ALFRED

A Benchmark for Interpreting Grounded Instructions for Everyday Tasks
Mohit Shridhar, Jesse Thomason, Daniel Gordon, Yonatan Bisk,
Winson Han, Roozbeh Mottaghi, Luke Zettlemoyer, Dieter Fox
CVPR 2020

ALFRED (Action Learning From Realistic Environments and Directives), is a new benchmark for learning a mapping from natural language instructions and egocentric vision to sequences of actions for household tasks. Long composition rollouts with non-reversible state changes are among the phenomena we include to shrink the gap between research benchmarks and real-world applications.

For the latest updates, see: askforalfred.com

What more? Checkout ALFWorld – interactive TextWorld environments for ALFRED scenes!

Quickstart

Clone repo:

$ git clone https://github.com/askforalfred/alfred.git alfred
$ export ALFRED_ROOT=$(pwd)/alfred

Install requirements:

$ virtualenv -p $(which python3) --system-site-packages alfred_env # or whichever package manager you prefer
$ source alfred_env/bin/activate

$ cd $ALFRED_ROOT
$ pip install --upgrade pip
$ pip install -r requirements.txt

Download Trajectory JSONs and Resnet feats (~17GB):

$ cd $ALFRED_ROOT/data
$ sh download_data.sh json_feat

Train models:

$ cd $ALFRED_ROOT
$ python models/train/train_seq2seq.py --data data/json_feat_2.1.0 --model seq2seq_im_mask --dout exp/model:{model},name:pm_and_subgoals_01 --splits data/splits/oct21.json --gpu --batch 8 --pm_aux_loss_wt 0.1 --subgoal_aux_loss_wt 0.1

More Info

• Dataset: Downloading full dataset, Folder structure, JSON structure.
• Models: Training and Evaluation, File structure, Pre-trained models.
• Data Generation: Generation, Replay Checks, Data Augmentation (high-res, depth, segementation masks etc).
• Errata: Updated numbers for Goto subgoal evaluation.
• THOR 2.1.0 Docs: Deprecated documentation from Ai2-THOR 2.1.0 release.
• FAQ: Frequently Asked Questions.

SOTA Models

Open-source models that outperform the Seq2Seq baselines from ALFRED:

Episodic Transformer for Vision-and-Language Navigation
Alexander Pashevich, Cordelia Schmid, Chen Sun
Paper, Code

MOCA: A Modular Object-Centric Approach for Interactive Instruction Following
Kunal Pratap Singh*, Suvaansh Bhambri*, Byeonghwi Kim*, Roozbeh Mottaghi, Jonghyun Choi
Paper, Code

Contact Mohit to add your model here.

Prerequisites

• Python 3
• PyTorch 1.1.0
• Torchvision 0.3.0
• AI2THOR 2.1.0

See requirements.txt for all prerequisites

Hardware

Tested on:

• GPU - GTX 1080 Ti (12GB)
• CPU - Intel Xeon (Quad Core)
• RAM - 16GB
• OS - Ubuntu 16.04

Leaderboard

Run your model on test seen and unseen sets, and create an action-sequence dump of your agent:

$ cd $ALFRED_ROOT
$ python models/eval/leaderboard.py --model_path <model_path>/model.pth --model models.model.seq2seq_im_mask --data data/json_feat_2.1.0 --gpu --num_threads 5

This will create a JSON file, e.g. task_results_20191218_081448_662435.json, inside the <model_path> folder. Submit this JSON here: AI2 ALFRED Leaderboard. For rules and restrictions, see the getting started page.

Rules:

1. You are only allowed to use RGB and language instructions (goal & step-by-step) as input for your agents. You cannot use additional depth, mask, metadata info etc. from the simulator on Test Seen and Test Unseen scenes. However, during training you are allowed to use additional info for auxiliary losses etc.
2. During evaluation, agents are restricted to max_steps=1000 and max_fails=10. Do not change these settings in the leaderboard script; these modifications will not be reflected in the evaluation server.
3. Pick a legible model name for the submission. Just "baseline" is not very descriptive.
4. All submissions must be attempts to solve the ALFRED dataset.
5. Answer the following questions in the description: a. Did you use additional sensory information from THOR as input, eg: depth, segmentation masks, class masks, panoramic images etc. during test-time? If so, please report them. b. Did you use the alignments between step-by-step instructions and expert action-sequences for training or testing? (no by default; the instructions are serialized into a single sentence)
6. Share who you are: provide a team name and affiliation.
7. (Optional) Share how you solved it: if possible, share information about how the task was solved. Link an academic paper or code repository if public.
8. Only submit your own work: you may evaluate any model on the validation set, but must only submit your own work for evaluation against the test set.

Docker Setup

Install Docker and NVIDIA Docker.

Modify docker_build.py and docker_run.py to your needs.

Build

Build the image:

$ python scripts/docker_build.py 

Run (Local)

For local machines:

$ python scripts/docker_run.py
 
  source ~/alfred_env/bin/activate
  cd $ALFRED_ROOT

Run (Headless)

For headless VMs and Cloud-Instances:

$ python scripts/docker_run.py --headless 

  # inside docker
  tmux new -s startx  # start a new tmux session

  # start nvidia-xconfig (might have to run this twice)
  sudo nvidia-xconfig -a --use-display-device=None --virtual=1280x1024
  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 ~/alfred/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

  # source env
  source ~/alfred_env/bin/activate
  
  # set DISPLAY variable to match X server
  export DISPLAY=:0

  # check THOR
  cd $ALFRED_ROOT
  python scripts/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.

Cloud Instance

ALFRED 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 $ALFRED_ROOT/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
$ cd $ALFRED_ROOT
$ python scripts/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

Citation

If you find the dataset or code useful, please cite:

@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}
}

License

MIT License

Change Log

14/10/2020:

• Added errata for Goto subgoal evaluation.

28/10/2020:

• Added --use_templated_goals option to train with templated goals instead of human-annotated goal descriptions.

26/10/2020:

• Fixed missing stop-frame in Modeling Quickstart dataset (json_feat_2.1.0.zip).

07/04/2020:

• Updated download links. Switched from Google Cloud to AWS. Old download links will be deactivated.

28/03/2020:

• Updated the mask-interaction API to use IoU scores instead of max pixel count for selecting objects.
• Results table in the paper will be updated with new numbers.

Contact

Questions or issues? Contact [email protected]