Autonomous vehicles and systems: modeling, estimation, and control

This line of work develops information-theoretic and control-theoretic methods for detecting and mitigating anomalous behavior in autonomous vehicles. Feedforward mutual-information detection schemes quantify how much an observed sensor or state trajectory deviates from the information expected under nominal dynamics, flagging faults and adversarial behavior without requiring an explicit fault model. Context-aware spatio-temporal anomaly detection extends this to networked cyber-physical systems, fusing measurements across space and time to distinguish genuine anomalies from benign disturbances as vehicles move through changing environments. Building on both, optimal control formulations fold anomaly minimization directly into the control law, driving the vehicle's behavior away from anomalous regions of its state space during operation rather than flagging deviations after the fact. This is the primary focus of the current research program.

Mutual-information anomaly detector flagging an induced sensor fault on a ground-vehicle testbed in real time.

This work develops Bayesian and information-based estimation methods for teams of autonomous vehicles and mobile sensors. For uncooperative radio-frequency targets, where the target does not broadcast a cooperative position estimate, this includes recursive Bayesian filters that fuse noisy signal-strength and bearing measurements across a moving sensor team to localize and track the emitter, coupled with ergodic exploration strategies that balance exploring uncertain regions against exploiting the current belief during search. For gas-leak source-term estimation, the mobile sensor forms and updates a probabilistic estimate of a leak's location and emission rate from noisy, intermittent concentration readings as it moves through the plume, using an infotaxis-style find-and-consume strategy to actively drive the search toward regions of highest information gain. This work sits at the intersection of estimation theory, optimal motion planning, and multi-vehicle coordination.

For uncooperative radio-frequency targets specifically, the target transmits an unknown waveform and offers no cooperative position report, so the sensor team must localize and track it from received-signal-strength measurements alone. This work casts that problem as information-theoretic Bayesian inference: each sensor's posterior belief over target position is updated recursively as measurements arrive, and each sensor's next motion is chosen to maximize expected reduction in that posterior's uncertainty rather than following a fixed search pattern. This concentrates the team's trajectories where they most improve the estimate and degrades gracefully as the number of cooperating sensors changes.

Information-theoretic multi-sensor search converging on an uncooperative radio-frequency target. See related publication.

The same infotaxis-style framework extends to source-term estimation for multiple, simultaneous chemical-gas leaks. Rather than assuming a single source, the estimator maintains a belief over the number, locations, and emission rates of an unknown set of plumes, updating that belief from sparse, intermittent concentration readings as the sensor moves through overlapping plume structure. The search alternates between finding new sources and consuming, or fully resolving, sources already located, so the vehicle does not abandon a partially localized leak to chase a stronger signal elsewhere.

Supplementary video from "Mobile-robotic sensor for estimation and localization of multiple chemical-gas leaks: A find-and-consume infotaxis approach," showing a mobile sensor searching for and localizing multiple gas leaks. ScienceDirect's host blocks the clip from playing embedded elsewhere. See related publication.

Running alongside the theoretical work above, this industry-facing project stream designs and validates exoskeleton evaluation systems built and tested to NIST standards, applying rigorous dynamic modeling and instrumentation to quantify how these physical systems perform under real operating conditions. Related applied work has included a delivery drone for emergency medicine, close-proximity flight control for quadrotors, and buried-sensor networks for soil-moisture monitoring in precision agriculture.

This foundational work designs haptic manipulators for mixed-reality systems, pairing mechanical and CAD design with Unity-based virtual integration, and conducts human-subject studies comparing interaction with real versus virtual objects, grounding the control work above in models of human perception of and response to physical feedback.