The Speakers

As the demands of real-time, energy-constrained AI outpace what conventional von Neumann architectures can deliver, neuromorphic hardware has emerged as a compelling alternative that promises efficient, event-driven computation. This talk traces a vertically integrated path toward neuromorphic hardware across four layers of abstraction. I begin at the materials layer, where AI, including generative models, helps identify candidate materials for neuromorphic devices. I then turn to device-level simulations of memristive elements that link material parameters to measurable device response. Moving up the stack, I examine circuit architectures that organize these devices into computational primitives for in-memory and event-driven processing. Finally, I connect this hardware foundation to system-level demands, using Advanced Air Mobility as a motivating application in which real-time sensing, computing, and control must coexist under severe size, weight, and power constraints. By following a single thread across these layers, the talk surfaces the co-design questions that arise only when materials, devices, circuits, and applications are considered together.

Short Bio

Dr. Mike Banad is an Associate Professor in the School of Electrical and Computer Engineering and an affiliated faculty member in the Materials Science and Engineering Program at the University of Oklahoma, where he leads the INQUIRE Laboratory. His research develops intelligent and efficient neuromorphic computing systems by connecting materials, devices, circuits, algorithms, and applications, with an emphasis on AI-driven design that spans the full hardware stack. His group's work ranges from generative and predictive models for materials discovery, to memristive device modeling, to in-memory and event-driven circuit architectures, and on to neuromorphic systems for autonomous platforms such as multi-UAV and advanced air mobility applications. His works are supported by federal research programs spanning AFOSR, AFRL, ONR, NASA, and NSF.

Sarah Sharif (University of Oklahoma)

Presentation Title

Thinking at the Speed of Light: Neuromorphic Photonic Hardware Beyond the von Neumann Wall

Abstract

The rapid evolution of modern science and engineering is increasingly shaped by four converging pillars: semiconductor devices, computing systems, heterogeneous integration, and AI-enabled design. Despite remarkable progress, fundamental limits in conventional computing architecture, and the von Neumann bottleneck in particular, continue to constrain efficiency, scalability, and raw computational speed. This talk will present photonics as an alternative paradigm for both computation and sensing, using reconfigurable nanophotonic devices to move past the bandwidth and energy ceilings of electronics. A second theme will be design itself: traditional physics-based methods are slow to scale, while purely data-driven AI often lacks interpretability and reliable training data. The talk will describe physics-informed inverse design, together with emerging hybrid quantum and machine-learning approaches, as a way to close that gap. It will close with a vision for integrated neuromorphic photonic hardware that delivers high bandwidth, low latency, and energy-efficient intelligence for sensing, computing, and control.

Shot Bio

Dr. Sarah Sharif is an Assistant Professor in the School of Electrical and Computer Engineering at the University of Oklahoma, with affiliated appointments in the Center for Quantum Research and Technology (CQRT) and the Materials Science and Engineering Program. She earned her Ph.D. in Electrical and Computer Engineering, with a minor in Physics, along with multiple master’s degrees spanning electrical engineering, physics, and materials science.

Her research focuses on nanophotonics, neuromorphic photonic hardware, and quantum-enabled optical systems, with an emphasis on programmable and intelligent photonic platforms for sensing, computation, and information processing. She has over a decade of experience across both academia and industry and has contributed to the development of advanced optical and quantum optical devices through numerous peer-reviewed publications and federally funded research projects supported by agencies such as the National Science Foundation (NSF), Air Force, and office of Navy.

Dr. Sharif is an IEEE Senior Member and currently serves as Chair of IEEE Young Professionals (Region 5). She is also an Associate Editor for Applied Optics and actively contributes to the community as a reviewer and committee member and session chair for major conferences and journals, including IEEE IPC, IEEE RAPID, Optica, and SPIE Photonics West.

Yu She (Purdue Univeristy)

Presentation Title

Toward Intelligent Manipulation: From Tactile Sensing and Embodied AI to General-Purpose Robotics

Abstract

Despite remarkable progress in robot perception and learning, today's robots still struggle with the kinds of contact-rich physical tasks that humans perform effortlessly. A central reason is that prevailing sensing and computing architectures remain heavily vision-centric, while the sense of touch, which is fundamental to dexterous manipulation in biology, is largely absent or underutilized. This talk presents the Purdue MARS Lab's research program aimed at closing this gap by co-designing tactile sensing hardware, multimodal perception, and learning-based control for general-purpose manipulation. I will introduce a family of high-resolution tactile sensors that provide rich geometric, force, vibration, and proximity information across scales from surgical tools to full robotic hands. Building on this sensory substrate, I will discuss our recent efforts on visuotactile policy learning, including the ManiFeel benchmark for visuotactile imitation learning, TacVLA for contact-aware vision-language-action models, tactile-aware quadrupedal loco-manipulation. I will conclude with a perspective on how tightly coupling tactile sensing, embodied AI, and hardware design can move robotics toward truly general-purpose physical intelligence.

Short Bio

Yu She is an Assistant Professor in the Edwardson School of Industrial Engineering at Purdue University, with courtesy appointments in the Elmore Family School of Electrical and Computer Engineering and the School of Mechanical Engineering. He directs the Purdue Mechanisms and Robotic Systems (MARS) Lab, where his research focuses on tactile sensing, mechanism design, intelligent control, and robot learning for manipulation in manufacturing, healthcare, agriculture, and underwater and space robotics. He received his Ph.D. in Mechanical Engineering from Ohio State University in 2018 and was a Postdoctoral Associate at MIT CSAIL from 2018 to 2021 before joining Purdue. His work has been recognized with the ASME Freudenstein Young Investigator Award (2025), the Showalter Early Investigator Award (2024), a Google Research Scholar Award (2022), the ASME DSCC Best Paper Award (2018), and Best Paper Finalist recognitions at RSS (2020). He is a Senior Member of IEEE. His research is supported by NSF, USDA, DOD, NASA, and industry partners including Google, Eli Lilly, NVIDIA, and Meta.

Pavel Salev (University of Denver)

Presentation Title

Neuromorphic functionalities enabled by electrical dynamics of a metal-to-insulator phase transition

Abstract

Insulator-to-metal transition (IMT) materials, such as VO2 and NbO2, exhibit volatile high-to-low resistance switching that produces an abrupt increase in current flow, resembling neuronal excitation. This phenomenon of electrical excitation has stimulated extensive research aimed at emulating biological neuronal spiking using IMT materials. Materials exhibiting the opposite metal-to-insulator transition (MIT), such as (La,Sr)MnO3, have been largely overlooked in the development of neuromorphic electronics. In this talk, I will discuss the resistive switching properties of MIT materials and demonstrate how dynamic MIT triggering can be used to generate and amplify neuron-like electrical oscillations. Electrical triggering of the MIT produces volatile low-to-high resistance switching that suppresses current flow, resembling neuronal inhibition. This electrical switching is accompanied by the development of a pronounced N-type negative differential resistance (NDR) in the I-V characteristics. Incorporating such an N-type NDR electrical switch into a simple RL circuit produces inhibitory self-oscillation dynamics under the application of a constant voltage, where each oscillation produces a brief current disruption in the circuit. Biasing the MIT switch into the NDR regime further allows for the amplification of small time-varying signals, such as spiking trains generated by artificial neuristors, effectively enabling the implementation of axon-like functionalities. Overall, MIT materials provide an interesting new platform for developing neuromorphic electronics. Combining MIT and IMT materials can greatly expand the design space for realistically mimicking excitatory and inhibitory behaviors of biological neuronal systems.

This work was supported by the U.S. Department of Energy, Office of Science, Basic Energy Sciences, under Award # DE-SC0026129.

Short Bio

Pavel Salev is an assistant professor in the Department of Physics & Astronomy at the University of Denver. His lab specializes in the development of correlated oxide electronics for applications in physics-based computing. He received bachelor’s and master’s degrees from Tver State University (Russia) and a Ph.D. in physics from the University of Tulsa. Prior to joining the University of Denver, he was a postdoctoral scholar at the University of California San Diego.

Xiaoqiu An (University of Michigan)

Presentation Title

Lithographically Generated 2D Bi₂Se₃ Grid Patterns as Physical Reservoir Computing Network Devices

Abstract

Physical reservoir computing offers a promising pathway toward energy- and resource-efficient temporal information processing by using the intrinsic nonlinear dynamics and short-term memory of materials and devices. In this talk, I will present our recent work on lithographically generated 2D Bi₂Se₃ grid patterns as physical reservoir computing network devices. These programmable layered-semiconductor networks are fabricated using rubbing-induced site-selective deposition, enabling controllable cross-linked structures with different grid densities.

We investigate how physical network structure regulates reservoir-mapping capability by systematically comparing Bi₂Se₃ networks ranging from grid-free channels to highly interconnected grid structures. Step-forward dynamic modeling confirms reservoir-like state evolution, where the output response depends on both the current input and previous device states. Mackey–Glass time-series prediction further shows that increasing network density and output-terminal number improves temporal prediction performance, suggesting that denser cross-linked networks generate richer and higher-dimensional reservoir states.

To connect this structure-regulated temporal-mapping capability to control applications, we further demonstrate Bi₂Se₃ physical reservoirs in robotic control tasks, including rover target tracking and LiDAR-based wall following. In these demonstrations, sensory signals are processed through the hardware reservoir, and a trained readout layer generates motor-control outputs. Overall, this work shows that programmable Bi₂Se₃ reservoir networks can serve as compute-in-physics platforms for temporal prediction, sensing, and control-oriented applications.

Jiao Suo (Purdue University)

Presentation Title

Organic Mixed Ionic-Electronic Conductor-based Optoelectronic Device for Artificial Retina

Abstract

The human retina integrates light sensing and preprocessing within a compact biological structure, offering inspiration for more efficient vision architectures than conventional systems that physically separate sensing and processing. To emulate retinal functions, artificial image sensors must reproduce one or more key retinal operations, such as temporal adaptation, short-term memory, spatial-temporal preprocessing, etc. These functions remain challenging to achieve using conventional silicon-based photodiodes without dedicated circuit design, since such materials rely on pure electronic conduction and typically exhibit fast optoelectronic responses with limited intrinsic temporal dynamics. Organic mixed ionic-electronic conductors (OMIECs) provide a promising materials platform for bio-inspired vision systems. The motion and redistribution of ionic species in OMIECs can generate rich temporal photoresponses dynamics, while their bandgap, molecular structures, morphology that could affect the optoelectronic properties are highly tunable. In addition, their compatibility with both ambient and aqueous environments could benefit various applications in robotic, wearable, and bio-interfacing visual systems. In this presentation, I will present our research progress in developing OMIECs-based artificial retinal devices by exploiting these intrinsic materials properties to realize retina-like functions. Specifically, we demonstrate device working mechanisms that introduce memory effects enabling temporal integration computing during sensing. We also achieve devices that have bipolar, i.e., positive and negative photoresponse, which can support spatial integration computing during sensing through appropriate pixel arrangement. We are further exploring strategies to combine temporal and spatial integration functions within a single device platform. Through the above work, we aim to establish an integrated organic retina model based on OMIECs for more efficient artificial visual system.

Short Bio

Jiao Suo is a Postdoctoral Researcher in James Tarpo Jr. and Margaret Tarpo Department of Chemistry at Purdue University, where her research focuses on retinomorphic device design, retinomorphic optoelectronic sensing mechanisms, and integrated system for more biomimetic and energy-efficient artificial vision. Before joining Purdue University, she received her Ph.D. in Mechanical Engineering from City University of Hong Kong and Master in Materials Science from Wuhan University.

Nina Mahmoudian (Purdue University)

Presentation Title

Toward Embodied Intelligence for Persistent Autonomous Systems

Abstract

Persistent autonomous systems must operate for extended periods in dynamic, uncertain, and resource-constrained environments while maintaining safety, adaptability, and mission effectiveness. Achieving this capability requires embodied intelligence, where perception, reasoning, control, and physical interaction are tightly integrated with a system’s sensing, energy resources, and environmental context. While recent advances in AI and foundation models have expanded autonomous capabilities, long-duration operation in the physical world continues to demand adaptation to uncertainty, limited communication, and changing conditions.

This talk presents recent advances in persistent maritime and cross-domain autonomy, highlighting autonomous underwater, surface, and aerial systems operating collaboratively in challenging environments. Examples include autonomous docking and recovery, adaptive autonomy, multi-vehicle coordination, and perception-driven decision making. These examples demonstrate how intelligent behavior emerges from the integration of sensing, computation, control, and environmental interaction, enabling long-duration operation with minimal human intervention. The talk concludes with future directions toward scalable embodied intelligence for persistent autonomous robotic teams.

Short Bio

Nina Mahmoudian is Professor of Mechanical Engineering and B.F.S. Schafer Scholar at Purdue University. She is a recipient of the NSF CAREER Award and the ONR Young Investigator Award and was a Fulbright Scholar in Portugal. In 2026, she is a Digital Futures Scholar-in-Residence at KTH in Sweden. Supported by NSF, DARPA, ONR, and industry partners, her research focuses on persistent autonomy for maritime and cross-domain robotic systems. She develops AI-enabled methods for adaptive control, mission planning, and collaborative autonomy that enhance the endurance, efficiency, and scalability of autonomous underwater, surface, and aerial vehicles operating in dynamic and uncertain environments.

John Hyatt (Army Research Office)

Presentation Title

Practical End-to-end Sensor-model-task Systems: Representing and Aligning Multimodal Data

Abstract

As “AI” adoption continues to accelerate across industry and government, more and more data is being fed into models that are expensive and unreliable – sometimes catastrophically so. Scientific advances continue to lag far behind engineering and product development. The result is a commercial landscape dominated by black-box machine learning algorithms, built on universal function approximation theorems, and optimized for benchmark performance. Such models parameterize complex statistical correlations in their training data (for example, specific configurations of pixels in an image rather than the determinative features of the system being observed), leading to high-dimensional and arbitrary internal representations. From a multimodal fusion standpoint, this means there is no easy way to align different representations, even though the underlying physical world is the same. From a hardware standpoint, large data representations require more compute, more storage, and more power, all of which are heavily restricted at the edge. From a communications standpoint – especially in contested environments, or when bandwidth is limited or reception is spotty – high-dimensional data representations are not only inefficient, they’re fragile: small perturbations in the transmitted representation can have a large impact on the outputs of downstream models. All of these concerns are highly relevant to the Army, but equally so in many practical civilian applications. This presentation is motivated by the idea of a distributed, heterogeneous system of sensors collecting multimodal data in an open-ended, resource-constrained environment and the end-to-end pipeline linking that system to effective performance on a downstream task. The goal is to identify relevant gaps in current state of the art and discuss potential ways to close them.

Short bio

John got his PhD in Physics from Georgia Tech in 2016 and joined the Army Research Laboratory as a member of their first AI/ML team in 2017. In 2022 he was invited to manage the Information Processing and Fusion program at the Army Research office, which funds fundamental research performed at academic institutions. His program focuses on developing solid mathematical foundations (especially for AI/ML) in such areas as: identifying and exploiting intrinsic data structure; active and collaborative sensing; effective alignment and representation of multimodal data; uncertainty quantification and propagation; and transparent, reliable, efficient, and principled model development. Outside of the program, his major professional goal is to narrow the understanding gap between widespread industry messaging (and the resulting unrealistic expectations common among non-experts in the Army and elsewhere) and the scientific and technical reality of current models, especially LLMs.

Ronald Ogden (University of Texas at Austin)

Presentation Title

Optics for Computation-free Event-based Vision Compression

Abstract

The use of remote vision sensors for autonomous decision-making poses the challenge of transmitting high-volume visual data over resource-constrained channels in real-time. In robotics and control applications, many systems can quickly destabilize, which can exacerbate the issue by necessitating higher sampling frequencies. We propose a novel sensing paradigm in which an event camera observes the optically generated cosine transform of a visual scene, enabling high-speed, computation-free video compression. In this study, we simulate this optically passive vision compression (OPVC) scheme and compare its rate-distortion performance to that of a standalone event camera (SAEC), finding that the rate-distortion performance of the OPVC scheme surpasses that of the SAEC.

Short bio

Ronnie Ogden is a PhD candidate in the Department of Aerospace Engineering at the University of Texas at Austin. He is currently researching how to leverage optics for real-time vision compression in robotics systems. His recent work utilizes the fields of stochastic control/estimation, information theory and physics. Prior to this, Ronnie was a flight test engineer at Wisk Aero, a company developing autonomous aircraft for urban air mobility. He has a passion for improving autonomous flight systems and exploring new problems. Ronnie received his BS in Aerospace Engineering and Mathematics from the Massachusetts Institute of Technology. In his free time, he enjoys climbing, language learning and solving puzzles.

Fei Xia (University of California, Irvine)

Presentation Title

Optical Machine Learning with Programmable Complex Optics

Abstract

Complex optics refers to light propagation in disordered, highly scattering structures which naturally mixes vast optical modes, with potential for high-dimensional, passive, and massively parallel transforms at optical bandwidths. This makes programmable complex optics an interesting platform to explore optical machine learning. Here, we report two advances along this direction trying to address two key challenges in optical machine learning: nonlinearity and scaling. To introduce nonlinearity in optical computing, we leverage a multiple-scattering cavity that provides structural nonlinearity, acting as a parallel optical encoder that maps inputs into rich nonlinear optical feature spaces without electronic overhead. To introduce larger scale computing, we introduce a reconfigurable, trainable optical neural networks use complex media as dense optical weights; with in-situ (hardware-in-the-loop) training, we directly optimize the system’s transfer function, reducing calibration and scaling naturally with aperture/mode count. Their computing performances are evaluated on industry and scientific datasets and show potential unconventional paths towards optical machine learning.

Short Bio

Fei is an Assistant Professor of Electrical Engineering and Computer Science (EECS) at the University of California, Irvine. She completed her Ph.D. at Cornell and her postdoc at École Normale Supérieure (ENS) in Paris. She leads a research team working at the interface of optics, computation, and biomedicine, with a particular interest in designing smart optical systems to image, sense, and process biomedical information probed by light. She aims to provide innovative solutions to challenging biomedical needs through the co-design of hardware and software. Fei has been recognized by several major awards and honors, including the Early Career Research Excellence Award from University of California, Nvidia Academic Research Grant, Scialog Fellow, Rising Star in Light, the Optica Foundation Challenge Award, the SPIE Women in Optics.

François Léonard (Sandia National Laboratories)

Presentation Title

Artificial Intelligence with Minimal Electronics

Abstract

Existing implementations of AI rely heavily on electronic components for computation and sensing. This reliance on digital electronics comes at steep energetic, temporal, and economic costs. In this presentation I will discuss alternative approaches to realize AI functionality where the physics of physical fields and their interactions with materials are harnessed as much as possible before involving any electronics. I will use optical systems as concrete examples to illustrate how materials can be designed/trained to perform AI functionality with high performance and how these systems have parallels and key differences with electronic neural networks. Finally, some key challenges and research opportunities for physical AI will be discussed.

Short Bio

Dr. François Léonard is Distinguished Member of Technical Staff in the Materials Physics Department at Sandia National Laboratories in Livermore, CA, where his research focuses on experimental and theoretical Nanoelectronics and Nanophotonics. He holds B.Sc., M.Sc. and Ph.D. degrees in Physics, and was a postdoctoral fellow at IBM Research before joining Sandia. He is the author of the book “The Physics of Carbon Nanotube Devices” and is a Fellow of the American Physical Society.

Mohammad Nadji (Neuraville Inc.)

Presentation Title

Bridging Classical Control and Neuromorphic Intelligence: FEAGI’s Path from Software to Silicon

Abstract

Modern autonomous systems cannot rely on neural intelligence alone, nor can they depend entirely on classical or model-based control. PID controllers, model predictive control, and other engineered control methods remain reliable, interpretable, and commercially proven for stability, constraints, and predictable behavior. Adaptive neural architectures, meanwhile, offer learning, perception, memory, and resilience in environments where fixed models and fixed control laws are insufficient. The challenge is not choosing between these paradigms, but creating an architecture where they can coexist.

This talk presents FEAGI as a cross-platform neuromorphic software architecture designed to unify deterministic control, adaptive neural computation, and embodied sensorimotor intelligence within a single deployable framework. FEAGI enables conventional control patterns, including PID-style feedback loops and MPC-oriented supervisory control, to operate alongside intelligent neural structures inside a cortical architecture, allowing systems to combine stability, responsiveness, learning, and adaptation without fragmenting the software stack.

The commercial value of FEAGI lies in its deployment path: from simulation and robotics to embedded hardware, edge systems, and ultimately ASIC-oriented neuromorphic implementations. By treating software architecture as the bridge between research hardware and industrial autonomy, FEAGI provides a practical path for translating neuromorphic materials, devices, sensing systems, and compute-in-physics research into real products.

The talk will discuss how this hybrid control-neural approach can reduce commercialization risk, preserve compatibility with proven engineering practice, and create a migration path from today’s software-defined embodied intelligence to tomorrow’s specialized neuromorphic hardware.

Short Bio

Dr. Mohammad Nadji is a software architect, AI researcher, and the Founder and CEO of Neuraville Inc.. He has more than two decades of experience designing and leading enterprise-scale software platforms across multiple industries, with a focus on complex distributed systems, software architecture, and applied artificial intelligence. His interdisciplinary work draws from AI, neuroscience, robotics, and systems engineering. He is the inventor of FEAGI, the Framework for Evolutionary Artificial General Intelligence, an open-source brain-inspired AI platform. Through Neuraville, he works to make advanced machine intelligence and robotics more accessible, trustworthy, and useful in real-world settings.

David Nolte (Purdue University)

Presentation Title

Fourier Phase Holography

Abstract

Interferometric holographic recording in the Fourier domain has distinct advantages over either image-domain or Fresnel holography. For instance, applications of real-time holographic films benefit from the physical optical Fourier transform of simple lenses, while digital holography with numerical reconstruction benefits from fast Fourier transform algorithms. This talk will introduce phase-sensitive holographic applications for laser-based ultrasound detection as well as timing-jitter compensation on femtosecond pulses using real-time holographic films composed of nonlinear optical semiconductor materials. Fourier-domain holography keeps the advantages of phase-sensitive detection with additional functionality associated with dynamic compensation of phase aberrations as well as spectral interferometry of ultrafast laser pulses. In addition, phase-sensitive Fourier holographic applications in digital holography include dynamic speckle detection with coherence-gating for laser ranging. A recent advance in our lab has pushed the limits on holographic stability for detection of ultra-low-frequency Doppler shifts down to 100 micro-Hertz corresponding to a detected velocity of 33 picometers per second or equivalent to displacing a third of a hydrogen atom per second.

Short Bio

David D. Nolte is the Edward M Purcell Distinguished Professor of Physics and Astronomy at Purdue University performing research in the fields of optical interferometry and holography. He received his baccalaureate from Cornell University, his PhD from the University of California at Berkeley, and was a post-doctoral member of AT&T Bell Labs before joining the physics faculty at Purdue. He is a Fellow of the National Academy of Inventors, the Optica society, the American Physical Society and the AAAS. In 2005 he received the Herbert Newby McCoy Award for scientific advances and in 2026 he received the Charles B. Murphy Award for undergraduate teaching at Purdue University. He has 25 issued patents and founded two biotech startup companies in diagnostic screening and high-content analysis. His most recent book, Interference: The History of Optical Interferometry (Oxford University Press, 2023), is a popular account of the scientists and engineers who developed optical interferometry and holography

Mel White (Naval Research Laboratory)

Presentation Title

Event-Based Wavefront Sensing for Adaptive Optics

Abstract

A major speed limitation of wavefront sensors such as Shack-Hartmann (SHWF) is the time it takes to read a full frame of data from the sensor. Event-based sensors (EBS) provide asynchronous streaming information at fine time resolution, posing a potential solution for real-time wavefront detection and adaptive optics. However, the relationship between the wavefront and sensor response is not well-characterized, and the few existing methods generally rely on data processing that is either too costly for real-time applications, or fails to fully exploit the information contained in neuromorphic data. This talk will discuss potential ways to integrate EBS into SHWF detectors using physics-informed priors and computational cost-aware algorithms, towards the goal of reducing latency in real-time systems.

Short Bio

Dr. Mel White is a postdoctoral research fellow at the Naval Research Lab in Optical Sciences in Washington, DC. Mel specializes in developing novel hardware for end-to-end computational imaging systems, fusing CMOS sensor design with optics, machine learning, and classical signal processing to image under challenging environmental conditions and resource constraints. She completed her PhD at Cornell University, where her work on custom imaging hardware was supported by an NSF Graduate Research Fellowship, and subsequently joined Rice University as a postdoc and member of the Rice Academy of Fellows. Prior to her research career, she served as a Peace Corps volunteer, and holds a B.A. in International Relations from William & Mary and a B.S. in Electrical Engineering from Rensselaer Polytechnic Institute.

Nak-seung Patrick Hyun (Korea University)

Presentation Title

Towards the Autonomy for Bio-Inspired Robotics with Extreme Behaviors

Abstract

Highly agile and extreme behaviors of many biological systems offer examples for future research directions to target similar mobility in bio-inspired robots, thereby understanding the complex dynamics and subsequently designing a robust and adaptive control framework. Examples of extreme behaviors in biological systems include the fast oscillation-driven maneuvers of bees flapping their wings at around 200 Hz and the rapid, impulsive strikes of mantis shrimp, which release their stored potential energy within milliseconds. The challenges in controlling robots exhibiting similar extreme behaviors stem from highly nonlinear dynamics operating over multiple timescales. Specifically, one has to account for fast dynamics (extreme motions) and slow dynamics (time-averaged motion or slower drift in the system), as well as the time-varying actuation model in the high-frequency regime (fast dynamics) vs the low-frequency regime (slow dynamics). This talk will address the control-theoretic aspects of addressing such challenges in bio-inspired robots, grounded in first principles of mathematical system theory. The first part of this talk will address recent progress in controlling a flapping-wing vehicle and the challenges in state estimation. In addition, recent findings on the nonlinear modeling of the dynamic principles underlying mantis shrimp impulsive strikes will be covered, enabling striking speeds of up to 27 m/s within a few milliseconds.

Short Bio

Nak-seung Patrick Hyun is an assistant professor in Electrical Engineering at Korea University and an adjunct faculty member at Purdue University. He was previously an assistant professor in the Electrical and Computer Engineering department at Purdue University from January 2023 to August 2025. He was formerly a research associate at the Harvard Microrobotics Laboratory, hosted by Robert J. Wood. He received a Ph.D. in Electrical and Computer Engineering in 2018, an M.S. degree in Mathematics in 2013, and an M.S. degree in Electrical Engineering in 2013 from the Georgia Institute of Technology. His research focuses on the control-theoretic aspects of bio-inspired robots, emphasizing systems with extreme behaviors, such as flapping vehicles and impulsive systems. His research program provides a cyclic learning cycle between biology, mathematical system theory, and robotics. His recent work has been published in multidisciplinary science domain journals, including Science Robotics, Nature, and PNAS, as well as in control theory journals such as Automatica and Nonlinear Analysis: Hybrid Systems. He is the recipient of the Granite Tower Teaching Award at the Korea University in 2026, and the Ruth and Joel Spira Outstanding Teacher Award at the Purdue ECE department in 2024.

Hritom Das (Oklahoma State University)

Presentation Title

Bio-Inspired Neuromorphic Circuits for Continuous Learning and Energy-Efficient Edge AI

Abstract

The rapid growth of edge AI is driving the need for intelligent systems that can continuously learn and adapt while operating under strict power and computational constraints. This talk presents a bio-inspired neuromorphic computing framework that draws inspiration from the complementary roles of the hippocampus and cerebral cortex in biological learning and memory. By combining programmable plastic neurons with heterogeneous memory technologies, including SRAM, eDRAM, and ReRAM, the proposed architecture supports fast adaptation, long-term knowledge retention, and energy-efficient continuous learning directly on-chip. The talk will introduce recent advances in mixed-signal neuron circuits, spike-timing-dependent and homeostatic plasticity, heterogeneous neuroprocessors, and event-driven architectures that reduce data movement and computational overhead at the edge. Finally, the presentation will discuss emerging opportunities and challenges in building scalable neuromorphic hardware for next-generation autonomous systems, robotics, healthcare, and intelligent sensing applications.

Short Bio

Dr. Hritom Das received the Ph.D. degree in Electrical and Computer Engineering from North Dakota State University, Fargo, ND, USA, in 2020. He was a postdoctoral research associate at the University of Tennessee, Knoxville, where he was affiliated with TENNLab. He is currently an Assistant Professor in the Department of Electrical and Computer Engineering at Oklahoma State University, Stillwater, OK, USA. His research interests include neuromorphic computing, low-power VLSI circuit design, memory devices, AR/VR systems, data privacy, and optimization techniques.

Kaushik Roy (Purdue University)

Presentation Title

Flight of the Micro-Brain: Unleashing AI Efficiency at the Edge

Abstract

Neuromorphic computing, a paradigm inspired by the biological brain, offers a blend of efficiency and performance that could be transformative for artificially intelligent systems operating in resource-constrained environments, such as edge computing. To harness this potential, we propose a system design approach that emphasizes the importance of appropriate input sensing, suitable learning algorithms, and specialized neuromorphic hardware to achieve "brain-like" efficiency. We illustrate this approach using vision-based drone navigation (VDN) as an exemplary application, drawing parallels with micro biological systems like the fruit fly. For such cognitive systems to function effectively, they must possess a detailed understanding of environmental structure and dynamics, involving complex perception tasks such as optical flow, depth estimation, and object detection/tracking.

Given the sequential and complex nature of these tasks, alongside the resource constraints typical of edge computing, we argue that efficient data acquisition methods—such as event-based cameras—and event-driven asynchronous computing models, like Spiking Neural Networks (SNNs) and hybrids of traditional analog neural networks (ANNs) and spiking networks, hold significant promise. Moreover, the true benefits of these brain-inspired algorithms cannot be fully realized on conventional deep learning hardware; instead, they require the development of specialized neuromorphic hardware. This hardware differs from traditional von Neumann architectures by employing near-memory and in-memory computing techniques to process core computational elements and support event-driven operations. Finally, we explore the domain-specific challenges related to large-scale integration and offer insights into algorithm-hardware co-design for emerging applications.

Imraan Faruque (Oklahoma State University)

Presentation Title

Physically-grounded AI via incorporating physical structure into neural networks and state estimators

Abstract

The inherent misalignment between generalist neural networks and the underlying structured found in 6 DOF physics, along with its associated uncertainty structure, has led to a growth in neural network size and attendant training data needed to provide physically grounded artificial intelligence. Incorporating physical structure into neural network design can significantly reduce the network scale and training data. This talk builds on parallel advances in state estimation through improved motion models for physical systems, illustrating how moving beyond generalist state estimation techniques like extended Kalman filters allows improved state estimate quality. Specific examples include 3DOF and 6DOF aerospace examples that incorporate the group structure of rigid body motions in SO(3) to provide improved state estimate quality.

Timothy Brown (Oklahoma State University)

Presentation Title

Designing dynamic compute-in-physics hardware for next-gen autonomous systems

Abstract

Dynamic vision sensor (DVS) based cameras leverage asynchronous in-sensor computation for low latency, highly temporally resolved machine vision, enabling autonomous systems to more accurately count quickly-moving parts, damp vibration, or give real-time driver assistance. These systems also highlight a crucial lesson: post von-Neumann architectures must be leveraged not only in software, but also in hardware to reach their full potential for low-energy edge computation. Until recently DVS cameras have been limited by needing to transport data to external NPU based chips to compute using in-software spiking neural networks. However, recent work shows that spiking neural networks implemented in hardware by embedding the computation in memristor physics to compute can lead to orders of magnitude improvement in compute per power expenditure. In this talk, we aim to demystify memristors, reviewing their history and some unfortunate terminology pitfalls, and demonstrating how cleverly chosen materials with strongly temperature-dependent transport and fading memory emulate the ion-gated switching physics that enable spatiotemporal information processing in biological neurons. In fact, since biological neurons and resistance switching materials both fall under the memristor device framework, we demonstrate that simple two terminal structures composed from various oxides and semiconductors are all universally capable of neuronal dynamics including tonic spiking, leaky integrate and fire, and axon-like amplification. We conclude by describing various materials-based approaches to develop fully CMOS-compatible compute-in-physics spiking neural network hardware compatible with DVS autonomous systems.

Short Bio

Dr. Timothy Brown graduated with a B.S. in Engineering Physics specialized in Analog Electronics in 2014 Engineering from the University of Tulsa. He obtained his PhD in Materials Science and Engineering from Texas A&M University in 2019. After graduation he transitioned to a postdoc in Electrical Engineering at TAMU and then to a second postdoc at Sandia National Labs. Dr. Brown is now an assistant professor at Oklahoma State University and leads the Brain Inspired Circuits for Computation and Communication (BraIn-C3) lab group engineering neuronal and nonlinear oscillator behaviors in analog devices. BraIn-C3’s research synthesizes Dr. Brown’s interdisciplinary background to characterize, model, and benchmark brain-inspired devices. Dr. Brown was a recipient of the NSF GRFP award in 2016 and was awarded the Excellence in Outreach award for his leadership roles during grad school. He loves cats and reading pop science and engineering histories, and greatly enjoys working with undergraduate and graduate students.

Page updated

Google Sites

Report abuse