Throughout my education and career, I have had a wide variety of experiences teaching me that the physical design of a robotic system is inherently multidisciplinary and requires extensive collaboration. I would like to share with you my research accomplishments working on underwater vehicles, quadrotors, legged robots, and miniature robots.

Dissertation: Information Gathering for Autonomous Vehicles

Project dates: January 2013-August 2018

Using autonomous vehicles for information gathering and coverage planning tasks present many unresolved challenges. Many current path planning techniques do not extend well to information gathering due to the path dependence of the reward. For reliable control, robotics systems require using a high fidelity motion model to generate feasible, smooth trajectories. For realistic scenarios, various sources of uncertainty such as motion uncertainty must also be accounted for.

My dissertation work investigates three distinct information gathering tasks: coverage planning for teams of underwater vehicles, maximizing mutual information for multipass coverage plans for unmanned aerial vehicles, and physics aware persistent monitoring tasks for unmanned surface vessels.

Topic 1: Maximizing Mutual Information for Multipass Coverage Planning

Consider a disaster scenario where search and rescue workers must search difficult to
access buildings during an earthquake or flood. Often, finding survivors a few hours sooner results in a dramatic increase in saved lives, suggesting the use of drones for expedient rescue operations.

Motion planning for multi-target autonomous search requires planning over an area with an imperfect sensor and may require multiple passes to accurately identify the locations of survivors. The algorithm I developed is based on best first branch and bound and is benchmarked against state of the art algorithms adapted to the problem in natural Simplex environments, gathering the most information in the given search time. We demonstrate that maximizing the mutual information of multipass coverage plans increases the speed of the search.

Topic 2: Physics-aware Persistent Monitoring Tasks for Gathering Recurring Reward

Consider the use of an unmanned surface vehicle (USV) conducting harbor patrols to detect intruders. It is reasonable to assume that possible intruders will enter the harbor from certain locations such as harbor entrances and shipping channels. This suggests the use of an “information value map” that signifies how some regions are more dynamic or interesting and should be observed more often. However, when close to shore, windy conditions or swift currents could cause the vessel to deviate from it’s intended path and run aground, requiring the planner to account the physics of the environment during it’s patrol route.

To solve this problem, I formulated the patrol problem as a persistent monitoring task and modeled the environment as an MDP, leveraging dynamic programming techniques to generate robust persistent coverage plans. This technique dramatically reduces the risk of following the patrol route by selecting waypoints that are both safe to reach and satisfy the monitoring task.

Topic 3: Coordinating Teams of Underwater Vehicles Conducting Large Scale Geospatial Tasks

I am currently investigating the use of path planning techniques to improve the performance of unmanned underwater vehicle (UUV) teams for large-scale geospatial tasks such as coverage planning. Major challenges to low cost long duration missions include expensive underwater positioning systems and energy storage for propulsion. It is advantageous to exploit the currents of the ocean to increase endurance. Forecast uncertainty must be accounted for, requiring the generation of feedback plans.

Demonstration Technologies with Quadrotors

The research group I work in at NRL focuses on conducting basic level autonomy research with quadrotors and ground vehicles. One of the major challenges in a basic research environment is that there are many different project requirements with a short turnaround time. As systems become more complicated by having more components, many more points of failure are introduced so the complexity needs to be managed to ensure that hardware experiments and demos can be relied upon to work. For over a year and a half I developed a modular software framework using Robot Operating System (ROS) where the components are reusable for a wide variety of tasks:

• deployed robust state estimation using Vicon motion capture data (using ETH Zurich’s modular sensor fusion framework
• tracking control for agile maneuvers
• take off and landing on a moving ground vehicle using visual servoing
• high fidelity simulation of experiments using hector, ROS and Gazebo.

Quadrotor Technologies

TODO youtube video

All of these capabilities are easily deployed to multiple platforms using a git repository. Code functionality is regularly tested using test driven development practices to ensure reliability.

Gait Generation and Stabilization for Quadruped Robotics

Project dates: June 2013–May 2015

Legged robots offer advantages over wheeled robotic platforms in that they can negotiate rough terrain. However, challenges exist in both stabilizing legged robotics and generating gaits to ensure reliable locomotion. Central pattern generators, which are found in the neural circuitry of many different animals, generate stable gaits and incorporate sensory feedback for stabilization.

For this project, I extended state-of-the-art task-based central pattern generators that account for ground contacts, enabling omnidirectional locomotion (forward, reverse, coordinated turn, sidestep and march in place). I developed real time joint controllers and trajectory generators in C++ using Eigen, and Rigid Body Dynamics Library for a real time model-based controller, controlling the Allegro Dog, an electrically powered 20kg quadruped robot.

In April 2015, As part of the National Robotics Week Kickoff Event, I got to present my work on the Allegro Dog at the Smithsonian National Air and Space Museum and give a live demonstration. That was a lot of fun, as I have visited the Air and Space Museum many times.

(My talk starts around 11:00)

I just discovered at a trade show that NRL is still using the CPG code I generated three years after I left the project!

Mixed Signal Computing for Miniature Robotics

Project dates: August 2010 — August 2012

For my M.S. Thesis in Electrical Engineering at the University of Maryland, College Park, I investigated mixed signal computing techniques for miniature robotics. Mixed signal computing offers distinct power advantages over digital computing by operating circuits in the subthreshold regime, using picoamps of current to represent continuously valued signals instead of continually charging and discharging logic gates.

For this project, I designed an analog circuit that computed the odometry of a miniature robot. This in turn enabled running an extended Kalman Filter and randomized receding horizon control on a miniature robotic platform with limited computational resources.
In addition, I developed an inter-robot distance sensor implementing Time Difference of Arrival between radio frequency packets and audio pulses. I also identified and characterized the dominant sources of noise which lead to the measurement error. This sensor was used to generate rendez-vous behaviors between agents.