Past Research:

Faster HJ-based Algorithms - UC Berkeley  [03/16 - 03/17]

A collaboration with Glen Chou and Mo Chen.

Hamilton-Jacobi reachability is a formulation that allows us to generate controls with certain guarantees, such as liveness and safety, for systems, independent of their linearity. However, the dynamic program underlying computing reachable sets has a running time complexity that grows exponentially with the number of states of a system. Several methods are used to approximate these reachable sets in order to avoid this slow computation, but these methods cannot always deliver our desired guarantees. We developed a technique that efficiently computes reachable sets and maintains the same guarantees that vanilla HJ-based algorithms provide.

UAV Traffic Management (UTM) - UC Berkeley  [02/16 - 03/17]

A collaboration with Mo Chen and Aparna Dhinakaran.

Many companies and groups are proposing projects like Amazon Prime Air and Google[x]'s Project Wing to use UAVs to provide civilian services. To prepare for these ideas, we need to implement UAV traffic management (UTM) systems to ensure guarantees for these drones, such as safety and timeliness. The algorithms used to compute control that guarantee safety or timeliness are computationally intractable for large multi-agent systems, so several structural assumptions have been proposed to make these computations tractable, one of which is platooning. With my collaborators, I worked on a hardware implementation of platooning as a proof of concept.

Human-Robot Shared Autonomy in Saftey-Critical Scenarios - UC Berkeley  [06/16 - Present]

A collaboration with Dexter Scobee.

Shared autonomy between human and robots is becoming more and more prevalent as new robotic systems intended for direct human use (self-driving cars, surgical arms, UAVs, etc.) continue to emerge. Little of the work done to address the problem of shared autonomy consider safety-critical scenarios where either the robot or the human may be more knowledgable than the other about the situation, enabling them to act more safely. We worked on and proposed methods for handling the human-robot interaction in these safety-critical situations.

Learning Quadrotor Dynamics with Neural Networks - UC Berkeley  [02/16 - 08/16]

A collaboration with Somil Bansal, Kene Akamatelu and Forrest Laine.

Models for quadrotor dynamics are well-established and can be used for generating good control. However, in general, controllers for quadrotors are derived from a linearization of the dynamics, which means that there are phenomenon that are purposely not accounted for in order to achieve a linear model for deriving control. This project aims to utilize neural nets to learn a higher fidelity model that includes previously unmodeled phenomenon in hopes to generate better control. [video]

TRACER - UCSF  [03/15 - 08/15]

Co-advised by Professor Noah Zaitlen and Professor Michael McManus, I developed algorithms for reconstructing cell lineage phylogenies. The primary purposes of the project were to prove that such phylogenies are actually computationally reconstructable and explore possible improvements to the underlying lineage tracing biosystem used to enable the reconstruction of these cell lineages.

Though my research focus is no longer on bioinformatics, feel free to discuss it with me, as I still find the field interesting.