In this project deep reinforcement learning is used to train multi-agent robotic systems to perfrom leader-follower formation control . The OpenAI's MADDPG environment is used after some modifications have been added for agent training.
The Framework used in this project is Python 3.6.9 installed on Ubuntu 18.04 LTS. Alongside with Numpy, Tensorflow.
The environment is the single most important element in the Reinforcement Learningprocess since it presents the physical world that the agent interacts with. In this project the environment used is Multi-Agent Particle Environments (MPE). based on OpenAI work. OpenAI is artificial intelligence research laboratory that develops free open-source tools and libraries that helps the Artificial Intelligence developers community in the Research and Industry fields. The original environment is a 2D world with a continuous observation and discreteaction space, along with some basic simulated physics. It was developed such that the agents are divided into 2 groups: Good Agents, Adversary Agents. Such that the good agents try to cooperate to cover certain goal landmarks so that the adversary agents can not cover these goals.
Many modifications are made to this environment in this project so that it can be used in the Leader-Follower Formation Control favor, including: