This repo contains code and data for running HELPER.

• Add HELPER-X support for ALFRED, Dialfred, and Tidy Task

(1) Start by cloning the repository:

git clone https://github.com/Gabesarch/HELPER.git

(1a) (optional) If you are using conda, create an environment:

conda create -n helper python=3.8

(2) Install PyTorch with the CUDA version you have. For example, run the following for CUDA 11.1:

pip install torch==1.10.0+cu111 torchvision==0.11.0+cu111 torchaudio==0.10.0 -f https://download.pytorch.org/whl/torch_stable.html

(3) Install additional requirements:

pip install -r requirements.txt

(4) Install Detectron2 (needed for SOLQ detector) with correct PyTorch and CUDA version. E.g. for PyTorch 1.10 & CUDA 11.1:

python -m pip install detectron2 -f https://dl.fbaipublicfiles.com/detectron2/wheels/cu111/torch1.10/index.html

(5) Install teach:

pip install -e teach

(6) Build SOLQ deformable attention:

cd ./SOLQ/models/ops && sh make.sh && cd ../../..

(7) Clone ZoeDepth repo

git clone https://github.com/isl-org/ZoeDepth.git
cd ZoeDepth
git checkout edb6daf45458569e24f50250ef1ed08c015f17a7

1. Download the TEACh dataset following the instructions in the TEACh repo

teach_download 

To our model on the TEACh dataset, you'll first need the GPT embeddings for example retrieval:

1. Download GPT embeddings for example retrieval: here. Unzip it to get the gpt_embedding folder in ./data folder (or in a desired foldered and set --gpt_embedding_dir argument). Alternatively, you can download the file with gdown (pip install gdown):

cd data
gdown 1kqZZXdglNICjDlDKygd19JyyBzkkk-UL
unzip gpt_embeddings.zip
rm gpt_embeddings.zip

TO run our model with estimated depth and segmentation, download the SOLQ and ZoeDepth checkpoints:

2. Download SOLQ checkpoint: here. Place it in the ./checkpoints folder (or anywhere you want and specify the path with --solq_checkpoint). Alternatively, you can download the file with gdown (pip install gdown):

cd checkpoints
gdown 1hTCtTuygPCJnhAkGeVPzWGHiY3PHNE2j

3. Download ZoeDepth checkpoint: here. Place it in the ./checkpoints folder (or anywhere you want and specify the path with --zoedepth_checkpoint). (Also make sure you clone the ZoeDepth repo: git clone https://github.com/isl-org/ZoeDepth.git) Alternatively, you can download the file with gdown (pip install gdown):

cd checkpoints
gdown 1gMe8_5PzaNKWLT5OP-9KKEYhbNxRjk9F

1. (if required) Start x server. if an X server is not already running on your machine. First, open a screen with the desired node, and run the following to open an x server on that node:

python startx.py 0

Specify the server port number with the argument --server_port (default 0).

2. Set OpenAI keys. Set Azure keys:

export AZURE_OPENAI_KEY={KEY}
export AZURE_OPENAI_ENDPOINT={ENDPOINT}

(If not using Azure) Important! If using openai API, append --use_openai to arguments. Then set openai key:

export OPENAI_API_KEY={KEY}

3. Run agent. To run the agent with all modules and estimated perception on TfD validation unseen, run the following:

python main.py \
 --mode teach_eval_tfd \
 --split valid_unseen \
 --gpt_embedding_dir ./data/gpt_embeddings \
 --teach_data_dir PATH_TO_TEACH_DATASET \
 --server_port X_SERVER_PORT_HERE \
 --episode_in_try_except \
 --use_llm_search \
 --use_constraint_check \
 --run_error_correction_llm \
 --zoedepth_checkpoint ./checkpoints/ZOEDEPTH-model-00015000.pth \
 --solq_checkpoint ./checkpoints/SOLQ-model-00023000.pth \
 --set_name HELPER_teach_tfd_validunseen

Change split to --split valid_seen to evaluate validation seen set.

All metrics will be saved to ./output/metrics/{set_name}. Metrics and videos will also automatically be logged to wandb.

To create movies of the agent, append --create_movie to the arguments. This will by default create a movie for every episode rendered to ./output/movies. To change the episode frequency of logging, alter --log_every (e.g., --log_every 10 to render videos every 10 episodes). To remove the map visualization, append --remove_map_vis to the arguments. This can speed up the episode since rendering the map visual can slow down episodes.

The following arguments can be removed to run the ablations:

1. Remove memory augmented prompting. Add argument --ablate_example_retrieval.
2. Remove LLM search (locator) (only random). Remove --use_llm_search.
3. Remove constraint check (inspector). Remove --use_constraint_check.
4. Remove error correction (rectifier). Remove --run_error_correction_llm.
5. Change openai model type. Change --openai_model argument (e.g., --openai_model gpt-3.5-turbo).

The following arguments can be added to run with ground truth:

1. GT depth --use_gt_depth. Reccomended to also add --increased_explore with estimated segmentation for best performance.
2. GT segmentation --use_gt_seg.
3. GT action success --use_gt_success_checker.
4. GT error feedback --use_GT_error_feedback.
5. GT constraint check using controller metadata --use_GT_constraint_checks.
6. Increase max API fails --max_api_fails {MAX_FAILS}.

To run with user feedback, add --use_progress_check. Two additional metric files (for feedback query 1 & 2) will be saved to ./output/metrics/{set_name}.

See the teach_edh branch for how to run the TEACh EDH evaluation.

If you like this paper, please cite us:

@inproceedings{sarch2023helper,
                        title = "Open-Ended Instructable Embodied Agents with Memory-Augmented Large Language Models",
                        author = "Sarch, Gabriel and
                        Wu, Yue and
                        Tarr, Michael and
                        Fragkiadaki, Katerina",
                        booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2023",
                        year = "2023"}