ICRA2023 SEIL: Simulation-augmented Equivariant Imitation Learning

PDF | Website | Video

Instruction

1. Clone this repo

git clone https://github.com/SaulBatman/SEIL.git
cd SEIL && pip install -r requirements.txt

2. Clone environment repo and export PYTHONPATH

git clone https://github.com/SaulBatman/BulletArm.git -b ibc
export PYTHONPATH=$PYTHONPATH:/YOURPATH/BulletArm/

3. Run simulation experiments in Block in Bowl

cd SEIL/scripts
# run SEIL
python main.py --env=close_loop_block_in_bowl --model=equi_d
# run CNN w/ TS
python main.py --env=close_loop_block_in_bowl --model=cnn
# run equivariant BC
python main.py --env=close_loop_block_in_bowl --model=equi_d --simulate_n=0
# run CNN baseline
python main.py --env=close_loop_block_in_bowl --model=cnn --simulate_n=0
# run implicit BC baseline
python main.py --env=close_loop_block_in_bowl --model=ibc --simulate_n=0

4. Run real-world experiments in Block in Bowl after unzip data into /scripts/buffer(dataset link)

cd SEIL/scripts
# run SEIL
python main.py --env=real_close_loop_block_in_bowl --model=equi_d --load_buffer=/YOURPATH/SEIL/scripts/buffer/bowl.npy
# run CNN baseline
python main.py --env=real_close_loop_block_in_bowl --model=cnn --load_buffer=/YOURPATH/SEIL/scripts/buffer/bowl.npy
# run implicit BC baseline
python main.py --env=real_close_loop_block_in_bowl --model=ibc --load_buffer=/YOURPATH/SEIL/scripts/buffer/bowl.npy

Note

• All saved models and parameters are defaultly saved in YOURPATH/SEIL/scripts/outputs
• There are other choices of environments (replace --env=close_loop_block_in_bowl):

  --env=close_loop_block_stacking
  --env=close_loop_shoe_packing
  --env=close_loop_drawer_opening
  

• You can select number of demonstrations by adding the following argument (DEFAULT=10):

  --planner_episode=INT_NUM
  

• To visualize a real-time pybullet renderer by adding the following argument:

  --render=t
  

• If you are using your own real-world data, please make sure that transition simulation has the access to your point-cloud-to-image projection function.