Skip to content

README.md

Latest commit

 

History

History
110 lines (86 loc) · 3.97 KB

README.md

File metadata and controls

110 lines (86 loc) · 3.97 KB

Using this repo to collect ALMI-X dataset. And we also upload our dataset at https://huggingface.co/datasets/TeleEmbodied/ALMI-X.

1. Select Motion

1.1 Create conda environments

create conda enviroment:

conda create -n almi-data_collection python==3.8
conda activate almi-data_collection
cd Data_Collection
pip install -r requirements_dc.txt

conda create -n create_text_annotation python==3.10
conda activate create_text_annotation
cd Data_Collection
pip install -r requirements_text.txt

1.2 Filter motions that are dominated by upper body movements

We reorganized the format of AMASS dataset and select required parts as:

  • motion_name.pkl: motion name of each motion; in KIT dataset, is like: "0-KIT_[idx][motion_simple_description][idx]_poses"
    • size: list[num_motions]
  • motion_dof_pos.pkl: dof pos of robot
    • size: list[num_motions] - list[num_frames] - numpy.ndarray[1] - numpy.ndarray[19](the order is following the H1-2 urdf(without 3 wrist joints for each arm))
  • motion_rg_pos.pkl: link position of robot
    • size: list[num_motions] - list[num_frames] - numpy.ndarray[1] - numpy.ndarray20 - numpy.ndarray[3](global position)
  • motion_root_rot.pkl: robot global orientation
    • size: list[num_motions] - list[num_frames] - numpy.ndarray[1] - numpy.ndarray[4](global orientation)
  • motion_root_vel.pkl: robot global linear velocity
    • size: list[num_motions] - list[num_frames] - numpy.ndarray[1] - numpy.ndarray[3](global linear velocity, x,y,z directions)
  • motion_root_ang_vel.pkl: robot global angular velocity
    • size: list[num_motions] - list[num_frames] - numpy.ndarray[1] - numpy.ndarray[3](global angular velocity, roll,pitch,yaw directions)

Put these files to Data_Collection/origin_data.

Filter satisfied motions from original dataset.

You can also download selected motions files at https://huggingface.co/datasets/TeleEmbodied/ALMI-X/blob/main/select_motions.zip, and unzip them into select_data.

cd Data_collection
conda activate almi-data_collection
python select_motion_first.py

Then the selected motions' data will be exported to Data_Collection/select_data following same format. And a new file named upper_motion_text remaining the text description for motion is also in select_data

1.3 Create text annotation for each motion

cd Data_collection
conda activate create_text_annotation
python save_text.py

Then the text descriptions will be saved in ./Data_Collection/text_no_expand, totally 34235 (41x835). And also ./Data_Collection/cmd/cmd_list.pkl and ./Data_Collection/npy_name_list.pkl used for collection trajectory dataset later.

1.4 Filter motions by hand

Then filter these 835 motions by hand:

  • delete motions which is illegible or excessive amplitude.
  • the subsequences of the actual completed actions in the original motion sequence are intercepted.
cd Data_collection
conda activate almi-data_collection
python ./mujoco/check_motions.py

Then there still 680 motions satisfying the requirements. A table file is saved in ./Data_Collection/mujoco/frame_text.csv, which the format is like:

264,396,cast a box
123,386,play the guitar using left hand
162,577,wave both
0,616,no
0,387,wave right
0,366,throw left
324,440,cast a box
...

where "no" in third column means delete this motion from dataset.

1.5 Expand the motions and texts

cd Data_collection
conda activate almi-data_collection
python get_motion_nums.py

This will create a Data_Collection/motion_num_dict.pkl file saving the number that each motion's expanding number. (If motion is deleted, the expanding number is 0).

Then expand the text description

cd Data_collection
conda activate almi-data_collection
python ./expand_text.py

Then the text total number is: 81549

2. Collect Mujoco Datasets

cd Data_collection
conda activate almi-data_collection
python ./mujoco/save_trajectory_data.py

The the trajectory data will be saved in ./Data_Collection/obs.