Skip to content

Latest commit

 

History

History
46 lines (45 loc) · 2.45 KB

README.md

File metadata and controls

46 lines (45 loc) · 2.45 KB

RL in Manipulation

This folder holds the code for the manipulation RL experiment

Requirement

Please skip if you already installed the required packages when running the code under rot_est

  1. Install anaconda
  2. Create and activate conda environment 
    conda create --name equi python=3.9
conda activate equi
  3. Install PyTorch (Recommended: pytorch==1.10.2, torchvision==0.11.3): 
    conda install pytorch==1.10.2 torchvision==0.11.3 torchaudio==0.10.2 -c pytorch
  4. Install other required packages 
    pip3 install -r requirements.txt

Running Equivariant SAC in Block Picking

python main.py --env=close_loop_block_picking --num_objects=1 --alg=sacfd --model=equi_both_d --max_train_step=5000 --planner_episode=50 --device=cuda:1 --view_type=camera_side_rgbd --aug=t --seed=1

Running the baselines

  • CNN + RAD: 
    python main.py --env=close_loop_block_picking --num_objects=1 --alg=sacfd --model=cnn_sim --max_train_step=5000 --planner_episode=50 --device=cuda:1 --view_type=camera_side_rgbd --aug=t --seed=1
  • CNN + DrQ: 
    python main.py --env=close_loop_block_picking --num_objects=1 --alg=sacfd_drq --model=cnn_sim --max_train_step=5000 --planner_episode=50 --device=cuda:1 --view_type=camera_side_rgbd --aug=f --seed=1
  • FERM: 
    python main.py --env=close_loop_block_picking --num_objects=1 --alg=curl_sacfd --model=cnn_sim_2 --max_train_step=5000 --planner_episode=50 --device=cuda:1 --view_type=camera_side_rgbd --aug=f --seed=1
  • SEN + RAD: 
    python main.py --env=close_loop_block_picking --num_objects=1 --alg=sacfd --model=sen_fc --max_train_step=5000 --planner_episode=50 --device=cuda:1 --view_type=camera_side_rgbd --aug=t --seed=1

Running in Other Environments

Replace --env=close_loop_block_picking --num_objects=1 with:

  • Block Pushing: --env=close_loop_block_pushing --num_objects=1. Also replace --planner_episode=50 with --planner_episode=20 to run with 20 expert demos.
  • Block Pulling: --env=close_loop_block_pulling --num_objects=2. Also replace --planner_episode=50 with --planner_episode=20 to run with 20 expert demos.
  • Drawer Opening: --env=close_loop_drawer_opening --num_objects=1
  • Block in Bowl: --env=close_loop_block_in_bowl --num_objects=2. Also replace --max_train_step=5000 with --max_train_step=10000 to run 10k time steps.