Publications

Conferences

[IAC’20] Skylar Eiskowitz, Sydney Dolan, Kir Latyshev, George Lordos, Matthew Moraguez, Alejandro Trujillo, Bruce Cameron, Oliver de Weck, Edward Crawley. “Quantifying the Impact of Cryo-Management, ISRU, and Fuel Cell Lunar Technology Infusion to a Notional Mars LOX/LH2 Architecture.” International Astronautical Congress, 2020. (here)

[IAC’20] Kir Latyshev, Sydney Dolan, Skylar Eiskowitz, George Lordos, Matthew Moraguez, Alejandro Trujilo, Bruce G. Cameron, Oliver de Weck, Edward F. Crawley. “Impact of the Lunar Gateway Location on the Human Landing System in case of Permanent Base at the Lunar South Pole.” International Astronautical Congress, 2020

[ASCEND’20] Sydney Dolan, Skylar Eiskowitz, Edward F Crawley, Bruce G Cameron. “Comparative Benchmarking of Crewed Lunar and Mars Mission Architectures.” ASCEND, 2020. (here)

Journals

[AA’21] Inigo del Portillo, Sydney Dolan, Bruce G. Cameron, Edward F. Crawley. “Architectural Decisions for Communications Satellite Constellations to Maintain Profitability While Serving Uncovered and Underserved Communities” Acta Astronautica, 2021. (pending approval)

Thesis

Control and Convolutional Neural Net Based Pose Estimation for On-Orbit Assembly. (link)

Talks

NASA XTM Meeting – June 2022

Pink Space Program – Skype A Scientist 2021

Orange County Public Schools – Skype a Scientist 2021

PhD Work That Wont Make the Final Cut But I Spent A Lot of Time On Anyways

During the first year of my PhD, I explored the potential of using control barrier functions for space traffic management. Control Barrier Functions are an appealing technique for risk averse domains like aerospace because CBFs can be used to provide certifications of a controller’s safety by relying on Nagumo’s property of forward invariance (i.e. if it starts in a safe set, it won’t leave it). However, CBFs are incredibly hard to derive for even simple dynamics. Recently, computer scientists have utilized machine learning to derive CBFs. With these neural network derived CBFs, we are able to determine control actions that always guarantee a safe action.

I’ve written two papers on this topic. The first was a literature review that surveys the existing state of the art for space traffic management collision avoidance maneuvers, and the state of CBFs. The second paper covers the development of a control Lyapunov barrier function that I designed for multi agent collision avoidance case. In the paper, I compared the impact of several different collision avoidance screening metrics on their resultant recommended collision avoidance maneuvers.

There are several reasons that this topic did not work out as a final PhD topic. There is a critical need for control barrier functions to be applied to more robust and realistic problems, as existing work hyperfixates on simple dynamic models with the occasional tuning variable to account for perturbation. However, for the field of space traffic management, there are several serious barriers (no pun intended) to applying CBFs on space flight hardware. First, the training of CBFs is extremely computationally intensive. For a single conjunction maneuver, the training took several hours. While this trained CBF could be extensible to other conjunctions in a similar area, preliminary research shows that it is unlikely that it can generalize to other conjunctions. Existing optimal control techniques that can optimized to be a maximum distance away from a satellite given a delta-V are much more preferable, generalizable, and computationally efficient. Second, while control barrier functions on ground can promise a set safety rate after dozens of tests, there is a lack of tractability and understanding around the neural network itself. While the neural network is trained to minimize violation of control barrier conditions, there is always the chance that it could output an unsafe action with just the right conjunction. In a field where a risk of collision of 10^-4 is considered unacceptable, space traffic management experts are reluctant to rely on the the nebulous inner workings of neural networks. Similarly, work in control barrier functions neglects to consider the training-testing data gap. Until this is addressed (and this could potentially be an fundamentally unsolvable problem), CBFs will never be implemented in the space traffic domain.

As a result, I did not continue on in the direction of control barrier functions for my PhD. CBFs are a neat trick to encode a more dynamic constraint on safety for ground robotics. However, they are ultimately too finick and poorly understood when compared to existing techniques in the field.