• Astrodynamics
• Attitude Dynamics and Control
• Space Situational Awareness (SSA)
• Machine Learning (ML) and Artificial Intelligence (AI)
  • Supervised Learning
  • Reinforcement Learning

Some detailed topics are included below.

The cislunar region (the space between Earth and the Moon and around them) is the focus of deep space exploration in the next decade. To support a long-duration mission, spacecraft need to have more autonomy than what they currently have. Recall that how crucial the GPS system is to our everyday life now. In the near future, a GPS-like positioning system in the cislunar region will be indispensable for all sorts of cislunar activities. Even though there is no immediate need from current missions, researchers have explored the usage of CR3BP for such a navigation system and found many difficulties due to the unstable nature of CR3BP. Our study using the rich dynamic characteristics in ER3BP for such a positioning constellation design is novel. Our research in this direction will bring some fresh ideas and options to this area.

• Hao Peng, and Xiaoli Bai, “A New Representation for Nominal Orbits of A Cislunar Navigation Satellite Constellation (under review)”, The Journal of the Astronautical Sciences, 2023.
• Hao Peng, and Xiaoli Bai, “Segmented Fitting Representations of Cislunar Navigation Satellite Constellation”, 2023 AAS/AIAA Astrodynamics Specialist Conference, Big Sky, Montana: 2023.

To our understanding, one great drawback of the reinforcement learning reinforcement is that they cannot be trusted for practical applications when the transition gap between the simulation environment and the reality cannot be ignored. However, this doesn't render RL useless if we can find a proper problem to solve.

In normal cases, the onboard system of a satellite will switch the system to safe mode when any anomaly occurs so that the ground control center can rescue the satellite promptly. The safe mode will ensure the solar panel and antenna orientations with highest priority to secure the power supply and the communication. However, if the malfunction occurs to the attitude control subsystem and beyond the expectation of safe mode, the satellite easily becomes a forever loss. The missing automation for the satellite here is to develop control laws by its own when facing unknown catastrophes.

This project will make use of reinforcement learning (RL) techniques and empower the satellite with surviving ability from catastrophic loss. The two major concerns here are still the power supply from solar panels and the communication from antenna, which both rely on a correct attitude and pointing. One appealing feature of this project is that the eventual output will be a general method that can fit any missions. The only requirement will be a computational unit for real-time training which may be already available on modern satellites.

Mega-constellation is a big trend in satellite communications. Starlink now has more than 3,000 satellites on orbit and more are being launched in a rapid pace. Meanwhile, space situational awareness (SSA) area is highly interested in using space-based optical or radar sensors to track all the resident space objects (RSOs). Due to the close proximity and less atmosphere interference, space-based sensors are expected to improve the overall SSA ability.

However, no mega-constellation is built or designed for the SSA purpose. Our research will fill the gap between matured mega-constellations and SSA, including the related orbital dynamical problems, sensor allocations, mission software and hardware requirements, expected outcomes, etc.

In most situations, people talk about mitigating and removing space debris, but our study focus on how to increase the accululation of space debris to eventually triger the famous Kessler Syncdrome, where cascading collisions occur. To accomplish this goal, we set up a simplified but real-enough environment based on the up-to-date space catelog. The dimension and mass of the debirs will also be modeled. Collaborative AI agents will be guided in this virtual reality environment to learn a policy to triger the catastrophe. Available actions include, for example, blowing up big objects (satellites or rocket bodies), adding new objects, etc.

The project goal is of course still to prevent the syndrome. We can use this AI to test against other mitigation resorts. Also, this AI will enlight us what operations should be definitely prevented and what objects should be removed before they break up.

Maneuver detection is a difficult problem for uncooperative resident space objects (RSOs). Common method is to detect the maneuver based on propagating of orbit estimates using a high-fidelity numerical model to a common intermediate epoch, and then analyze the statistics of all the deviations. This process is resource-consuming. My previous work has already established a ML approach to improve orbit prediction accuracy. ML models are trained using historical data to directly improve orbit predictions without modeling additional orbital dynamics or the satellite. For this study, I plan to create a method based on the performance of the trained ML models that are used to improve orbit prediction accuracy.