This repository is a fork from the original iGibson repository, with modifications to create a environment with a dexterous hand and a drawer for training tasks like drawer opening/closing.

Almost all files are the same as the original iGibson repository despite some additions and modifications. See my commit history to find all files that I have changed. I will described the major changes and additions below. For questions about iGibson related code, see the original github repo and iGibson documentation.

The legacy_3d folder in the root directory is my previous attempts of opening drawers using robot arm/free gripper/dexterous hand, with 3D observation instead of robot vision. One important file is hand_drawer_open_env_new.py where an environment with a dexterous hand and a drawer is implemented. The environment follows the OpenAI Gym API. We can use train_policy.py to use PPO to train a MLP with parallel sampling, and render_policy.py to render a trained policy. We use garage library implementation of PPO. Using the hyperparameters in the code, a decent policy can be trained, but feel free to try others. Other environments like gripper_drawer_open_env.py and sawyer_drawer_open_env.py can also be used but they are deprecated and training convergence is not guaranteed.

Two main files that define this environment are:

• gibson2/envs/hand_drawer_env.py: define the environment in OpenAI Gym style.
• gibson2/core/physics/robot_locomotors.py: the DexHandRobot class defines the dexterous hand as well as APIs about it.

Other subtle code changes can be found in the commit log. In examples/demo folder, we can use hand_drawer_example.py to test the environment.

The algorithms that are used to train the environment in iGibson are CURL and DrQ, which are both located in the root directory, and are copied from the open source implementation of their original papers. The code CURL and DrQ maintains very similar structure because they are both based on SAC-AE implementation. See the original repo of CURL, DrQ, and SAC-AE for more details.