This document contains information for running Hand Manipulation Suite experiments reported in the paper. The current directory contains code and command scripts for our REvolveR.
We provide the expert policies trained from on the source robot. Please download the from the Google Drive link. The RL algorithms and tasks can be inferred from the directory names.
The definition of the MuJoCo models of the source and target robot are in make_generalized_envs.py.
To visualize the models and expert policies, please first download the expert policies files and change the values of generalized_env and policy_file in the code.
Then use python make_generalized_envs.py.
The command scripts for launching REvolveR are the command_train_revolver_*.sh files. For example, command_train_revolver_hammer.sh is the script for launching REvolveR on Hammer task.
After downloading the expert policy models from Google Drive, please change the value of policy and baseline in the command_train_revolver_*.sh files to the path of the policy and baseline files.
The users can easily develop their own code for visualizing the trained policies based on our implementation in the make_generalized_envs.py.