Skip to content

Latest commit

 

History

History
184 lines (140 loc) · 10.6 KB

README.md

File metadata and controls

184 lines (140 loc) · 10.6 KB
Raw

SPOC 🖖: Imitating Shortest Paths in Simulation Enables Effective Navigation and Manipulation in the Real World

License Python 3.9+ Code style: black

This repository contains the code and data for the paper "Imitating Shortest Paths in Simulation Enables Effective Navigation and Manipulation in the Real World".

🐍 Setting up the Python environment 🐍

🛠 Local installation 🛠

pip install -r requirements.txt
pip install --extra-index-url https://ai2thor-pypi.allenai.org ai2thor==0+5d0ab8ab8760eb584c5ae659c2b2b951cab23246

🐳 Docker 🐳

Please see the README.md in the docker directory for instructions on how to build and run the docker image.

🏋️ Training 🏋️

📥 Downloading the training data 📥

We have two different types of training data, fifteen and all. The fifteen type has the agent navigating and fetching one of fifteen possible object types, the all type has the agent navigating to and fetching any object type (of many hundreds). To download the training data for the fifteen type, run the following command:

python -m scripts.download_training_data --save_dir /your/local/save/dir --types fifteen

and change fifteen to all to download the all type training data. By default, you will download data corresponding to all 10 task types (ObjectNav, Pickup, Fetch, etc); if you'd only like to download data for a subset of task types, look at the the --task_types flag.

📁 Dataset format 📁

Once you run the above command, you will have a directory structure that looks like this

/your/local/save/dir/<fifteen OR all>_type
    <TASK_TYPE>
        house_id_to_sub_house_id_train.json # This file contains a mapping that's needed for train data loading
        house_id_to_sub_house_id_val.json   # This file contains a mapping that's needed for val data loading
        train
            <HOUSEID>
                hdf5_sensors.hdf5 -- containing all the sensors that are not videos
                    <EPISODE_NUMBER>
                        <SENSOR_NAME>
                raw_navigation_camera__<EPISODE_NUMBER>.mp4
                raw_manipulation_camera__<EPISODE_NUMBER>.mp4
        val
            # As with train

The raw_*_camera_*.mp4 files contain videos of the agent's trajectories in the environments. The hdf5_sensors.hdf5 contains all sensors values that are not camera frames. Below is the definition of all the sensors found in the hdf5_sensors.hdf5 files:

Sensor Name Definition
an_object_is_in_hand Indicates whether an object is currently being held in hand.
house_index Indicates the index of the house in the dataset.
hypothetical_task_success Indicates whether the task will be successful if the agent would have issued done at this step.
last_action_is_random Indicates if the most recent action taken was chosen randomly.
last_action_str A string representation of the last action performed.
last_action_success Indicates whether the last action performed was successful.
last_agent_location Indicates agent's location in the world coordinate frame.
manip_accurate_object_bbox Indicates the bounding box for the target object in the manipulation camera, based on one method of calculation in simulation.
manip_task_relevant_object_bbox Indicates the bounding box for the target object in the manipulation camera, based on a second method of calculation in simulation.
minimum_l2_target_distance The minimum Euclidean (L2) distance to the target object or location.
minimum_visible_target_alignment Measures the minimum degree the agent needs to turn to center the object in the navigation camera frame (if object is visible).
nav_accurate_object_bbox Indicates the bounding box for the target object in the navigation camera, based on one method of calculation in simulation.
nav_task_relevant_object_bbox Indicates the bounding box for the target object in the navigation camera, based on a second method of calculation in simulation.
relative_arm_location_metadata Arm proprioceptive, relative location of the arm in the agent's coordinate frame.
room_current_seen Indicates whether this room has been seen before or not.
rooms_seen Count of rooms that have been visited by the agent.
templated_task_spec A dictionary of the task information, which can be used to generate the natural language description of the task.
visible_target_4m_count The count of targets visible within a 4-meter radius or distance.

🤖 Actions 🤖

Our discrete action space is as follows:

  • move_ahead (m): Move the agent’s base forward by 0.2 meters.
  • move_back (b): Move the agent’s base backward by 0.2 meters.
  • rotate_left (l): Rotate the agent’s base left by 30°.
  • rotate_right (r): Rotate the agent’s base right by 30°.
  • rotate_left_small (ls): Rotate the agent’s base left by 6°.
  • rotate_right_small (rs): Rotate the agent’s base right by 6°.
  • move_arm_up (yp): Move the arm up by 0.1 meters.
  • move_arm_down (ym): Move the arm down by 0.1 meters.
  • move_arm_out (zp): Extend the arm outward by 0.1 meters.
  • move_arm_in (zm): Retract the arm inward by 0.1 meters.
  • move_arm_up_small (yps): Move the arm up by 0.02 meters.
  • move_arm_down_small (yms): Move the arm down by 0.02 meters.
  • move_arm_out_small (zps): Extend the arm outward by 0.02 meters.
  • move_arm_in_small (zms): Retract the arm inward by 0.02 meters.
  • wrist_open (wp): Rotate the wrist counterclockwise by 10°.
  • wrist_close (wm): Rotate the wrist clockwise by 10°.
  • end (end): Signal the end of a task.
  • sub_done (sub_done): Mark a sub-task as complete.
  • pickup (p): Initiate a grasp action to pick up an object.
  • dropoff (d): Execute a release action to drop an object.

💪 Running training 💪

For training commands refer to the TRAINING_README.md file.

📊 Evaluation 📊

In order to run evaluation you'll need:

  1. The processed/optimized Objaverse assets along with their annotations.
  2. The set of ProcTHOR-Objaverse houses you'd like to evaluate on.
  3. A trained model checkpoint.

Below we describe how to download the assets, annotations, and the ProcTHOR-Objaverse houses. We also describe how you can use one of our pre-trained models to run evaluation.

💾 Downloading assets, annotations, and houses 💾

📦 Downloading optimized Objaverse assets and annotations 📦

Pick a directory /path/to/objaverse_assets where you'd like to save the assets and annotations. Then run the following commands:

python -m objathor.dataset.download_annotations --version 2023_07_28 --path /path/to/objaverse_assets
python -m objathor.dataset.download_assets --version 2023_07_28 --path /path/to/objaverse_assets

These will create the directory structure:

/path/to/objaverse_assets
    2023_07_28
        annotations.json.gz                              # The annotations for each object
        assets
            000074a334c541878360457c672b6c2e             # asset id
                000074a334c541878360457c672b6c2e.pkl.gz
                albedo.jpg
                emission.jpg
                normal.jpg
                thor_metadata.json
            ... #  39663 more asset directories

🏠 Downloading ProcTHOR-Objaverse houses 🏠

Pick a directory /path/to/objaverse_houses where you'd like to save ProcTHOR-Objaverse houses. Then run:

python -m scripts.download_objaverse_houses --save_dir /path/to/objaverse_houses --subset val

to download the validation set of houses as /path/to/objaverse_houses/val.jsonl.gz. You can also change val to train to download the training set of houses.

🛣 Setting environment variables 🛣

Next you need to set the following environment variables:

export OBJAVERSE_HOUSES_DIR=/path/to/objaverse_houses
export OBJAVERSE_DATA_DIR=/path/to/objaverse_assets

You can see some usage examples for the downloaded data in this jupyter notebook.

🚀 Running evaluation with a pretrained model 🚀

For evaluation commands refer to the TRAINING_README.md file.

📚 Attribution 📚

We use Open English WordNet 2022 dataset in our work, attribution to Princeton WordNet and the Open English WordNet team. See the english-wordnet repository for more details.

📝 Cite us 📝

@article{spoc2023,        
    author    = {Kiana Ehsani, Tanmay Gupta, Rose Hendrix, Jordi Salvador, Luca Weihs, Kuo-Hao Zeng, Kunal Pratap Singh, Yejin Kim, Winson Han, Alvaro Herrasti, Ranjay Krishna, Dustin Schwenk, Eli VanderBilt, Aniruddha Kembhavi},
    title     = {Imitating Shortest Paths in Simulation Enables Effective Navigation and Manipulation in the Real World},
    journal   = {arXiv},
    year      = {2023},
    eprint    = {2312.02976},
}