Embedded Files
Active Projects
Integration of control and deep learning with safety guarantee
I am developing a control algorithm that ensures probabilistic safety in the presence of uncertainties in deep learning. In particular, I have formulated it as a chance-constrained problem and developed a computationally efficient, sampling-based approach to solve the problem with high confidence. One of the main results is that I have quantified the number of samples required to ensure a specified level of safety with high confidence. I am currently validating the closed-loop systems of control and deep neural networks on the realistic vehicle simulator CARLA.
Computation of reachable sets
I am developing a computationally efficient method to compute robust controlled invariant sets using zonotopes.
Previous Projects
Safety verification and control at road intersections
I developed controllers that override vehicles at road intersections only when necessary to prevent collisions. To determine the timing of overrides in real-time (called safety verification), I developed computationally efficient, yet approximate, approaches based on abstractions of the concrete dynamical model and mixed-integer linear programming formulations. I also quantified approximation error bounds to measure the quality of the approximate solutions.
Reachability-based decision making for autonomous driving
I developed a decision-making algorithm for autonomous driving in city scenarios, such as lane changes and coordination at unsignalized intersections. I proved that the algorithm guarantees safety and liveness; that is, the autonomous vehicle eventually reaches a destination without collisions. I also developed its interface with a lower-level motion planner, and validated the whole algorithm via hardware experiments.
Control projects in Internet of Things (IoT)
I participated in several control projects that utilize connected environment.
We designed centralized control algorithms to improve fuel consumption, air pollutant emission, and throughput at highway merging and a network of intersections.
Also, we developed a sensor selection method that prioritizes the most informative sensors for state estimation.
Moreover, I participated in a warehouse automation project to improve the robustness of mobile robot control when their commands are transmitted through a nondeterministic wireless communication link. Based on predicted link quality, we developed a method that determines the optimal transmission number and sequence of the commands.
Google Sites
Report abuse