Aligning Text and Embodied Environments for Interactive Learning
Mohit Shridhar, Xingdi (Eric) Yuan, Marc-Alexandre Côté,
Yonatan Bisk, Adam Trischler, Matthew Hausknecht
ICLR 2021
ALFWorld contains interactive TextWorld environments (Côté et. al) that parallel embodied worlds in the ALFRED dataset (Shridhar et. al). The aligned environments allow agents to reason and learn high-level policies in an abstract space before solving embodied tasks through low-level actuation.
For the latest updates, see: alfworld.github.io
Install with pip (python3.6 or higher, but >3.7 might require updating dependencies):
$ pip install https://github.com/MarcCote/downward/archive/faster_replan.zip
$ pip install https://github.com/MarcCote/TextWorld/archive/handcoded_expert_integration.zip
$ git clone https://github.com/alfworld/alfworld.git alfworld
$ cd alfworld
$ pip install -r requirements.txt
$ pip install .
# PyPi install is broken for now :( See https://github.com/alfworld/alfworld/issues/27
# $ pip install alfworldDownload PDDL & Game files and pre-trained MaskRCNN detector:
$ export ALFWORLD_DATA=<storage_path>
$ alfworld-downloadUse --extra to download pre-trained checkpoints and seq2seq data.
Play a Textworld game:
$ alfworld-play-twPlay an Embodied-World (THOR) game:
$ alfworld-play-thorGet started with a random agent:
import numpy as np
import alfworld.agents.environment as environment
import alfworld.agents.modules.generic as generic
# load config
config = generic.load_config()
env_type = config['env']['type'] # 'AlfredTWEnv' or 'AlfredThorEnv' or 'AlfredHybrid'
# setup environment
env = getattr(environment, env_type)(config, train_eval='train')
env = env.init_env(batch_size=1)
# interact
obs, info = env.reset()
while True:
# get random actions from admissible 'valid' commands (not available for AlfredThorEnv)
admissible_commands = list(info['admissible_commands']) # note: BUTLER generates commands word-by-word without using admissible_commands
random_actions = [np.random.choice(admissible_commands[0])]
# step
obs, scores, dones, infos = env.step(random_actions)
print("Action: {}, Obs: {}".format(random_actions[0], obs[0]))Run python <script>.py configs/base_config.yaml
Installing from source is recommended for development.
Clone repo:
$ git clone https://github.com/alfworld/alfworld.git alfworldInstall requirements:
# Note: Requires python 3.6 or higher
$ virtualenv -p $(which python3.6) --system-site-packages alfworld_env # or whichever package manager you prefer
$ source alfworld_env/bin/activate
$ pip install --upgrade pip
$ pip install -r requirements.txt
$ cd alfworld
$ python setup.py developDownload PDDL & Game Files and pre-trained MaskRCNN detector:
$ export ALFWORLD_DATA=<storage_path>
$ python scripts/alfworld-downloadUse --extra to download pre-trained checkpoints and seq2seq data.
Train models:
$ python scripts/train_dagger.py configs/base_config.yamlPlay around with TextWorld and THOR demos.
- Data: PDDL, Game Files, Pre-trained Agents. Generating PDDL states and detection training images.
- Agents: Training and evaluating TextDAgger, TextDQN, VisionDAgger agents.
- Explore: Play around with ALFWorld TextWorld and THOR environments.
- Python 3.6
- PyTorch 1.2.0 (later versions might be ok)
- Torchvision 0.4.0 (later versions might be ok)
- AI2THOR 2.1.0
See requirements.txt for all prerequisites
Tested on:
- GPU - GTX 1080 Ti (12GB)
- CPU - Intel Xeon (Quad Core)
- RAM - 16GB
- OS - Ubuntu 16.04
Pull vzhong's image: https://hub.docker.com/r/vzhong/alfworld
OR
Install Docker and NVIDIA Docker.
Modify docker_build.py and docker_run.py to your needs.
Build the image:
$ python docker/docker_build.pyFor local machines:
$ python docker/docker_run.py
source ~/alfworld_env/bin/activate
cd ~/alfworldFor headless VMs and Cloud-Instances:
$ python docker/docker_run.py --headless
# inside docker
tmux new -s startx # start a new tmux session
# start nvidia-xconfig
sudo nvidia-xconfig -a --use-display-device=None --virtual=1280x1024
# start X server on DISPLAY 0
# single X server should be sufficient for multiple instances of THOR
sudo python ~/alfworld/docker/startx.py 0 # if this throws errors e.g "(EE) Server terminated with error (1)" or "(EE) already running ..." try a display > 0
# detach from tmux shell
# Ctrl b then d
# source env
source ~/alfworld_env/bin/activate
# set DISPLAY variable to match X server
export DISPLAY=:0
# check THOR
python ~/alfworld/docker/check_thor.py
###############
## (300, 300, 3)
## Everything works!!!You might have to modify X_DISPLAY in gen/constants.py depending on which display you use.
ALFWorld can be setup on headless machines like AWS or GoogleCloud instances. The main requirement is that you have access to a GPU machine that supports OpenGL rendering. Run startx.py in a tmux shell:
# start tmux session
$ tmux new -s startx
# start X server on DISPLAY 0
# single X server should be sufficient for multiple instances of THOR
$ sudo python ~/alfworld/scripts/startx.py 0 # if this throws errors e.g "(EE) Server terminated with error (1)" or "(EE) already running ..." try a display > 0
# detach from tmux shell
# Ctrl b then d
# set DISPLAY variable to match X server
$ export DISPLAY=:0
# check THOR
$ python ~/alfworld/docker/check_thor.py
###############
## (300, 300, 3)
## Everything works!!!You might have to modify X_DISPLAY in gen/constants.py depending on which display you use.
Also, checkout this guide: Setting up THOR on Google Cloud
18/12/2020:
- PIP package version available. The repo was refactored.
ALFWorld
@inproceedings{ALFWorld20,
title ={{ALFWorld: Aligning Text and Embodied
Environments for Interactive Learning}},
author={Mohit Shridhar and Xingdi Yuan and
Marc-Alexandre C\^ot\'e and Yonatan Bisk and
Adam Trischler and Matthew Hausknecht},
booktitle = {Proceedings of the International Conference on Learning Representations (ICLR)},
year = {2021},
url = {https://arxiv.org/abs/2010.03768}
}
ALFRED
@inproceedings{ALFRED20,
title ={{ALFRED: A Benchmark for Interpreting Grounded
Instructions for Everyday Tasks}},
author={Mohit Shridhar and Jesse Thomason and Daniel Gordon and Yonatan Bisk and
Winson Han and Roozbeh Mottaghi and Luke Zettlemoyer and Dieter Fox},
booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
year = {2020},
url = {https://arxiv.org/abs/1912.01734}
}
TextWorld
@inproceedings{cote2018textworld,
title={Textworld: A learning environment for text-based games},
author={C{\^o}t{\'e}, Marc-Alexandre and K{\'a}d{\'a}r, {\'A}kos and Yuan, Xingdi and Kybartas, Ben and Barnes, Tavian and Fine, Emery and Moore, James and Hausknecht, Matthew and El Asri, Layla and Adada, Mahmoud and others},
booktitle={Workshop on Computer Games},
pages={41--75},
year={2018},
organization={Springer}
}
- ALFWorld - MIT License
- TextWorld (Jericho) - GNU General Public License (GPL) v2.0
- Fast Downward - GNU General Public License (GPL) v3.0
Questions or issues? File an issue or contact Mohit Shridhar