Congratulations to Aerospace Assistant Professor Alex Gorodetsky, who has been awarded a three-year contract of $450,000 from the Air Force Office of Scientific Research (AFOSR) through its Young Investigator Program (YIP). His proposal entitled, “Online Reinforcement Learning in Partially Observed Physics-based Environments with Compression-based Computation” discusses the need for progress in computational mathematics and algorithms for reinforcement learning within autonomous aerospace systems. Professor Gorodetsky explains:
“Reinforcement learning, or the process of learning to act by interacting with the environment, is traditionally an extremely computationally- and data-intensive procedure. If you look at the latest results – for instance, in the area of AI for computer games – significant training time (sometimes on the order of days) with large scale computational resources and extremely large amounts of data is still required. In aerospace applications, these requirements are prohibitive.”
Thus, the goal of Professor Gorodetsky’s research is to develop new computational mathematics to enable autonomous systems such as dynamic soaring aircraft or drone delivery systems to “rapidly adjust to unforeseen conditions in order to extend their applicability to these important next-generation applications.” The algorithms will be deployed on these two applications of autonomous systems navigating complex aerial environments. These applications are challenging for current approaches because data is often limited, the environment is rapidly changing, and the autonomous platform may have limited onboard computational power. Professor Gorodetsky provides the example of a high-altitude autonomous glider, which could have atmospheric science research applications like the Perlan Project aircraft:
“High altitude autonomous gliders have the capacity to enable long-duration missions for monitoring and/or tracking. The way they can do this is by identifying thermals and up-drafts in the atmospheric flow, and then leveraging these processes to enable optimal performance. Traditional CFD-based approaches to model these processes with high-fidelity simulations are too expensive to embed on systems that often have limited computational and communication resources. As a result, these systems must be given the ability to learn these physical processes and then to act accordingly. Furthermore, learning and acting must occur in real-time. We will verify these approaches through simulation, and potentially in limited experimentation.”
Ultimately, Professor Gorodetsky hopes to use his YIP grant to “develop a plug-and-play computational framework to enable reinforcement learning for autonomous systems” in order to “reduce the requirement of humans to participate in low-level decision making regarding how to execute a mission.” With the objective of reducing human error, “efficient computational mathematics approaches and algorithms can serve as a foundation of the ‘brain of the autonomous system,’ and enable humans to concentrate on providing the high-level mission specifications. ” Autonomous systems are widely-regarded as a safer alternative to manned missions in warzones or in unpredictable environments. By improving the robustness of unmanned aerial systems, Professor Gorodetsky’s research can help mitigate the need for dangerous manned missions, a key step in propelling the aerospace industry forward while protecting its pilots.