Application of Learning from Demonstration to a Simulated Mining Inspection Task

Robotic inspection systems are of paramount importance in modern industries.
The mining industry is no exception and has a great need for
inspection of ore bodies and product. Unfortunately it has the added
problems of a potentially dangerous environment for both equipment
and human operators. This environment requires rugged robotic systems
to deliver appropriate sensor inspection instruments using robust
algorithms to minimise safety problems and maximise product throughput.

Teleoperation is a common method of controlling robots used in the
mining environment. It is time consuming and open to human controller
fatigue and consequent lack of concentration by the operator leading
to expensive and dangerous accidents. A method that supplants teleoperation
for some tasks is desirable.

In this thesis Learning from Demonstration (LfD) techniques are developed
and applied to a simulated mining tunnel inspection task. This task
has been proposed in this thesis as a simulated teaching task, where a
robot learned to follow a predetermined path mapped out by a human
teleoperator. The work covered in this thesis shows that it is possible
to replace teleoperation for some tasks. Specifically the thesis presents
the problems and possible solutions that may be encountered in applying
machine learning techniques to an inspection task in an underground
mining context. The application of LfD to train a robot allows a robot to
be quickly deployed in a dynamic environment with minimal retraining.
The thesis presents an improved stereo system for use in teleoperation
as well as an improved method for training robots using LfD. The work
in the thesis presents multiple algorithms based on Gaussian Mixture models, Continuous Hidden Markov models and Active Discrete Hidden
Markov models (See 5.2 in Chapter 5 ) to implement an LfD strategy for
an inspection robot in a confined environment. The results demonstrate
that certain LfD strategies are superior to others as well as presenting
new work in the implementation of these algorithms.