AI Verification and Validation
for Trustworthy Systems

Rare-event estimation, sequential decision making, and systems assurance for autonomous vehicles, energy systems, and critical infrastructure.

Explore Research Apply to Join

What We Study

The AI V&V Lab investigates how much trust an AI system has actually earned before it is allowed to run something expensive, critical, or dangerous — and how to produce the evidence that proves it.

Safety Validation and Rare-Event Estimation

Estimating failure probabilities for black-box, learning-enabled systems at rates where naive Monte Carlo is intractable. Adaptive importance sampling with normalizing-flow and diffusion proposal distributions, and estimators that carry finite-sample guarantees rather than empirical pass rates.

Sequential Decision Making under Partial Observability

MDP and POMDP formulations, online planning, and stochastic programming for decisions taken on incomplete information — from geological uncertainty in mineral supply chains to socially-equitable energy allocation — with provable performance or regret bounds.

Systems Assurance and Certification

Machine-readable verification and validation specifications, requirement-to-test traceability, and runtime monitoring, following the traceability obligations imposed by standards such as DO-178C and ISO 26262 — structured so an assurance argument can be re-executed against a changed system rather than re-asserted.

View all research areas

Current Projects

Some selected current projects led by our lab members, supported by various collaborators and sponsoring organizations.

Trustworthy AI via V&V, Explainability, and Certification

A certification authority needs a traceable safety argument, not a benchmark score. In this project, we build and document that end to end: machine-readable V&V requirements in vnvspec, a scriptable closed-loop testbed in Duckietown.jl, and a runtime monitor that reasons over inter-vehicle influence paths to contain a spoofed message with roughly 3x fewer quarantines than blanket isolation. Our studies also showcase the potential and challenges of trustworthy AI in various context, including capability loss and recovery in compressed driving policies and auditable Kellgren-Lawrence grading in Knee-xRAI.

vnvspecDuckietown.jlAV Failure PathsCompressed AV Capability LossKnee-xRAI

AI Safety & Verification/Autonomous Systems & Robotics

Robust Decision Making under Uncertainty

Autonomy should not make failure more frequent, or more surprising. We study how to design safe policy, e.g., by revisiting the use of Dijkstra for stochastic shortest paths, tracking nonnegative reduced costs with one confidence bound per state-action pair and no transition kernel at all. We also devise safe planning for human-robot collaboration in Petro-HRCD-FLP, keeping human-robot supervision ratios and service-level agreements inside the optimization rather than patching them on afterwards.

DORADORASolvers.jlPetro-HRCD-FLP

Optimization & Decision Making/Autonomous Systems & Robotics

Scalable Rare-event Simulation

Rare failures are expensive to estimate. Because of that, crude Monte Carlo often needs millions of runs or even billions to see one, and its relative error grows as the event gets rarer. Our work uses importance sampling, and extends it to learn a proposal for a black-box system and relaxes it into a certified upper bound on failure probability, instead of a sinbgle point estimate that depends on the sampler having found the right tail. We also explore other variance reduction techniques, including control variates, which cut an EV charging network design's sample size tenfold, cheap enough to put station reliability inside a standard MIP.

Rare-event SPCControl Variate EV Charging

Simulation & Rare-Event Analysis/AI Safety & Verification

Optimization Models in Supply Chains, Critical Minerals, and Energy Systems

If a neural network sets your reorder quantities, what is the worst it can do to you? A forecaster inside a multi-echelon chain is a composed piecewise-linear function, so the machinery built to verify image classifiers can certify inventory policy against demand shocks. To simulate this, we build deepbullwhip, an open benchmark for complex supply chains. We also study how to model lithium supply chain under geological uncertainty, forced co-production in polymer planning, stroke diagnosis sequencing, and the various other societal applications.

deepbullwhipLithium POMDPCoupled Plan OptDSA-POMDPLLM ePuskesmas

Supply Chain & Logistics/Critical Minerals & Energy Resilience/Optimization & Decision Making

Recent News

2026

Selected Publications

View all publications

Interested in joining?

We are looking for motivated researchers at all levels: postdocs, PhD students, master's students, interns, and visiting faculty.

Apply to join