| Abstract (Eng): | The creation of robotics that can evolve has been a hot topic for the last decade or two. This is because it can make the
workplace safer and more efficient environment, which will result in better competition for manufacturers implementing
robotics in their production processes. Current robots strictly follow programmed actions.
The goal of this thesis is to learn how difficult it is to stack blocks (cubes) onto each other. This will be done with Reinforce-
ment Learning in a simulated environment using MuJuCo. The environment consists of a table with a set of cubes, a robotic
arm and an RGBD camera to observe its state of it.
The Artificial Intelligence (AI) is implemented through a model-free Deep Q-Learning (DQL) algorithm, which incorporates a
Convolutional Neural Network (CNN) trained with depth camera data. The model is trained using multiple reward functions
to encourage specific behaviours, including placing blocks at higher altitudes than previous ones, making newly placed
blocks the highest points in the action space, the highest point or the number of neighbouring blocks and accounting for the
absolute height of the placed block.
Learning to stack cubes will give insight into how the robot can interact with the environment and understand how the
placement of these blocks affects structural stability.
|
|---|