This repository contains the scripts and results of running HER+DDPG experiments on the shadow dexterous hand.

The Report is available in the reporsitory with the name AER1517_Project.pdf, while the video is available here.

Please follow the instructions to install the OpenAI gym in your system. As part of the OpenAI gym installation, make sure to install the mujoco_py wrapper as well with the provided instructions.

The OpenAI Baselines can be installed by following the instructions here. Make sure you are able to run the example 1 given in the instructions that runs PPO with mujoco humanoid.

Copy the the run_shadow_experiments.sh into the baselines folder.

Run the following command to train all 6 agents for 200 epochs.

$ bash run_shadow_experiments.sh -b <path to baselines folder>

In order to plot the results from the logs run the following python script

$ python ./logs/hand_manipulator_lc_plotter.py -i < Dir path [touch logs]> -i <Dir path [no touch logs]> -t <title> -o <output file path>

The run_shadow_experiments.sh script can also be used to visualize the agent behavior. The following command template runs the the provided trained model in the provided mojovco environement. You can run bash run_shadow_experiments.sh -h for the usage message.

$ bash run_shadow_experiments.sh -b <path to baselines folder> -r HandManipulateBlock-v0 -f <path to baselines>/models/HandManipulateBlock-v0_her_2M_t0