Deep Reinforcement learning for real autonomous mobile robot navigation
HTML-код
- Опубликовано: 4 ноя 2018
- 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 learn robot navigation in an unknown environment. The input for the robot is only the fused data from a 2D laser scanner and a RGBD camera and the orientation to the goal. The map is unknown. The output is an action for the robot (velocities, linear, angular). The navigator (small GA3s) is pretrained in a fast, parallel, self implemented simulation environment and then deployed to the real robot. To avoid over fitting we use only a small network and we 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 real 3D obstacle avoidance and without environmental interventions in contrast to other approaches. To further avoid over fitting we use different levels of difficulties of the environment and train 32 of them simultaneously.
Paper: arxiv.org/abs/2005.13857
Winner of the Show & Tell 2019 Master software projects at the University of Applied Science Gelsenkirchen.
www.w-hs.de/informatik/aktuel...
www.w-hs.de/fileadmin/public/...
www.trikon-online.de/fileadmin... pp. 33ff
... to be continued...
"Background Music by: • ♩♫ Epic Chase Music ♪♬... "
Code at:
github.com/RoblabWh/RobLearn 
Наука
Great production value on the video!
Nice work done!
Inspirational!
Nice intro!
Good work! Can you please tell me which Software did you use for Simulation?
on which simulator you trained your agent?
so robot doesn't load the map in map_server ? amazing !
Cool! How do you replace the navigation stack or put your RL planner on the real robot?
On the robot
Hey, Great Work I just wanted to know whether can this be achieved without the use of laser scanner, just by achieving basic ultrasonic sensors?
yes an no. Laser scanner are much faster and more dense then ultrasonic sensors. The simulator should be updated to this constraint and the resulting network would be different from that with the laser scanner. The principal approach could work.
So basically I'll be just using my camera for navigation and the ultrasonic sensors are merely for avoiding crash?
And can I achieve the same thing that you've done just with camera?
Hey Mukesh, now you can get good quality LiDAR at cheap rate at Centauri Robotics, we also have our tutorial series centaurirobotics.in/dev/
Good work, how it know its target
Hi. Great work. After training an saving your model, how did you use it with the real robot?
O robô precisa de um cérebro para rodar o algoritmo. No caso ele usa uma Nvidia Jetson TX1. Então ele treinou no ambiente virtual, e depois implementou o algoritmo no Nvida Jetson TX1 para usar o aprendizado no robô real.
NIce work, Can you share the git repo please for us.
good work please can you send to me the code source of simulation
Code at:
github.com/RoblabWh/RobLearn
Hello, I am a senior in IT department and my graduation project is almost the same of your Master's degree project
and i need your help if you are not busy,and need any resources or videos to understand your code....
any publication?
is under review
May I know how to apply the trained model in real autonomous mobile robot navigation? which file should I refer to ?
lol die musik