Author: CAO RUI
A implementation for solving reach target task based on Twin Delayed DDPG(TD3) with Hindsight Experience Replay(HER) using PaddlePaddle.
Python: Python 3.6+
PaddlePaddle : Deep learning framework
PARL : Reinforcement learning toolbox based on PaddlePaddle
gym : Universal environment builder for RL tasks
RLBench: RL tasks extension for robotics researches.
First, create a virtual environment by virtualenv, in it, install PaddlePaddle, gym and PARL by
pip install requirements.txt
Then install RLBench via RLBench.
python rlbench_reach_td3_train.py
python rlbench_reach_td3_eval.py
4 stages (initial model, after 40000-episode trained model, after 80000-episode trained model, fianl model) of training model are uploaded and the corresponded render results are recorded in the folder records.
The success rate as shown in the following figure, here every epoch equals to 200 training episodes.
