I apply DDPG+HER for a Nachi_mz07 robot arm with ROS, Gazebo and OpenAI_ROS framework My work is the extension of the work of Tianhong Dai: https://github.com/TianhongDai/hindsight-experience-replay
I changed the output from torque value to delta position value. I want the robot to learn how to move based on position controller instead of torque controller. It's more safety in an industrial application.
I changed the HER buffer in TianhongDai/OpenAI work from a static to dynamics size, it is now applicable to the case of early ending episode
- ROS Noetic
- Python 3.8
- Pytorch 1.12.0 + Cu116
- Ubuntu 20.04.5 LTS focal
- Gym 0.21.0
- MoveIt 1
- Will be added later