Final Projects Reinforcement Learning Applications WS21/22

The practical course gives an introduction to reinforcement learning using Unity. In a seven-day hackathon, the students can work full time on their projects in teams of up to 4 people. The fantastic results of the hackathon in the winter of 2022 are shown below.

Competitive Parking

Competitive Parking is a ML-Agent project with Unity where many car agents compete with each other to find a parking lot. The car agents have been successively trained with Curriculum Reinforcement Learning to adapt to increasingly large and more complex parking lots, and an increasing number of competing agents. Parking lots are procedurally generated and can grow to any size. While we do not want the cars to park perpendicularly and crash everywhere, the agents have been trained to park and drive fast. After all, we do not want the player to got any free space, do we. :)

Team

Parking was developed by
Rulu Liao, Human-Machine Interaction Master
Marcel Quanz, Informatics Master

Guide Dog

This project aims to explore the functionalities of Unity Machine Learning Agents (ML-Agents) by letting a dog guide a blind person through a 3D city scene. Rather than coding the behavior of the dog to guide the human safely, the dog is an intelligent agent which learns through reinforcement learning. The blind person knows the goal to go to next, but isn’t aware of the obstacles on the way. With the help of ray perception sensors the dog has an idea about its close surroundings and can help the human reach the goal. Collider and trigger are added to the city scene’s objects to detect if the dog misguides the human. By receiving rewards for reaching the human’s goal and staying close to the human the dog agent learns. The dog is “punished” for incorrect behavior. Due to the game’s navigation complexity the trained dog navigates through the city more quickly and elegantly than a human player.

Team

Guide Dog was developed by
Konstantin Hegestweiler, Master Computer Science
Hasan Turalic, Master Computer Science
Sina Schnebelt, Master Media Informatics
Pia Hammer, Master Media Informatics

Garbage Collector

The intelligent trash agents are programmed to collect the waste they are responsible for, depending on the type of waste. Their task is to steer through a predefined environment without collisions and to collect their rubbish without taking other objects, such as toys, with them. Another goal is to collect the rubbish as quickly as possible. (within an episode)

Team

Garbage Collector was developed by
Anna Fischhaber, Master Computer Science
Johanna Prinz, Master Media Informatics

KitchenRobo

Tired of cooking by yourself? KitchenRobo is here to help. Our kitchen robot brings items from the kitchen to a specified drop point after a voice command. During the movement of the robot, it is important that it does not touch the human, itself, the floor, or any kitchen furniture. Since the robot should learn independently how to move its six axes correctly and efficiently, we used reinforcement learning with Unity ML-Agents to train the model.

Team

KitchenRobo was developed by
Maximilian Tränkler, Master Computer Science
Jonathan Haudenschild, Master Human-Computer Interaction & Media Design
Marion Botsivali, Master Human-Computer Interaction

Drone Mechanics

In this project, we try to create an autonomous drone agent which controls the rotors, so that the drone can fly in a stable manner and also accomplish various tasks such as flying to defined locations in 3D space. The trained neural network is also eligible for deploying to microcontrollers that is beneficial for real drones.

Team

Drone Mechanics was developed by
Bora Kunter Sahin, Master Informatics
Janina Mattes, Master Informatics

RoboVolley

RoboVolley is a multi-agent game implemented with ML-Agent in Unity. In this project we have built robot arm agents that learns to control its six axes correctly to play volleyball against each other. The Agents were trained with both cooperative and competitive strategies by using reinforcement learning algorithm. The goal of cooperative strategy is preventing the ball fall on the ground and play with each other as long as possible. In contrast, when trained with competitive strategy, the individual agent aims to win the game and behaves more aggressively. For detecting the placement of ball and controlling the game episode, Collider and trigger are added to corresponding scene’s objects. Which agent starts, will be chosen interchangeably and the ball will be started at a random position within the correct side.

RoboVolley

Tennis was developed by
Florian Rieß, Master Computer Science
Heling Cai, Master Computer Science
Hyerim Park, Master human computer interaction
Mustafa Yasin, Master Computer Science