SYLVESTER KACZMAREK

ASSURED AUTONOMY / RESEARCH

Spacecraft
AI safety.

Verifiable limits, independent monitoring and fallback control for autonomous spacecraft and mission-critical AI. Research grounded in neuromorphic safety monitoring and cislunar rover simulations.

INDEPENDENT RUNTIME SAFETY MONITOR CONTROL PATH
AI CONTROLLER AUTONOMOUS SUPERVISOR POLICY + MONITORING ALLOW FALLBACK RUNTIME ASSURANCESAFETY DECISION
AI authority remains bounded by independent controlRESEARCH CONCEPT
+84 ppMission success in simulation
RuntimeDecision monitoring
FallbackSafety intervention

01 / THE CHALLENGE

Autonomy needs defined operating limits.

An AI model can perform well in testing while still making an unsafe choice under sensor failures, unseen inputs or conflicting objectives.

For spacecraft and robotic missions, assurance requires a way to limit permitted actions, detect unsafe decisions in time and retain control when the model can no longer be trusted.

02 / THE CONTRIBUTION

Independent monitoring and intervention.

I developed assurance methods that bound AI authority, monitor decisions at runtime, trigger fallback controls and capture verification evidence.

The work includes a published probabilistic framework for neuromorphic safety monitoring and KRI-STD-001, a public onboard-AI assurance standard.

03 / THE APPROACH

From AI decisions to controlled actions.

01 / DEFINE

Bound authority

Specify which control actions are permitted and the operational conditions under which they may occur.

02 / DETECT

Monitor at runtime

Assess model decisions and detect situations that call for intervention rather than continued autonomous operation.

03 / RESPOND

Fallback and verify

Trigger a safer control response and retain evidence of the detection and intervention for subsequent evaluation.

04 / RESEARCH EVIDENCE

Evaluated under combined mission faults.

A high-fidelity cislunar rover simulation combined adversarial and environmental faults to examine the effect of timely safety intervention.

In that simulation, intervention improved mission success by 84 percentage points. The published framework also relates monitor response time and detection errors to bounded control-loop risk.

The reported result is from simulation and should not be interpreted as spacecraft flight qualification.

MORE ABOUT THE RESEARCHER

View my wider work.

AI assurance, space systems, neuromorphic computing and applied research.

MAIN WEBSITE