NSF Award
1735004
This project will integrate the capabilities of deep learning networks into a biologically inspired architecture for sensorimotor control that can be used to design more robust platforms for complex engineered systems. Studies of the flexibility and contextual aspects of sensorimotor planning and control will extend existing paradigms for human-robot interactions and serve as the foundation for creating personal assistants that are able to operate in natural settings. Growing understanding of how these layered architectures are organized in the brain to produce highly robust, flexible, and efficient behavior will have many applications to rapidly evolving technologies in complex environments, including the Internet of Things, autonomous transportation, and sustainable energy networks.<br/><br/>The nervous system is a layered architecture, seamlessly integrating high-level planning with fast lower level sensing, reflex, and action in a way that this project aims to both understand more deeply and mimic in advanced technology. The central goal is to develop a theoretical framework for layered architectures that takes into account both system level functional requirements and hardware constraints. There are both striking commonalities and significant differences between biology and technology in using layered architectures for active feedback control. The most salient and universal hardware constraints are tradeoffs between speed, accuracy, and costs (to build, operate, and maintain), and successful architectures cleverly combine diverse components to create systems that are both fast and accurate, despite being built from parts that are not. Recent progress has made it possible to integrate realistic features and constraints for sensorimotor coordination in a coherent and rigorous way using worst-case L-infinity bounded uncertainty models from robust control, but much remains to explore to realize its potential in neuroscience and neuro-inspired engineering. Another application of this framework will be to software defined networking, which explicitly separates data forwarding and data control, and provides an interface through which network applications (such as traffic engineering, congestion control and caching) can programmatically control the network. This makes it a potential "killer app" for a theory of integrated planning/reflex layering; research collaborators will be eager to deploy new protocols on testbeds and at scale.
