Research

I study how robots move and how robots think about movement.

The Robotic Space Exploration (RoSE) lab advances methods for mobile autonomous robots in diverse space environments, from satellites in orbit to underwater vehicles in extraplanetary oceans to planetary rovers on lunar or Martian soft soil. Our algorithms enable robots to navigate and conduct scientific exploration in extreme, scientifically-rich settings. My work leverages dynamics and control theory integrated with machine learning models and planetary science domain knowledge to equip robots with the intelligence to interpret unpredictable and unknown terrain conditions in real time, ensuring safe and efficient traversal. Ultimately, we aim to create robotic explorers capable of replacing or complementing human missions, digitizing the roles of both scientists and explorers in high-risk environments.

If you’re curious about what my day-to-day activities look like, I keep a research log.

Data-Driven Identification of Soft-Soil Rover Dynamics

This project aims to develop a comprehensive dataset to enhance understanding of vehicle-terrain interactions, particularly in off-road environments. The project seeks to improve predictive models for vehicle mobility and performance by collecting and analyzing data on various terrains and vehicle responses. This dataset is intended to support research in terramechanics, assist in designing more efficient off-road vehicles, and contribute to advancements in autonomous vehicle navigation systems.

Exploring Planetary Surfaces: A Machine Learning Approach Using Gaussian Processes for Resource Mapping with Autonomous Rovers

This project investigates the performance of a rover equipped with an ASD FieldSpec 4 Spectrometer tasked to map the spectral composition of an analog planetary environment on Mauna Kea, Hawai’i. The rover's onboard computer runs a Gaussian Process active learning algorithm to comprehensively map the environment while reducing reliance on human intervention — a critical advantage for operations in space that limit communications and emphasize ambitious mission timelines. Key performance metrics, including model accuracy, convergence, and samples collected, are evaluated to assess the algorithm's effectiveness in intelligent exploration. Mechanically, the rover provides mobility to a location, calibrates the spectrometer with automated white referencing, and measures payload readings with an active light source shielded by an awning. This research demonstrates the rover's adaptability and reliability in analog extraterrestrial environments, moving us closer to a rover with increased autonomy suitable for space operations.

Developing Rover Intelligence Faster with Cyber-Physical Twins

Cyber-physical twins are particularly important and useful when physical systems are hard to build and data is expensive to collect. These two qualities exemplify space systems and missions.

NASA EPSCoR research conducted by the Robotic Space Exploration (RoSE) lab, led by Dr. Frances Zhu, is developing a simulated planetary surface environment and rover with representative physics to accelerate developing and testing machine learning algorithms. Machine learning hinges upon the quantity and quality of data available. Simulations allow developers to toggle these aspects of data to detect the sensitivity of algorithms to the availability and noisiness of data. 

A simulated environment, once established, can be trusted to behave like a real, physical system, but unlike a physical system, a simulation enables rapid iteration to try algorithms that push the limits of a physical system without damaging expensive assets. By leveraging a simulated environment, software engineers are able to encode intelligence into a robot and play out potentially dangerous scenarios so that we may avoid and protect against failure in real mission deployment.

The RoSE lab’s research focus resides on imbuing intelligent behavior in robots exploring planetary surfaces that we have never seen before. Aspects of intelligence in autonomous exploration include localization (knowing where we are), path planning (knowing where to go), dynamic system identification (understanding how we move), and control (knowing how to get there). By partnering with NASA JPL and Astrobotic, the RoSE lab is developing open-source simulation and software packages to advance autonomous robotic space exploration. 

Vertically Integrated Project (VIP) Robotic Space Exploration (RoSE) Team

VIP Team RoSE (Robotic Space Exploration) is the first group of students at the University of Hawai’i at Manoa (UHM) to participate in the University Rover Challenge (URC), a competition sponsored by The Mars Society. We are a multidisciplinary team consisting of business, biology, astrogeology, astrophysics, computer science, and electrical, mechanical, and computer engineering undergraduate students.

Artemis CubeSat Kit: a Low-Cost, Spaceflight-Ready 1U CubeSat and Educational Materials in the Public Domain

Our mission is to create 1) a low-cost satellite kit that can be used as a space flight mission, suborbital payload, avionics on a rocket, or as a tabletop data collector and 2) an online course that teaches undergraduates with no prior aerospace engineering experience the fundamentals to designing and building a small satellite.

Microgravity Flight Experiment

On Earth's surface, we are stuck between a rock and a hard, fast natural phenomenon known as gravity. Spacecraft in orbit do not follow same physical constraints but operate in microgravity environment. To prove the performance of spacecraft technology, we should test in environments much like the final physical environment: microgravity. A cost-efficient way to accomplish weightlessness is to fly in an airplane that follows a parabolic trajectory, lovingly known as the vomit comet. The flux pinning research team from JPL and Cornell flew with ZeroG in March of 2017 and 2018.

Learned Trajectory Tracking

Exploring extreme terrain is important for:

• Search and rescue missions

• Planetary exploration

Exploring extreme terrain is difficult because:

• Hard to track a trajectory

• Hard to simulate terramechanics

• Analytical or numerical approaches

Learning a dynamics model (includes terramechanics) and control policy in an interpretable form is valuable and insightful. This work explores a new method for learning a terramechanics model and a control policy using a model-based genetic algorithm.

Flux Pinning Air Bearing Experiment (Cornell)

I led a team of undergraduates to:

• conceive of a spacecraft docking mechanism utilizing flux pinning dynamics

• design and build an experiment centered on the analogue spacecraft

• conduct over 100 experiments that characterize the docking performance

Flux Pinning Sandbox

Dynamics model for flux-pinned single magnet and single superconductor interaction based on Kordyuk's frozen image model and magnetic moment dipole model

Flux Pinning Air Bearing Experiment (JPL)

The proof-of-concept experiment at Cornell showed huge potential so we transitioned the analogue spacecraft to JPL with several modifications.

• superconductor system is more flight traceable

• flat floor is leveled to within one thousandth of a degree

• Vicon camera system measures position and attitude to submilllimeter precision

We conducted five different experiments in over 500 experiments.

Flux Pinning Simulation

Dynamics model for twelve magnet and trio of superconductor interaction based on Kordyuk's frozen image model and magnetic moment dipole model. Highly coupled and nonlinear dynamics.

Mapping Magnetic Field Embedded in Superconductor

The strength of the flux-pinned interface directly depends on the magnitude and shape of the magnetic field. In this experimental effort, I spatially map the magnetic field the frozen image and mobile image sources generate. Spoiler, reality is not as strong as theory predicts.