ECE_7202_Final_Project

1. Path planning using value iteration

If both the map and model are know, we use value iteration to find the optimal value functions V* and the optimal path is derived using a greedy policy.

Model

Action space S: up/down/left/right/stay​

Reward: 0 for each action; +10 if enter the goal position (discounting factor = 0.99)​

Task ends when the agent reaches the goal

Single agent, static enviroment

Single agent, dynamic enviroment

The location of dynamic obstacles is initialized randomly, and move randomly for each step​

Robot is aware of the nearby obstacles, but doesn't know where they will move

2.Path planning using TD method

If the map is unknow, the robot needs to explore the enviroment and find the optimal path

Single Agent (SARSA/Q-learning, n-steps SASA/Q-learing)

Mulitple Agents (simultaneously Q-learining)

3. Real application

Preprocess the image learned path