After the successful completion of the course, the student:
- Design, develop, test and evaluate deep learning algorithms for different tasks of autonomous robots
- Design, develop, test and evaluate reinforcement learning algorithms for different tasks of autonomous robots
- Experiment on how to design and embed these algorithms on on-board units hosted on the robot and extract real-time information for the autonomous deployment of the robot.
|
|
|
This course will discuss the principles of deep learning and reinforcement learning. It will introduce the use of Convolutional Neural Network (CNN) architectures with specific regard to supervised, semi-supervised, and unsupervised training techniques. Semantic segmentation of images (2D) and point clouds (3D) using deep learning algorithms as well as object detection and tracking using deep learning algorithms from images will be debated as specific applications. The potential of Visual SLAM, stereo and single-image 3D reconstruction using deep learning will be discussed too, with specific regard to their use in collision avoidance and autonomous navigation. The strategies on how to embed and optimize these algorithms on onboard units for their real-time deployment will be given. This course will also introduce the classic reinforcement learning algorithms, which employ a system of rewards and penalties to compel the autonomous robot to solve a problem by itself. In particular, Markov Decision Process and Q-learning and their applications in autonomous navigation will be presented.
|
 |
|