Space West Spotlight - UWE: AI for Autonomous Systems
Welcome to Space West Spotlight, a series celebrating the people, research and businesses that make up the Space West community.
Through this series, we'll profile the individuals driving innovation, the pioneering research shaping the future and the local businesses powering growth across our region.
This edition shines the spotlight on the University of the West of England. We would like to thank Dr Jonathon Lancelot for sharing this research with us.
What is the focus of your research?
My research focuses on resource-efficient and verifiable AI for autonomous systems. I am interested in a fundamental question: how much useful intelligence can we place at the edge without continually increasing the computation required to achieve it?
I approach this through neural-network design, computer vision and anomaly detection. My current research includes Dynamic Fractional Stride (DFS), binary neural networks (BNNs), and architectures for autonomous visual perception and anomaly detection. The aim is to make AI systems computationally lighter while retaining the ability to recognise and respond to meaningful changes in their environment.
Space is a natural proving ground for this research. A spacecraft cannot always assume continuous connectivity, abundant power or access to large computing infrastructure. Intelligence therefore needs to operate closer to the sensors and increasingly onboard the system itself. These constraints are particularly pronounced in space, but the same underlying problem exists in aviation, robotics, autonomous vehicles and other safety-critical cyber-physical systems.
What challenge or opportunity does it address?
There is a tendency in AI to approach increasingly difficult problems with larger models and more computation. That becomes much harder when AI leaves the data centre and operates inside a physical system.
At the edge, computation, memory, energy and communications all become constraints. Autonomous systems may also encounter situations that were not explicitly represented during training. In safety-critical environments, we therefore need to consider more than accuracy alone: computational cost, robustness, anomaly detection and how system behaviour can be verified are also important.
My research sits at that intersection: autonomy, computational constraint and verification. Rather than treating limited computing resources as a problem to solve after an AI system has been designed, I treat the constraint as part of the architecture from the beginning.
What makes this work novel or significant?
One direction I am investigating combines Dynamic Fractional Stride with binary neural networks as part of a wider minimum-compute approach.
BNNs are attractive because binary representations can substantially simplify neural-network operations. That efficiency, however, introduces trade-offs in training and information representation. My work investigates whether DFS-based approaches can help manage information differently within these constrained architectures and contribute to more efficient learning and operation.
My computer-vision research approaches the problem from another direction. Rather than requiring an autonomous system to know every possible object or abnormal event beforehand, anomaly-detection models can learn representations of an expected environment and identify when observations begin to depart from it.
These strands share a common design philosophy. Rather than making an AI system large and then asking how it can be squeezed onto smaller hardware, I want to investigate what happens when computational efficiency is part of the neural architecture from the start.
How does it benefit industry or society?
Reducing the computation required for useful autonomous intelligence could expand the range of platforms on which AI can realistically operate.
For space systems, this could mean more processing taking place onboard, closer to the sensors generating the data. A spacecraft could potentially identify anomalous information locally and prioritise what needs to be processed, communicated or examined further.
This is particularly relevant where power, communications and computing capacity are limited, including smaller spacecraft and distributed autonomous platforms where simply adding more computing hardware may not be the best engineering solution.
The principle also extends beyond space. Aircraft, robots, autonomous vehicles and industrial systems face variations of the same problem. Reducing unnecessary computation also has the wider potential to reduce the energy and hardware required to deploy AI.
What are the next steps for your research?
The next step is integration.
I have been developing different parts of this problem through neural-network design, anomaly detection and computer vision. The objective now is to bring these components together into demonstrable architectures and evaluate them under realistic computational constraints.
That means measuring more than whether a model is accurate. I want to understand the relationship between accuracy, computational cost, robustness, latency and the ability to detect when something unexpected is happening.
Space systems provide an excellent environment in which to test these questions because the constraints are real and demanding. The underlying methods, however, are deliberately being developed for transfer to other autonomous and safety-critical systems.
What opportunities are there for collaboration or knowledge transfer?
This research benefits from collaboration because laboratory experimentation alone cannot reproduce every constraint found in a real autonomous platform.
I am interested in working with organisations developing spacecraft, satellites, onboard computing, sensors, embedded AI, computer vision, robotics and other autonomous or safety-critical technologies. Collaboration could involve representative hardware, datasets, simulation environments, engineering use cases or opportunities to test experimental architectures against real operational requirements.
I am particularly interested in hardware-software co-design. If the objective is genuinely resource-efficient AI, there is an opportunity to consider the neural architecture and the hardware that executes it as parts of the same engineering problem.
Ultimately, I want to build systems that can be demonstrated and tested, not simply described theoretically. Space is one of the most demanding environments in which to do that, and lessons learned there can transfer into many other industries.