Dataset#

Furniture assembly is a complex, long-horizon manipulation task, which is very challenging to solve using reinforcement learning. To make our benchmark tractable, we provide 219.6 hours of 5100 successful demonstrations collected using an Oculus Quest 2 controller and a keyboard.

Each furniture assembly task has three different levels with respect to the randomness in task initialization: low, medium, and high.

Download Dataset#

The FurnitureBench dataset is available on our Google Drive:

low               # Low randomness
  |- cabinet      # Demonstration files for cabinet
    |- 0.pkl
    |- 1.pkl
    |- ...
  |- chair        # Demonstration files for chair
  |- ...
med               # Medium randomness
  |- cabinet
  |- ...
high              # High randomness
  |- cabinet
  |- ...

For easy downloading, we also provide compressed datasets (*.tar.gz) for each furniture model and randomness level under low_compressed, med_compressed, and high_compressed directory. You can download our data with commands below.

Download with gdown#

pip install gdown

python furniture_bench/scripts/download_dataset.py --untar --randomness [low|med|high] --furniture <furniture> --out_dir <path/to/data>

# E.g., download lamp data with low randomness
python furniture_bench/scripts/download_dataset.py --untar --randomness low --furniture lamp --out_dir ./furniture_dataset

# E.g., download all furniture data with medium randomness
python furniture_bench/scripts/download_dataset.py --untar --randomness med --furniture all --out_dir ./furniture_dataset

Note! If you are facing download issues related to Too many users have viewed or downloaded..., please refer to the following Download with rclone section.

Download with rclone#

Sometimes gdown might be slow or reject your download request due to access quota limitations. In this case, you can utilize rclone to download the dataset.

How to use rclone#

Expand to see instructions

Install rclone.
Run rclone config to setup Google Drive remote:

No remotes found, make a new one?
n) New remote
s) Set configuration password
q) Quit config
n/s/q> n
--------------------

Enter name for new remote.
name> furniture
--------------------

Choose a number from below, or type in your own value
Storage> 18
--------------------

Two double "Enter" to skip client_id and client_secret
--------------------

Choose a number from below, or type in your own value.
Press Enter to leave empty.
scope> 2
--------------------

Enter a value. Press Enter to leave empty.
service_account_file> "Enter"
--------------------

Edit advanced config?
y) Yes
n) No (default)
y/n> n
--------------------

Use web browser to automatically authenticate rclone with remote?
* Say Y if the machine running rclone has a web browser you can use
* Say N if running rclone on a (remote) machine without web browser access
If not sure try Y. If Y failed, try N.

y) Yes (default)
n) No
y/n> n
--------------------

Option config_token.
For this to work, you will need rclone available on a machine that has
a web browser available.
For more help and alternate methods see: https://rclone.org/remote_setup/
Execute the following on the machine with the web browser (same rclone
version recommended):
        rclone authorize "drive" "<Your config_token>"
Then paste the result.
Enter a value.
config_token>

*Writer's note*
# Copy and past `rclone authorize "drive" "<Your config_token>"` in a machine with web browser
# Login to your Google account
# Allow rclone to access your Google Drive
# Past the result to `config_token` in the terminal
--------------------

Configure this as a Shared Drive (Team Drive)?

y) Yes
n) No (default)
y/n> n
--------------------

Keep this "furniture" remote?
y) Yes this is OK (default)
e) Edit this remote
d) Delete this remote
y/e/d> y
--------------------

Current remotes:

Name                 Type
====                 ====
furniture            drive

e) Edit existing remote
n) New remote
d) Delete remote
r) Rename remote
c) Copy remote
s) Set configuration password
q) Quit config
e/n/d/r/c/s/q> q
--------------------

Connect to Google Drive remote by opening the dataset Google Drive and clicking “Add a shortcut to Drive”, “My Drive”, and “Add”.

The following Python script will download our dataset. Use --untar flag to decompress files after downloading.

python furniture_bench/scripts/download_dataset.py --randomness [low/med/high] --furniture <name of furniture> --out_dir <path/to/data> --use-rclone <--untar>

# E.g., download lamp data with low randomness
python furniture_bench/scripts/download_dataset.py --randomness low --furniture lamp --out_dir ./furniture_dataset --use-rclone

# E.g., download all furniture data with medium randomness
python furniture_bench/scripts/download_dataset.py --randomness med --furniture all --out_dir ./furniture_dataset --use-rclone

Dataset Size#

The size (in GB) of demonstrations of raw .pkl files for each furniture in each level is summarized below:

Furniture	low	med	high
lamp	26	27	11
square_table	76	75	25
desk	46	57	25
drawer	43	39	11
cabinet	38	36	17
round_table	25	26	15
stool	37	42	19
chair	54	68	31
one_leg	112	129	69
Total	457	499	223

Demonstration File Format#

Each demonstration is stored in a .pkl file, containing a sequence of sensory inputs, actions, rewards, and other metadata:

'furniture': Furniture name, e.g., 'lamp'
'observations': List of observation dicts
  {
    'color_image1': Wrist camera image (224, 224, 3)
    'color_image2': Front camera image (224, 224, 3)
    'robot_state': {
      'ee_pos': EEF position (3,)
      'ee_quat': EEF orientation (4,)
      'ee_pos_vel': EEF linear velocity (3,)
      'ee_ori_vel': EEF angular velocity (3,)
      'joint_positions': Joint positions (7,)
      'joint_velocities': Joint velocities (7,)
      'joint_torques': Joint torques (7,)
      'gripper_width': Gripper width (1,)
    }
  }
'actions': List of 8-D actions
'rewards': List of rewards (1 if a furniture part is assembled; otherwise, 0)
'skills': List of skill completion flags (1 if a skill is completed; otherwise, 0)

FurnitureBench

Dataset

Contents