metadata
dataset_info:
features:
- name: observation.state
sequence: float32
length: 6
- name: action
sequence: float32
length: 6
- name: observation.images.webcam
dtype: video_frame
- name: episode_index
dtype: int64
- name: frame_index
dtype: int64
- name: timestamp
dtype: float32
- name: next.done
dtype: bool
- name: index
dtype: int64
splits:
- name: train
num_bytes: 4736267
num_examples: 35051
download_size: 1729226
dataset_size: 4736267
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
Move a blue cube and a pink cylinder between target zones.
Always start and finish with the blue cube and pink cylinder on the right, with the blue cube closer to the arm. This allows us to generate demonstrations without having to intervene for environment resets.
Always start and finish with the arm in rest position.
Always move the blue cube across first. This makes it so that the task is Markovian.
Only do one cycle per episode.
First 25 episodes with only natural lighting. Last 25 episodes with added warm LED room lighting.
The orientation of the gripper (thumb-left as seen in the video) is consistent throughout.
