REPOSITORY > RESULTS

Abstract

The pursuit of autonomous robot learning is aimed at enabling robots to seamlessly integrate
into our daily lives, enhancing our efficiency and convenience. On the other hand,
in the industrial sector, autonomous robot learning plays a vital role in keeping up with
the rapidly evolving market demands. As products and processes change swiftly, robots
equipped with autonomous learning capabilities can quickly adapt and optimize their performance,
leading to increased productivity and competitiveness.
In this thesis, robot behaviour generation is considered at different levels of complexity.
Initially, we address autonomous optimization methods, then progress to autonomous
adaptation, and finally explore the potential of fully autonomous behaviour generation.
Installation or even just modification of robot-supported production and quality inspection
is a tedious process that usually requires human expert engagement. The resulting
parameters, e.g. robot velocities specified by an expert, are often subjective and suboptimal.
That is why, in the first part of the thesis, we demonstrate an autonomous trajectory
optimization method for the repetitive task. We propose a new approach for specifying
visual inspection trajectories based on CAD models of workpieces to be inspected. The
expert is required only to select the desired points on the inspection path along which the
robot should move the camera. The rest of the approach is fully automatic.
In the second part of the thesis, we extend the autonomy of robot behaviour to adapting
the whole industrial process. Robotic visual inspection is based on manually pre-defined
postures between the camera and the object, and often a robot has to move the in-hand
camera around of the object. The path of the robot is typically always the same, determined
in advance. But, in order to check only the selected subset out of all possible aspects of
the product, we need to generate all possible transitions between all possible aspects. This
is a slow and tedious process. Therefore, we propose a method able to use neural networks
in order to classify parts only according to possible defects. The classification results
allow for the exclusion of certain viewpoints, effectively saving time on the production line.
The emphasis lies on the adaptive nature of this industrial task, showcasing the ability to
accelerate processes and improve overall efficiency in industrial settings.
Trajectory generation for dynamic tasks, which involve complex problem-solving and
require more than simple motion repetition, is explored in the third part of the thesis. Due
to the time-consuming nature of robot learning, simulations are often employed to accelerate
the process. While complete skill learning in simulations is extensively researched
and falls outside the scope of this part, our focus lies on effectively transferring knowledge
from simulation to the real-world. Initially, a deep neural network encodes the skill in one
domain, which is later adapted for a target domain by partial retraining using the real
data. Our proposed approach combines backpropagation and reinforcement learning for
retraining, demonstrating significant speed improvements compared to individual methods
and a tenfold acceleration compared to learning from scratch. As an example for this
approach, we apply it to the challenging task of robot throwing, which involves dynamic
elements and is not reliant only on the final position of the robot motion.