Transfering virtual policies to robot rovers
In this student project, we examined the transfer of agent policies being trained in virtual environments and simulations to physical systems. The policy of the robot was trained using a simple two layer neural network which controlled the two motors of the robot an received rewards from a camera and two IR collision sensors. The learned policy was transfered to the robots raspberry pi and fulfilled the task of finding a red patch after several epochs of training on the virtual environment.
A video on the robots progress over multiple training epochs can be downloaded here: robot video
- Sebastian Schmoll
- Matthias Schubert