Prof. Sylvester Kaczmarek

CEO and Applied Scientist

Building AI to expand what humanity can discover and create

View of Earth and the Moon from the Artemis I mission, symbolizing the frontier of space exploration and trustworthy AI

Hardware can reach new worlds. What might we discover when machines can reason and act beyond our reach? My work connects advanced AI to the physical world.

Portrait of Prof. Sylvester Kaczmarek

My general fields are advanced artificial intelligence and cybersecurity for safety-critical cyber-physical systems. I like to start with the assumptions behind a design. If two sensors disagree, for instance, what makes us trust one over the other? The answer can change how we build the system.

As Featured In

Other publications

My work combines centralised safety assurance with distributed edge intelligence across spacecraft, robots, sensors, and computing nodes. These systems use onboard processing, robust planning, hardware-software co-design, and advanced computing architectures to operate under uncertainty, limited communications, constrained energy, and cyberattack.

This approach follows an established architectural pattern for complex safety-critical systems: centralised command provides strategic coherence and defines safety constraints, while decentralised execution provides adaptability and resilience in dynamic environments. Supported by onboard safeguards and verification, this architecture can enable increasingly capable autonomous missions while maintaining meaningful control and accountability.

Selected Work and Collaborations

Selected Non-Confidential Work

My work in this domain is supported by bodies including the UK Space Agency and the UK's National Cyber Security Centre (NCSC), and includes engagement with the U.S. Space Force's SDA TAP Lab. My contributions to artificial intelligence for space exploration were recognised with an Award of Merit from ESA and FDL Europe. My written evidence on investment and research commercialisation was cited in the House of Lords' report on the UK space economy . I also co-authored a 2026 CETaS research report at The Alan Turing Institute . I am also an Arm Ambassador and have published original research on SME2-accelerated satellite signal acquisition. My professional training and certifications include programmes from NATO, CISA, FEMA, IAEA, and the Cloud Security Alliance.

My work advances the safety, efficiency, and sustainability of space-based robotic intelligence, cislunar mobility, and autonomous physical-world operations. AI can propose more ideas than we have time to test. Which ones matter when those ideas must guide physical machines? I bring together advanced AI, verifiable control, and neuroscience-inspired computing to turn promising ideas into systems that can be measured and tested.

This research supports critical infrastructure and space exploration through work with organisations including NASA, ESA, UKSA, NATO, The Alan Turing Institute, the European Space Resources Innovation Centre, and the Luxembourg Space Agency. It also advances OrbiSky’s robotic intelligence programme and contributes to new standards for autonomous systems that can operate reliably and safely far beyond Earth’s immediate vicinity.

Selected Leadership, Research and Policy

For agencies, primes, and space companies, I provide a focused Trustworthy Onboard AI Audit covering autonomy architecture, assurance, cybersecurity, and mission readiness.

Click for full audit overview →

In parallel to these technical pursuits, my approach is aligned with the principles of machine alignment, including the integration of human values such as fairness, responsibility, and sustainability into AI systems. This involves developing systems grounded in verifiable mechanisms derived from first principles, with hard operational constraints enforced at the system's dynamic core. As AI becomes more capable, I think scientific judgement matters more. Which questions deserve our time, and what evidence would make us change our minds? My objective is to build intelligence that remains transparent, amenable to expert knowledge, and subject to meaningful human control.

I also emphasise ethical frameworks and stakeholder engagement throughout the development of advanced AI. Geoffrey Hinton highlighted the urgency of this challenge in his Nobel address, warning that advanced AI systems "are no longer science fiction" and that "we urgently need research on how to prevent these new beings from wanting to take control". This concern resonates strongly with my work. As autonomous intelligence becomes more capable and operates farther from direct human oversight, safety, alignment, and control must be treated as foundational engineering requirements.

The iconic 'Earthrise' photograph taken by William Anders during the NASA Apollo 8 mission, showing a fragile Earth rising over the lunar horizon.

Credit: NASA / William Anders, Apollo 8

The "Earthrise" photograph from Apollo 8, often called the most influential environmental photo ever taken. This view of our fragile home is the ultimate "why" for my work. It is a constant reminder that the intelligent systems we build must be embedded with the wisdom and responsibility this perspective demands.

Driving questions such as "How can we build robust and trustworthy AI for complex autonomous systems in space?" define the immediate focus of my work. A broader question also continues to shape my thinking: "When does consciousness emerge from AI's perceptual abilities?"

Through interdisciplinary research and collaboration, I seek to develop AI systems that are safe, secure, interpretable, and beneficial as their capabilities increase. These systems are intended not only for current missions, but also as foundations for a future in which increasingly autonomous intelligence can operate responsibly alongside humanity, including far beyond Earth.

This endeavour continues the challenge articulated by Alan Turing in 1947: "a machine that can learn from experience." It also builds on the hypothesis put forward in the 1955 Dartmouth proposal that learning and other aspects of intelligence could be described precisely enough for machines to simulate them. What draws me to those early ideas is the scale of the question those researchers were willing to ask. The question I want to spend my career on is what we could discover and build if we could trust machines to take on work beyond our reach.

The mission is just beginning.

Kaczmarek Editions

IN-ORBIT MANUFACTURING

SkyYield

Engineering a manufacturing payload for high-performance ZBLAN fibre in microgravity.

The challenge

ZBLAN has major potential for ultra-low-loss optical fibre, but fibre quality depends on tight thermal control and preventing crystallisation. A microgravity manufacturing payload must integrate thermal management, sensing and reliable process control within spacecraft constraints.

My contribution

As OrbiSky CEO, I led the SkyYield programme, bringing together payload architecture, manufacturing requirements, instrumentation, process control and commercial strategy. I directed the end-to-end engineering concept and verification approach for microgravity ZBLAN processing.

Programme deliverables

Under a UK Space Agency contract, SkyYield established the payload concept, process-control requirements and onboard measurement strategy for manufacturing ZBLAN fibre in microgravity.

ASSURED AUTONOMY

Making autonomous AI accountable

Verifiable limits, independent monitoring and intervention in mission-critical control.

The challenge

An AI model can perform well in testing while still making an unsafe choice under sensor failures, unseen inputs or conflicting objectives. Operational assurance requires control over what a model is permitted to do.

My contribution

I developed assurance methods that bound AI authority, monitor decisions at runtime, trigger fallback controls and capture verification evidence. My work includes a published probabilistic framework for neuromorphic safety monitoring and KRI-STD-001, a public onboard-AI assurance standard.

Research results

In a high-fidelity cislunar rover simulation combining adversarial and environmental faults, timely safety intervention improved mission success by 84 percentage points. The published framework also relates monitor response time and detection errors to bounded control-loop risk.

NEUROMORPHIC COMPUTING

Low-power AI for spacecraft

Radiation-resilient spiking neural networks evaluated on low-power Akida hardware.

The challenge

Spacecraft electronics must contend with tight power budgets, limited compute capacity and radiation-related faults. New AI architectures need demonstrable robustness as well as efficient inference.

My contribution

I developed and evaluated reliability models and protective methods for spiking neural networks, including radiation-inspired fault injection and hardware-in-the-loop experiments with event-driven computing hardware.

Research results

In Akida hardware-in-the-loop tests under severe simulated radiation-event rates, the Hierarchical Temporal Defense maintained an anomaly-detection F1 score above 0.72, compared with a baseline below 0.48, and increased mean time to failure by 4.5×.

FDL LUNARLAB

LunarFM

A multimodal representation of the Moon's surface for planetary science and resource prospecting.

The challenge

Lunar orbital observations are fragmented across instruments and missions. Resource mapping must combine imagery, spectral data, gravity, topography and thermal measurements, often with very limited labelled samples.

My contribution

As Research Lead for FDL LunarLab's Foundation Model for Lunar Resources, I shaped the research roadmap, guided multimodal data integration and evaluation, and co-authored the LunarFM paper with an interdisciplinary team.

Research results

LunarFM integrates 18 orbital data layers into 768-dimensional embeddings and compresses the input data by approximately 300×. A titanium dioxide resource-mapping study generated a global prediction using only eight expert-labelled Apollo samples.

ARM SME2

Accelerated Satellite Signal Acquisition

Faster satellite signal acquisition and diagnostics on phone-class Arm hardware.

The challenge

Satellite receivers need to acquire signals accurately across uncertain timing and carrier frequencies before frames can be decoded. Portable devices have limited time and compute capacity for that search.

My contribution

I developed and published a reproducible C++ satellite acquisition and diagnostics pipeline comparing scalar reference, Arm NEON and SME2 implementations. The workflow combines synthetic IQ replay, signal synchronisation, decoding, integrity checks and validation.

Benchmark results

On a vivo X300, SME2 reduced medium-workload acquisition latency from approximately 3.59 ms to 1.58 ms and delivered 2.14–2.42× speedups over NEON across five runs using reproducible test signals.

THE ALAN TURING INSTITUTE

Future AI Paradigms

New AI paradigms, national security and UK strategic readiness.

The challenge

Frontier AI advances may come from changes in architectures, learning methods and computing hardware. Policymakers need to evaluate promising directions and their national-security risks without assuming a single technical path.

My contribution

I conducted research for and co-authored The Next Frontier, a 2026 CETaS report at The Alan Turing Institute examining emerging AI architectures, learning approaches, computing hardware and their implications for national security and UK strategy.

Research findings

The report evaluates 15 emerging AI approaches against eight technical challenges and identifies UK priorities in specialist skills, research infrastructure and adoption of next-generation AI.