Autonomy is driving today.
We model the risk.

Surveys the technological challenges of surveillance, decision making, and validation in aircraft collision avoidance systems, presenting their regulatory-accepted solutions as instructive case studies for safety-critical systems more broadly.

Develops a suite of methods for safe decision-making under uncertainty, combining learned surrogate models with belief-space policy iteration, chance-constrained planning, and Bayesian failure-probability estimation for black-box safety-critical systems.

Presents a probabilistic-inference framework for validating safety-critical autonomous systems, using adaptively trained deep generative models and state-dependent importance sampling to efficiently sample failures and estimate their probabilities.

Introduces DB-KSVD, a scalable alternating-optimization dictionary learning algorithm that disentangles high-dimensional transformer embeddings into interpretable features, matching sparse-autoencoder performance for mechanistic interpretability at millions of samples.

Proposes Kov, a black-box LLM red-teaming attack that frames jailbreaking as a Markov decision process and uses Monte Carlo tree search over perplexity-regularized adversarial suffixes to elicit naturalistic, transferable harmful behaviors.

Introduces a chance-constrained POMDP planner that learns neural network surrogates of value, policy, and failure probability to guide safe Monte Carlo tree search, achieving target safety levels without hand-tuning reward-cost tradeoffs.

Introduces ConstrainedZero, a chance-constrained POMDP policy-iteration algorithm that learns a neural failure-probability surrogate and uses adaptive conformal inference within Monte Carlo tree search to meet target safety levels without trading off rewards against costs.

Presents three probabilistic estimators for real-time pose estimation from image features, along with closed-form multivariate calibration and sharpness metrics, enabling well-calibrated visual predictions for safety-critical aircraft systems.

Presents a Bayesian optimization approach to black-box safety validation that uses a probabilistic surrogate model and three acquisition functions to estimate failure probabilities of safety-critical systems with far fewer samples than baseline methods.

Introduces a method for uncovering heterogeneous effects in computational sustainability models, enabling decision-makers to identify how outcomes vary across conditions rather than relying on aggregate averages.

Frames estimation of the failure-trajectory distribution for sequential autonomous systems as Bayesian inference, using differentiable dynamics rollouts and Hamiltonian Monte Carlo to improve sample efficiency and parameter-space coverage over black-box baselines.

Develops backward reachability techniques for neural feedback loops that leverage forward neural network analysis tools to efficiently over-approximate backprojection sets, certifying safety for linear and nonlinear closed-loop systems with neural network controllers.

Presents closed-loop analysis methods for designing and formally verifying safe machine learning-based perception systems that process high-dimensional camera images for safety-critical tasks previously performed by human operators.

Introduces smoothing bandits, a black-box estimation method combining Gaussian processes and threshold bandits, to efficiently determine sets of safe perception-system performance requirements from high-level closed-loop safety specifications.

Presents AVOIDDS, a photorealistic benchmark for vision-based aircraft detect-and-avoid that pairs a 72,000-image labeled dataset spanning varied environmental conditions with a closed-loop simulator for evaluating object detectors on the downstream collision avoidance task.

Proposes a reinforcement learning method that uses physics-based LiDAR data augmentation to stress-test autonomous vehicle perception, finding high-likelihood tracking and trajectory-prediction failures under adverse weather with smaller input disturbances than baselines.

Proposes training a generative adversarial network to map states to plausible input images, yielding a low-dimensional input space so existing closed-loop verification tools can provide formal safety guarantees for image-based neural network controllers.

Develops a risk-driven approach to designing perception systems that quantifies how perceptual errors affect closed-loop safety, reducing collision risk by 37% on a vision-based aircraft detect-and-avoid task.

Presents ZoPE, a fast optimizer over the outputs of low-dimensional-input ReLU networks that eagerly splits the input space and bounds objectives with zonotope propagation, achieving large speedups over mixed-integer programming baselines.

Develops an MDP-based collision avoidance logic that issues aircraft speed change advisories and evaluates it via Monte Carlo simulation, finding that speed advisories reduce collision risk but are neither as safe nor as efficient as horizontal and vertical maneuvers.

Introduces a method for verifying inverse-model neural networks by overapproximating a nonlinear stochastic forward model with piecewise-linear constraints and encoding both as a mixed-integer program, demonstrated on an airplane fuel-gauge case study.

Surveys algorithms for black-box safety validation of cyber-physical systems through simulation, spanning optimization, path planning, reinforcement learning, and importance sampling methods for finding failures and estimating their probabilities.

Presents POMDPStressTesting.jl, a Julia package that uses reinforcement learning and stochastic optimization to find likely failures in black-box safety-critical systems via adaptive stress testing, with pluggable solvers built on the POMDPs.jl interface.

Examines certification considerations for applying adaptive stress testing to airborne software, discussing how the technique can find likely failures in safety-critical aviation systems within the context of software certification requirements.

Develops a method to convert neural network verification results into probabilistic safety guarantees for neural network controllers, combining adaptive policy overapproximation with MDP model checking and state abstraction for aircraft collision avoidance.

Presents techniques for efficient stress testing of computationally expensive black-box safety-critical systems, including surrogate-model cross-entropy variants, episodic adaptive stress testing, and an adaptive framework that learns and exploits system weaknesses.

Presents a method for generating representative small UAS trajectories from open-source geospatial and map data, evaluating its scalability and diversity to support Monte Carlo collision-risk encounter modeling for airspace integration.

Extends adaptive stress testing to episodic sequential decision-making and applies Monte Carlo tree search to uncover likely, high-probability failures in a commercial flight management system's trajectory predictor, outperforming Monte Carlo and cross-entropy baselines.

Presents ACAS Xu, an integrated collision avoidance and detect-and-avoid system for UAS that uses Markov decision process logic and rollout-based risk anticipation, demonstrating substantial safety benefits and alert reductions in fast-time Monte Carlo simulation.

Proposes learning an urban air mobility encounter model from expert preferences rather than data, using preference-based learning to tune an MDP reward from pairwise expert queries and generate realistic encounter trajectories with minimal expert time.

Develops a parametric Bayesian network model of pilot response to TCAS resolution advisories that varies with individual encounter properties, showing that conventional nonparametric models underestimate collision risk in safety-critical encounters.