Adaptive AI Safety
Toward AI safety as a capability maintained throughout operation, not merely a property specified at design time.
Research at a Glance
My work follows a simple safety loop: anticipate change, adapt the safety mechanism, and assure behavior at runtime.
Predictive & Proactive Safety
Infer changing context, forecast safety-relevant conditions, quantify uncertainty, and construct constraints before risk materializes.
Adaptive Safety Constraints
Modify concrete safety requirements as operating conditions change while keeping them anchored to stable higher-level safety objectives.
Runtime Safety Control
Intervene when learned adaptation is insufficient through projected constraints, shielding, fallback behavior, and conservative runtime control.
Research Focus
My research focuses on AI safety for adaptive and autonomous systems, with particular interests in safe and continual reinforcement learning, safety under nonstationarity, runtime safety, and functional safety for safety-critical systems.
How can an autonomous learning system remain safe when its environment, knowledge, operating conditions, and the concrete constraints required for safe behavior change over time?
My long-term goal is to develop AI systems for which safety is not merely a static property specified at design time, but a capability maintained throughout operation. Such systems should recognize safety-relevant changes, anticipate emerging risks, adapt their safety mechanisms when appropriate, and act conservatively when reliable adaptation is not possible.
Current Research: Adaptive Safety under Nonstationarity
My current research focuses on safe continual reinforcement learning in nonstationary environments. Two questions are central: whether a safety constraint remains valid as operating conditions change, and whether the environment is changing faster than an agent can safely adapt.
I address these problems through adaptive safety constraints anchored to more stable safety objectives. My work combines context-aware constraint adaptation, adaptation-feasibility checks, proactive context forecasting, and runtime enforcement. I also develop evaluation methods that distinguish stationary, seen-nonstationary, and unseen-nonstationary regimes and measure constraint validity, physical safety, intervention burden, and adaptation behavior.
Selected Research
Concrete systems and methods that operationalize the broader adaptive-safety agenda.
Proactive Context-Forecasted Safety Constraints
Problem. Reactive adaptation can occur only after risk has already increased. Approach. Infer latent context, forecast its evolution, quantify predictive uncertainty, and construct anticipatory safety constraints.
Result: 87.2% collision-rate reduction on the main benchmark and 92–97% across three held-out road layouts.
LILAC+
Problem. Static safety constraints can become obsolete, overly conservative, or unsafe under nonstationarity. Approach. Combine context-based, adaptation-speed, and soft-to-hard mechanisms for policy adaptation and runtime protection.
Evaluation: reduced safety violations relative to unconstrained and fixed-constraint baselines across stationary and nonstationary regimes while maintaining competitive reward.
CPSS — Constraint Projection Safety Shield
Problem. Cumulative safety requirements do not directly prevent unsafe individual actions. Approach. Project the remaining safety budget into enforceable action-level runtime constraints.
Result: 91.6% average collision-rate reduction (0.082 → 0.007) versus an unshielded DQN baseline across four highway-env driving tasks.
SafeShiftRL
Goal. Provide a modular benchmark for evaluating safety under stationary, seen-nonstationary, and unseen-nonstationary conditions, with explicit measures of constraint validity, physical safety, intervention burden, and adaptation behavior.
Research Directions
My current work motivates several connected directions toward a broader methodology for adaptive AI safety.
Predictive and Proactive AI Safety
How can an AI system reason about future contexts and the future validity of its own safety assumptions? I am interested in forecasting, calibrated uncertainty, and anticipatory constraint construction that act before unsafe conditions arise.
Adaptive Runtime Safety and AI Control
I am interested in runtime monitoring, shielding, dynamic constraints, fallback behavior, and intervention policies that restrict autonomy when environmental change exceeds an agent's safe adaptation capacity.
Safety under Novelty and Unmodelled Change
A central challenge is detecting unfamiliar contexts, quantifying epistemic uncertainty, and deciding when learned predictions or safety constraints should no longer be trusted.
How should an autonomous system recognize the limits of its own safety model?
Long-Horizon Agent Safety
For increasingly autonomous systems, safety must account for accumulated risk over trajectories. I am interested in trajectory-level monitoring, dynamic safety budgets, and intervention policies that coordinate local decisions with longer-term requirements.
Multi-Agent Safety under Nonstationarity
Multiple learning agents create endogenous nonstationarity. I am interested in forecasting safety-relevant changes in other agents, adapting constraints to changing interaction regimes, and detecting emerging collective risks.
From Functional Safety to Adaptive AI Safety
My AI-safety research grows from a broader background in functional safety and safety-critical software-intensive systems, including autonomous and automotive systems. Classical functional-safety engineering contributes concepts such as specification, architectural assurance, runtime monitoring, fault handling, and safe fallback.
Learning-enabled systems introduce an additional challenge: system behavior and the conditions relevant to safety can continue changing after deployment. I am interested in bringing these perspectives together to develop safety-engineering principles for adaptive systems whose behavior and operating conditions cannot be completely specified or validated in advance.
Research Vision
My broader objective is to contribute to a science and engineering methodology for adaptive AI safety. Reinforcement learning provides a useful setting because learning, sequential decision-making, environmental change, and safety interact explicitly, but the underlying problem extends to increasingly capable autonomous and agentic AI systems operating over long periods in changing environments.
I aim to develop theory, algorithms, runtime mechanisms, and evaluation methodologies that connect adaptive learning with dependable safety assurance. The long-term goal is AI that can recognize when its assumptions are changing, anticipate safety-relevant consequences, adapt its safety mechanisms when justified, and remain conservative when reliable adaptation cannot be assured.
I want adaptation decisions themselves to be governed by explicit safety assumptions, evidence, and runtime controls.