Explore GEN family of websites, social media experiences, blogs and more.

Unleashing ideas.
Strengthening startups.

Back to search results
    Date submitted
  • 23-Jan-2017



Applying innovative machine learning concepts to spacecraft trajectory design.

Conventional spacecraft trajectory optimization techniques are rather computationally expensive and must be done ad hoc. New revolutions in machine learning have shown great promise in the subject area of continuous optimal control of dynamical systems. Applying both machine learning and other computational intelligence techniques in tandem enable the spacecraft trajectory design process to be expedited, saving both time and money.


Additional Questions

Who is your customer?

This software library is marketed at customers necessitating intelligent and expedient solutions to space mission design. Among these customers may be: the National Aeronautics and Space Administration, the European Space Agency, the Japanese Aerospace Exploration Agency, and various other private astrodynamics companies. Of course pricing will differ on a case by case basis. Factors drawing into pricing are the costs of: computing resources, transportation, and time-involved.

What problem does this idea/product solve or what market need does it serve?

Mitigates computational expense and improves solution optimality in spacecraft trajectory optimization for space mission design.

What attributes will make this idea/product successful? Why do you believe that those features will create success?

A housefly is a rather simple organism, yet it is able to independently make decisions to achieve its goals, such as navigating to a food-source and avoiding obstacles. Inspecting closer, a housefly is able to make these decisions instantaneously, such as in the case of being swatted at by a human. If one thinks about the descent of a lander onto the Martian surface, the nature of the situation is quite the same. Because communication with Earth is prolonged, the lander must make decisions on its own in order to safely land on the surface. If a common housefly can independently make decisions in real-time, in uncertain dynamic environments, than surely a spacecraft should be able to do the same in an environment where the objective is clearly outlined. Astro.IQ will mitigate the computational expense of conventional spacecraft trajectory optimization and enable real-time autonomous control. As of now, the interplay between machine learning and spacecraft trajectory optimization has been largely unexplored. Recently, there have been a few developments suggesting that the use of machine learning, particularly deep learning, in trajectory optimization may prove to be advantageous in many aerospace applications. Machine learning has proven to be quite successful in applications such as facial recognition, language translation, and game playing. Recently, machine learning has been quite successful in many applications of control, most significantly in beating the world champion in the game of GO. It is quite obvious that machine learning should be quite successful in a variety of control scenarios. Astro.IQ will implement these methods in a variety of astrodynamics scenarios.

Explain how you (your team) will execute to make this idea/product successful? What gives you (your team) an advantage over others already in the market or new to this market?

Within the software library, several scenarios of spacecraft trajectory optimization will be examined. Additionally, several different paradigms of machine learning will be investigated and integrated; among them: recurrent neural networks, reinforcement learning, deep learning, and others. Three distinct scenarios of spacecraft trajectory optimization will be examined, namely: cislunar trajectories (e.g. from a parking orbit about Earth to the $L_2$ Lagrange point in the Earth-Moon system), planetary landings (e.g. upright landings of reusable rockets (like SpaceX)), and interplanetary trajectories (e.g. from Earth to Mars). It should be noted that each of these scenarios are distinctly different, as their dynamics have varying levels of complexity, and thus it is my hope that these models will become benchmarks for those in the astrodynamics community who want to validate their own machine learning algorithms. Each of the different scenarios and machine learning paradigms will manifest as periodically scheduled deliverables and milestones of this project. As both machine learning and spacecraft trajectory optimization are rather numerically intensive, the great majority of the software library's back-end will be built on a framework of C++ and Fortran. On the other hand, in order to maximize the intuitiveness of the user-experience, the front-end will be built on a framework of Python. This will allow the software to both function expeditiously and be intuitive enough to allow fast prototyping.