CQN-maniskill

This repository is the simple application of CQN agent for the ManiSkill2 Environment

How to set Maniskill2 demonstration data

To download the asset and demo datasets:

# Download the asset
python -m mani_skill2.utils.download_asset PickSingleYCB-v0

# Download the demonstration dataset for a specific task
python -m mani_skill2.utils.download_demo PickCube-v0
python -m mani_skill2.utils.download_demo PickSingleYCB-v0
python -m mani_skill2.utils.download_demo PegInsertionSide-v0
python -m mani_skill2.utils.download_demo PlugCharger-v0
python -m mani_skill2.utils.download_demo StackCube-v0

To convert the demo datasets to the rgbd image mode:

python -m mani_skill2.trajectory.replay_trajectory --traj-path \
demos/v0/rigid_body/PickCube-v0/trajectory.h5 --save-traj --obs-mode rgbd --target-control-mode pd_ee_delta_pose --num-procs 10

python -m mani_skill2.trajectory.replay_trajectory --traj-path \
demos/v0/rigid_body/PegInsertionSide-v0/trajectory.h5 --save-traj --obs-mode rgbd --target-control-mode pd_ee_delta_pose --num-procs 10

python -m mani_skill2.trajectory.replay_trajectory --traj-path \
demos/v0/rigid_body/StackCube-v0/trajectory.h5 --save-traj --obs-mode rgbd --target-control-mode pd_ee_delta_pose --num-procs 10

# For the task PickSingleYCB-v0 with cups
python run_traj_replay.py

The control mode can be changed to pd_joint_delta_pos or pd_ee_delta_pose depending on the task. The number of processes can be adjusted according to the number of CPU cores available.

How to train a model

To train a model for a specific task, please modify 'ms2_task@global' in "cfgs/config_maniskill2.yaml" The following is an example of how to train a model for the task PickSingleYCB-v0:

defaults:
  - _self_
  - ms2_task@_global_: picksingleycb_v0

Other task yaml files can be found in cfgs/ms2_task.

To train a model:

    python train_ms2.py

Single YCB Object Pick and Place results

Cube Stacking results

Reference

Continuous Control with Coarse-to-fine Reinforcement Learning

ManiSkill2