Date & Time: June 11th, 14:00-15:00
Location: Seminar Room M 3170-73 at Dept. of Automatic Control, LTH
Author: Teodor Åstrand
Title: Robot Reinforcement Learning for Object Isolation
Supervisor: Yiannis Karayiannidis, Dept. Automatic Control, LTH
Examiner: Björn Olofsson, Dept. Automatic Control, LTH

Abstract: This thesis employs deep reinforcement learning, a branch of machine learning, to carry out robotic tasks. The objective centers around teaching an agent controlling a 7-axis robot arm with a gripper tool, to complete an object isolation task. For this task, a robot manipulates a cluttered environment in such a way that a predetermined target object becomes isolated. Sub-tasks were developed to explore simpler robot tasks to evolve and combine them into more complex tasks, where the goal was the object isolation task. Agent training took place in a simulated robot learning environment with the use of primarily a proprioceptive low dimensional state-space, where reward-shaping was the primary tool to teach a given task.

The reinforcement learning algorithm Proximal Policy Optimization (PPO) implemented with a neural network architecture was used to train agents for the robotics tasks and the robot arms joint velocities were used as the action-space the agents. Multiple experiments were conducted for agents practicing different tasks and their performance was evaluated by measuring their task completion success rate and rendering their behavior among others.