Implementation of Neural Policy Style Transfer with Twin-Delayed DDPG (NPST3) framework for Shared Control of Robotic Manipulators.
The algorithm can be executed with the policy_exec_nostyle_incremental.py script. Each numbered trained model in the repo corresponds to a different Style.
Trained Styles
0: Happy
1: Calm
2: Sad
3: Angry
For example: python policy_exec_nostyle_incremental.py --style 0 for the happy style.
You can install this repo running python setup.py install within the root of the repo. Python3 is required.
In case of compatibility issues you can check the requirements-stable.txt within setup.py for the last tested versions.
The Autoencoder is trained running the script conv_AE.py
The trained model can be executed with conv_AE_predict.py.
The TD3 network is trained with td3_st_train_no_style_incremental.py
The dataset is available here
In order to use it, download it and save it in the root directory of the repo within a new "dataset" folder.