Embedded Files
Gonzalo Villagrán Corrales
Supervised by Prof. Luca Iocchi
Introduction
Introduction
The motivation for this project is the use of RL to solve a generation trajectory problem in a
simulation platform in which we UAVs as the agents of the performance and the agents themselves
will learn how to generate these trajectories. The goal is to coordinate the agents performances
taking in consideration the obstacle avoidance problem and the interaction with the other agents
of the simulation. This approach to the problem of trajectory generation gives us a wide range of
possibilities to experiment with UAVs performances allowing us to develop different applications
and methodologies to reach the goal. In our case we decided to develop a method for the drone
performance coordination using reinforcement learning in an outdoor environment.
Agent
Agent
This UAV represents our agent in the performance. A multi-agent system can be defined in the platform PEDRA.
Environment
Environment
The environment is composed by elements as buildings and trees which can represent an obstacle for the UAV trajectory. The environment is build on the platform Unreal Engine.
Problem
Problem
The solution scope is to implement a Reinforcement Learning model for which the UAV can generate autonomously a trajectory taking in consideration the actuation protocols between present agents in the environment.
We have studied different situations in a multi agent environment for which the agents reach a solution for the proposed problem. A single Agent learning in a multi-agent situation and a simultaneous agent learning in a multi-agent situation.
Frameworks
Frameworks
AirSim is an open-source, cross platform simulator for drones, built on Unreal Engine developed by Microsoft. It can be used to experiment with deep learning, computer vision and reinforcement learning algorithms for autonomous vehicles.. PEDRA is a programmable engine for Artificial Intelligence applications. The goal of PEDRA is to create a tool that allow us to develop different kinds of algorithms that can be applied to a multirotor on a simulation.
Thesis
Presentation
Page updated
Google Sites
Report abuse