Episodic TD3

Introduce the movement primitives into actor-critic algorithms

Our framework is based on TD3 framework from Stable-baselines3 and ProMP framework from Autonomous Learning Robots (ALR) Lab.

Stable-baselines3: https://github.com/DLR-RM/stable-baselines3

Autonomous Learning Robots (ALR) Lab: https://alr.anthropomatik.kit.edu/

Description of the implementation structure

Name	Description
train.py	An example for training an environment.
enjoy.py	An example for enjoying an environment.
continue.py	An example for continue training an environment.

FOLDER config:

The yaml file which stores the parameters. When you want to use our framework, please add an environment configuration into context.yml or non_context.yml

FOLDER utils:

Create the environment, algorithm model, and the callback.

Load the hyperparameters from yaml file.

Please see train.py. It is an example for using utils.

FOLDER model:

The main structure of our algorithm.

Training and Sampling:

Name	Description
episodic_td3.py	The main algorithm framework of Episodic TD3.
base_algorithm.py	The base class of EpisodicTD3 class.

Actor:

Name	Description
detpmp_model.py	Build the ProMP model based on Gaussian Basis Functions.
deppmp_wrapper.py	Wrap the ProMP model to generate reference trajectory based on the weights.
controller.py	The controller of ProMP.

Critic:

Name	Description
td3_policy.py	Provide the critic networks, contextual actor network.
base_policy.py	The base class of td3_policy.py.
replay_buffer.py	The Replay Buffer with normalized timestep information