Intervention Design for Effective Sim2Real Transfer by

Melissa Mozifian, Amy Zhang, Joelle Pineau and David Meger.

[Paper] [Webpage]

If you use this repo in your research, please consider citing the paper as follows

@article{mozifian2020intervention,
    title={Intervention Design for Effective Sim2Real Transfer},
    author={Melissa Mozifian and Amy Zhang and Joelle Pineau and David Meger},
    year={2020},
    eprint={2012.02055},
    archivePrefix={arXiv},
    primaryClass={cs.RO}
}

ssh

git clone --recurse-submodules git@github.com:melfm/ibit.git

https

git clone --recurse-submodules https://github.com/melfm/ibit.git

For all python package dependencies run:

pip install -r requirements.txt

The hydra dependency gets updated frequently which can potentially break the config style, the version used is 1.0.3

pip install hydra-core

Install the third_party dependencies by running

bash setup_thirdparty.sh

The third parties have further dependencies. For instance Meta-World is based on MuJoCo, which has a proprietary dependency. Please follow the instructions in the mujoco-py package for help. And DMControl, there might be some differences for instance it expects the mujoco path to be ~/.mujoco/mujoco200_$PLATFORM/.

From ibit dir, you can run the followings

cd ibit

python train.py --config-name='configs/reach_v1'
python train.py --config-name='configs/button_press_v1'

Running DBC Agent This is the default agent but in case you want the experiment ID to reflect this:

python train.py --config-name='configs/reach_v1' agent.name=dbc agent.cls=agents.dbc_agent.DBCAgent

Running DRQ Agent

python train.py --config-name='configs/reach_v1' agent.name=drq agent.cls=agents.drq_agent.DRQAgent

Running Jaco dmsuite

python train.py --config-name='configs/jaco_reach_site_features' agent.name=drq agent.cls=agents.drq_agent.DRQAgent

Running with hyperparameter options

python train.py --config-name='configs/reach_v1' penalty_type='rex' env_resample_rate=50000 num_envs=4

• Make sure the viewing camera is set correctly per env. You can change this inside metaworld/envs/mujoco/mujoco_env.py Line: 83 - viewer_setup()
• If you run the same config file twice, things will be overwritten. To avoid this, make sure you overwrite some parameters which will create experiment file with appended overwrites.

• apply_mod : (bool) Applies random modifications to the env (for now only visual changes)
• num_envs : (int) Number of randomized environments
• internvention : (list) Type-1 -> Image Augmentation, Type-2 -> Rendering changes
• penalty_type : (str) Switch between IRM and REX penalties
• agent.name : (agent) DRQ, DBC

1. If running remotely, pass the flag

python train.py --remote_render=True

Although at the moment you still need to modify the mujoco_env.py GlfwContext instantiation. See this&this

2. On the host computer, export display for headless rendering and switch backend rendering engine to osmesa:

export DISPLAY=:0

and

MUJOCO_GL="osmesa" LD_LIBRARY_PATH=/usr/lib/x86_64-linux-gnu/nvidia-opengl/:$LD_LIBRARY_PATH

In general, if you notice hanging, check the rendering or check if the env is being created even. Sometimes the errors thrown from mujoco or OpenGL via metaworld are suppressed so you can check on the command line to see if you can import the environments and create the env.

raise MujocoException('Got MuJoCo Warning: {}'.format(warn)) mujoco_py.builder.MujocoException: Got MuJoCo Warning: Nan, Inf or huge value in QACC at DOF 0.
The simulation is unstable. Time = 1.4625.

This happens with DeepMDP Agent or DBC, they are meant to be run with more than one env so check num_envs is more than 1.

For now, hydra is slightly annoying with current file arrangement setup. Just create a symlink to jaco_physics which lives under jaco_real and run it from inside ibit directory. So from ibit, run:

ln -s ../jaco_real/jaco_physics.py jaco_physics.py