William Curran

Education

Oregon State University, Corvallis, OR

Dual PhD, Robotics and Computer Science, June 2017

• Advisors: Dr. Bill Smart and Dr. Prasad Tadepalli
• Focus of Study: Reinforcement Learning, Learning Based Controls, and Human-Robot Interaction
• Dissertation: High-Dimensional Reinforcement Learning with Human Feedback.

M.S., School of Electrical Engineering and Computer Science, June 2013

• Advisor: Dr. Kagan Tumer
• Focus of Study: Multiagent Systems

B.S., School of Electrical Engineering and Computer Science, June 2011

• Emphasis on Machine Learning and Artificial Intelligence
• Cum Laude

CV

projects

Learning-Based Controls for MERLIN (2017-2018)

The goal of the Meso-scale Robotic Locomotion Initiative (MERLIN) is to develop a small robot that a marine could carry in a backpack and deploy to conduct intelligence, surveillance, and reconnaissance missions. Due to the small size requirements, MERLIN uses hydraulics that have much more energy density, but are a harder engineering and modeling challenge. I developed Control Theoretic/Deep Reinforcement Learning hybrid techniques with the goal of robust, adaptive, and model-free control.

• Merlin Overview

Dimensionality Reduced Reinforcement Learning (2015-2017)

The complexity of state-of-the-art personal robots lead to large dimensional state spaces, which are difficult to learn in. To alleviate this issue, we developed a technique called Dimensionality Reduced Reinforcement Learning. This technique leverages concepts from dimensionality reduction, transfer learning and reinforcement learning to learn high-dimensional policies quickly.

• Open loop control of the Yaskawa Motoman MH5F robot arm balancing a ball in three separate starting scenarios.
• Closed loop control of the Yaskawa Motoman MH5F robot arm balancing a ball.

Project Chiron (2014-2016)

Project Chiron is a self-driving wheelchair project at Oregon State University. We worked with individuals with ALS and the ALS foundation to develop a small package that can be mounted on a powered wheelchair to provide self-driving capabilities. In 2016, I was a part of a 4-man team sent to the Robots for Good competition in Dubai. We placed 7th out of over 700 entries.

• In-house video demonstration of full system with Google Glass.

ROSPalantir (2013-2016)

The ROS ecosystem is an interconnected web of packages, nodes and people with no efficient means to compare, assess or visualize them. We've developed a set of tools consisting of various metrics, a data visualization web app, and an active monitoring system. With these tools, we aimed to elucidate the current state of the ecosystem as well as determine where community efforts should be directed.

• Invited talk at ROSCon
• Github

As Autonomous As Possible (2014-2015)

In this work we introduce a risk-aware task-level reinforcement learning algorithm that adapts an end-user's risk tolerance. A3P learns a task-level policy where states are tasks and actions are approaches in accomplishing that task.

Investment Casting Optimization for Precision Castparts Corporation (2014-2017)

We applied Evolutionary techniques to optimize the workflow for the Small Structurals Business Operation for Precision Castparts Corporation Investment Casting throughput.

Multiagent Modeling of the United States National Air Space (2012-2013)

Worked at NASA and Oregon State University to apply multiagent reinforcement learning techniques to reduce congestion and delay in the National Airspace System (NAS). Developed a novel automated partitioning system that allowed agents to learn quickly and efficiently by separating the state space. This work eventually scaled to over 30,000 learning agents, simulating the entire NAS. This work resulted in my MS in Computer Science.

Machine Learning for the Autonomous Detection of Bird Roosts (2010-2012)

Sparrows are extremely clustered in a small area while nesting overnight. In the morning they fly away from their nest in such large numbers that they show up as strange clouds on weather radar. I used computer vision techniques to analyze this weather radar data and autonomously classify where these bird roosts are located. Ornithology researchers could then use these locations to study bird habits.

Learning How to use Adjectives to Classify Images (2011)

In this work I explored how adjectives could beneficial to computer vision algorithms. Adjectives are rich descriptors, but are very subjective and hard to acquire. I led a user study determining if there are a common set of adjectives people use for determining whether a car is cool, cute, or classic. We found that there are a large number of overall adjectives that people use to determine a style of car, but there are a select few that all participants used. We then Incorporated these adjectives in a computer vision algorithm.

Service

• 2019 Reviewer for the Journal of Autonomous Robots
• 2018 Panelist in the Machine Learning for Fully Autonomous Maritime Systems Workshop at Unmanned Maritime Systems Summit
• 2018 Panelist in the Verification and Validation Workshop at ICRA
• 2018 Reviewer for the Transactions on Human Robotics Interaction Journal
• 2018 Reviewer for HRI Pioneers Workshop
• 2017 Treasurer and Founding Member of the Oregon State Robotics Graduate Student Association
• 2017 Program Committee Member for IJCAI Special Track on AI and Autonomy
• 2016 Program Committee Member for IJCAI Main Track
• 2016 Reviewer for AAAI Multiagent Interaction without Prior Coordination Workshop
• 2013-2018 Reviewer for AAMAS Adaptive Learning Agents Workshop

Skills