Welcome to the Intelligent Systems Research Laboratory at Texas A&M University. We focus on developing advanced algorithms and analytical methods to design next-generation autonomous systems operating in uncertain dynamic environments. We achieve these goals using theoretical tools from stochastic dynamical systems, robust control, nonlinear estimation, and convex optimization. Our recent work is summarized below.
Certification of Autonomous Flight with Theory & Data Driven Machine Learning
We expect Urban Air Mobility to rely more on autonomy than any other transportation sector, and consequently, it is critical to quantify their safe flight envelope accurately. Computational tools for critical analyses of the system, such as determining region-of-attraction, establishing local performance bounds, and identifying the largest uncertainty set for which robust performance is guaranteed, are necessary to establish trust in this increased autonomy. This research focuses on developing new formulations and algorithms for determining the region of attraction of uncertain nonlinear system dynamics with various gust models, establishing bounds on exogenous signal attenuations and quantifying the largest uncertainty set for which the control and estimation algorithms perform satisfactorily. These formulations incorporate simulation data (black box) and physics-derived vehicle dynamics, resulting in constrained PDEs in high dimensional space, which are solved using recent developments in scientific machine learning.
We are interested in determining sparse architectures for control and estimation for large-scale dynamical systems in this work. For large-scale systems it is nontrivial to determine location and precision of sensors and actuators to achieve the desired closed-loop system performance. We also consider the tradeoff between utility and privacy in these problems. These problems are considered in the Kalman filtering framework including ensemble and unscented variants, and \(\mathcal{H}_2\) and \(\mathcal{H}_\infty\) framework. Applications include tracking of aerospace vehicles, structural health-monitoring of wind-turbine blades, and battery thermal management.
Space domain awareness is concerned with tracking space objects and classifying them with respect to specific characteristics. In this research, we are developing novel algorithms for uncertainty propagation and state estimation. Challenges include non-Gaussian uncertainty supported on cylindrical coordinate systems \(\mathbb{R}^5\times\mathbb{S}\), sparse sensing, and unknown sensor characteristics. These algorithms are used for conjunction analysis, which predicts upcoming object encounters to notify satellite operators and avoid high-risk encounters.
N. Das, R. Bhattacharya, Sparse Sensing Architecture for Kalman Filtering with Guaranteed Error Bound, 1st IAA Conference on Space Situational Awareness (ICSSA), Orlando, FL, USA, 2017.
Future exascale machines are expected to have \(10^5-10^6\) processors, providing a deep hierarchy of systems and resources. However, many challenges exist, which must be overcome before exascale systems can be utilized as a useful tool to further understanding critical scientific inquires. Among the main obstacles to scaling code to exascale levels is the communication necessary in tightly coupled problems, such as in uncertainty propagation, turbulence flow simulations at high Reynolds numbers, and large-scale convex optimization. The synchronization across processors can cause 50-80% processor idle time. Our research focuses on applying systems theory to asynchronous numerical algorithms that do not wait for data to be synchronized. Communication between processors is modeled as a stochastic channel, and the behavior of the numerical algorithm is analyzed in a stochastic jump dynamical system framework. In addition, we present a new systems theory for finite-difference schemes that guarantee a given spectral error, a given approximation order, and numerical stability for a given linear PDE. The formulation treats the coefficients in a finite-difference formulation as control gains and the finite-difference scheme is synthesized as a controller for a linear dynamical system.
Cyber-physical systems have strong coupling between physics, communication, and computation. In our research, we develop algorithms for quantifying uncertainty in system behavior due to uncertainties in the physics (unmodeled dynamics, process and sensor noise), communication (irregular channels, packet loss, etc.), computation (jitter in real-time tasks, CPU transients, etc.). The system-level behavior is modeled as a stochastic jump system, and new uncertainty propagation algorithms for such jump systems are developed. New stochastic scheduling algorithms have been developed that switch between computational tasks to ensure system-level robustness.
Our lab has expertise in designing custom aerial platforms for various needs. Our research integrates aerodynamics, structural design, and flight control design in a single unified framework. The objective is to develop next-generation tools for rapid custom design of high confidence unmanned air vehicles for various industries, including defense, oil & gas, and precision agriculture. The vision is to codesign much of the system engineering aspect by integrating state-of-the-art in computational fluid dynamics, structural mechanics, robust control theory, CAD software, and 3D printing. The application focus is currently on aerospace systems but can be extended to general autonomous systems.
A. Halder, K. Lee, and R. Bhattacharya, Optimal Transport Approach for Probabilistic Robustness Analysis of F-16 Controllers, AIAA Journal of Guidance, Control, and Dynamics, 2015.
R. Bhattacharya, S. Mijanovic, E. Scholte, A. Ferrari, M. Huzmezan, M. Lelic, M. Atalla, Rigorous Design of Real-Time Embedded Control Systems, IEEE Advanced Process Control Applications for Industry, Vancouver, May, 2006.
Hypersonic flight leading to entry descent landing of a large spacecraft on the surface of Mars has been identified as a research area by NASA. The requirement is to land within a few kilometers of the robotic test sites. One of the significant concerns of high mass entry is the mismatch between entry conditions and deceleration capabilities provided by supersonic parachute technologies. In such applications, there are uncertainties present in initial conditions and other system parameters. Estimating these systems’ parameters is a challenging problem because of the nonlinearities in the system and the lack of frequent measurements. The evolution of uncertainty (as shown in the figure) is non-Gaussian. In our work, we develop new algorithms for UQ, state-estimation, and guidance algorithms. The controlled descent ensures robustness with respect to system uncertainties and guarantees landing at the desired site with high accuracy.