aniruddha and Mottaghi, Roozbeh},
booktitle={NeurIPS},
year={2022}
}
@inproceedings{eftekhar2023selective,
title={Selective Visual Representations Improve Convergence and Generalization for Embodied AI},
author={Eftekhar, Ainaz and Zeng, Kuo-Hao and Duan, Jiafei and Farhadi, Ali and Kembhavi, Ani and Krishna, Ranjay},
booktitle={ICLR},
year={2024}
}
@article{AllenAct,
author = {Luca Weihs and Jordi Salvador and Klemen Kotar and Unnat Jain and Kuo-Hao Zeng and Roozbeh Mottaghi and Aniruddha Kembhavi},
title = {AllenAct: A Framework for Embodied AI Research},
journal = {arXiv preprint arXiv:2008.12760},
year = {2020},
}