Deep Reinforcement Learning has been successfully applied in various computer games. But it is still rarely used in real-world applications, especially for continuous control of real mobile robot navigation. In this video we present our first approach to learning robot navigation in an unknown environment. The input to the robot is only the merged data from a 2D laser scanner and an RGBD camera and the orientation to the target. The map is unknown. The output is an action for the robot (speeds, linear, angular). The navigator (small GA3s) is pre-trained in a fast, parallel, self-implemented simulation environment and then deployed on the real robot. To avoid overfitting, we use only a small network and add random Gaussian noise to the laser data. The sensor data fusion with the RGBD camera allows the robot to navigate in real environments with true 3D obstacle avoidance and without environmental interventions, unlike other approaches. To further prevent overadaptation, we use different difficulty levels of the environment and train 32 at the same time.

Paper: https://arxiv.org/abs/2005.13857

Winner of the Show & Tell 2019 Master software projects at Gelsenkirchen University of Applied Sciences.
https://www.w-hs.de/informatik/aktuelles/neuigkeiten-und-termine/nachricht-lesen/news/detail/News/rueckblick-show-tell-2019/
https://www.w-hs.de/fileadmin/public/dokumente/erkunden/fachbereiche/FB3-Informatik/Veranstaltungen/ShowAndTell2019/RobLearn.pdf
http://www.trikon-online.de/fileadmin/mediadaten/pdf_ausgabe/Trikon2019_2.pdf page 33ff
… to be continued…

/"Background music by: https://www.youtube.com/watch?v3pZdZiPnsGs/"

Code by:
https://github.com/RoblabWh/RobLearn

Please take the opportunity to connect and share this video with your friends and family if you find it useful.