Unmanned Autonomous Vehicles (UAV) are expected to play a central role in many applications of great interest, including urban and infrastructure monitoring, precision agriculture and delivery services. However, making the operations of these airborne platforms autonomous requires the execution of algorithms for the real-time analysis of the surrounding environment and mission planning. The complexity of these algorithms clashes with the inherent constraints of UAVs, whose embedded systems have limited computing power and energy supply. The proposed research aims at the development of techniques to make distributed computing ultra reliable in the context of UAV systems. The project establishes a layer of intelligence that controls in real-time how information is propagated and processed across the layers of the systems to transform sensorial input into decisions, as well as a semantic form of neural compression to significantly reduce the amount of data transported over weak wireless links. This project will have a broad impact in terms of education, mentorship and outreach. The impact of this research effort will be broadened by providing unique outreach and educational opportunities to students at the collegiate and high school level.
The proposed project approaches the problem from two complementary angles: (1) designing deep reinforcement learning (RL) agents that learn to optimally communicate data across noisy channels, and (2), building novel lossy compression algorithms based on probabilistic deep learning that are specifically designed for distributed machine learning without a human in the loop. For (1), one of the key challenges resides in the multi-scale nature of the stochastic processes driving the system dynamics that present important geographical and temporal trends at different scales. An innovative hierarchical learning approach will be used to make the RL agent effective as the system dynamics evolve across time and space. For (2), novel kinds of extreme neural lossy compression algorithms based on Variational AutoEncoders (VAE) with additional supervision will be designed. Instead of focusing on signal reconstruction, the resulting compressor will aim at preserving only relevant information needed for a specified supervised learning task, leading to the new paradigm of supervised (or semantic) compression.
This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.