Abstract
"Machine Learning (ML) is spreading into more application areas and facilitating a step change
in autonomous perception and comprehension capabilities. Within the space sector it is currently
used in the ground segment and its use in the space segment is being actively investigated. ML is
especially good at perception tasks that have traditionally been difficult for computers to master,
improvements in these perception capabilities have facilitated a wide range of new terrestrial
applications such as autonomous vehicles and drones as well as language comprehension and
translation. Deep space probes and rovers are reliant upon their on-board autonomy since
communication opportunities are sporadic, low bandwidth, and high latency. The levels of
autonomy these craft have directly affects their capabilities, enabling them to perform activities
without direct commands from ground controllers.
ML models have the potential to increase the level of spacecraft autonomy, expanding mission
capabilities, science returns, and returns on investment however two formidable barriers to
adoption exist. Firstly confidence in ML as a discipline and of the performance of specific
models is lower than that usually expected in the aerospace community, careful mission design
is required to demonstrate and especially utilise ML on active space missions. Secondly limited
processing power of space qualified radiation tolerant processors presents challenges not seen in
many terrestrial applications to date. On-board anomaly detection and robotic perception are two
applications where the use of ML on-board space vehicles and rovers is currently undergoing
active research and development.
To the authors knowledge this thesis presents the first investigation into the use of ML for the
estimation of terrain navigability on-board planetary rovers. The suitability of both Convo-
lutional Neural Network (CNN)s and encoder-decoder models is evaluated in terms of their
accuracy and computational performance using two planetary terrain data-sets. Their accuracy
was found to match that of existing state of the art navigability estimators, while surpassing
their performance. Deployment of ML models onto radiation hardened processors is identified
as barrier to adoption, since no software tools existed which targeted these processors.
Experimental tools automating the deployment and validation tasks are developed, which are
the first of their kind to target radiation hardened processors. Enabling new insights in the study
of low level ML implementation on current and next generation radiation hardened processors.
Using these tools novel techniques are found that significantly reduce the amount of memory
required to perform inference on a wide range of contemporary benchmark models, while using
state of the art techniques to optimise execution time. The memory requirement of MobileNet
v1 is reduced by 33% while MobileNet v2 is reduced by 53%. The impacts of new techniques
are been analytically characterised, and automated tools developed which allow rapid evaluation
and adoption in the wider ML community.
New techniques discovered during this work are informing the current development of the Mars
Sample Fetch and Return rover at Airbus as well as other machine learning groups in the space
industry. These tools are enabling transfer of existing techniques from the ML community into
the space sector. While these methods have been developed for space applications on radiation
hardened processors, they are equally applicable to low power terrestrial computing. The spread
of ML onto micro-controllers in embedded Internet of Things (IoT) devices is using these
techniques impacting the performance a wide range of applications outside of the space sector."