Irfan Ahmad Ganie, a doctoral candidate in electrical engineering, will defend their dissertation titled “Human-Robot Teaming Using Safe Lifelong Learning Based Optimal Adaptive Control Framework.” Their advisor, Dr. Jagannathan Sarangapani, is a curators distinguished professor in the electrical and computer engineering department. The dissertation abstract is provided below.

This dissertation presents a unified, safe lifelong optimal learning control framework for human–robot teaming that addresses the challenges of uncertain nonlinear systems across multiple domains. We first develop a lifelong integral reinforcement learning (LIRL)–based optimal trajectory tracking scheme for continuous-time affine systems subject to state constraints. In this methodology, a critic multilayer neural network (MNN) approximates the value function while an auxiliary neural network identifier generates optimal control policies. The critic’s weights are updated online using a novel singular value decomposition (SVD)-based method, extendable to MNNs with multiple hidden layers, and an integrated lifelong learning scheme mitigates catastrophic forgetting. The effectiveness of this approach is demonstrated on a two-link robotic manipulator, achieving a 47% cost reduction compared to existing methods.

Building on this foundation, we propose an integral reinforcement learning (IRL)–based optimal tracking controller for strict feedback systems, applicable to mobile robots, UAVs, and similar platforms. By combining backstepping with dynamic surface control, the controller relaxes the need for repeated derivative computations of virtual controllers. Novel online SVD-based weight update laws for both actor and critic deep neural networks (DNNs) are derived to minimize a discounted value function, while an online lifelong learning technique addresses the vanishing gradient and catastrophic forgetting problems. Simulation results in mobile robot tracking reveal a 76% total cost reduction relative to prior art.

Recognizing that in many practical scenarios system states are not fully measurable, we extend the framework to an output feedback control scheme. This approach employs a scalable multilayer neural network observer alongside an actor-critic MNN via an IRL/adaptive dynamic programming strategy. The observer and controller weights are updated using SVD-based methods combined with continual learning, while output constraints are enforced via barrier Lyapunov functions and Karush–Kuhn–Tucker (KKT) conditions. Validation on a two-link robotic manipulator shows an 80% performance improvement over existing techniques.

Since robots increasingly operate alongside humans, we further incorporate human–robot teaming through an online safety-aware Stackelberg game–theoretic control framework for human–robot–object manipulation. In this setting, human intention is estimated in real time using neural networks whose weights adapt via continual learning. The resulting control architecture employs an actor-critic MNN framework with SVD-based weight updates, integrating the human’s objectives into the robot’s control policy while maintaining safety via dual-component control enforced by KKT conditions.

Finally, we address cooperative manipulation in multi-robot teams with unknown dynamics by developing a distributed deep neural optimal adaptive observer–based control framework. A multilayer neural network–based distributed observer estimates and transforms human-intent–based reference trajectories from limited local information, eliminating the need for precise grasp matrices or globally shared states. At the lower level, a cooperative game theoretic controller ensures Pareto-optimal coordination, with safety rigorously maintained through composite control barrier functions and enhanced KKT conditions. Simulation results confirm the approach’s real-time efficacy and a 42% cost reduction compared to recent methods.