Research Areas

Core Research Topics

1) Rehabilitative Systems

We design and evaluate control frameworks and therapies for rehabilitation platforms (e.g., exoskeletons),  which involve human-in-the-loop dynamic models, with a goal of improving human health. A particular focus is on the rehabilitation of individuals with neurological conditions that cause movement disorders. An important consideration whenever robots interact with a human is that the robots are capable of causing harm (particularly when the human is in a fragile, vulnerable, or unpredictable state), which requires stability to be ensured during all stages of control design and implementation. We develop control frameworks that emphasize safety-aware operation and reliable interaction in the presence of uncertainty. Example application areas follow:

• FES-based Rehabilitation Control:
  To achieve rehabilitative outcomes, individuals with movement disorders must perform repetitive exercises with a sufficient intensity and duration. However, these individuals may be unable to properly control their body to sufficiently perform these exercises due to their condition. Functional electrical stimulation (FES), which applies an electrical stimulus to artificially contract a muscle, can be used to assist an individual in performing their rehabilitative exercises. In the CARE Lab, we design control and estimation techniques for FES-based systems, including hybrid exoskeletons that merge FES with exoskeletons, which are inherently complex, uncertain, nonlinear, and time-varying. Our developments include strategies to handle time delays, input saturation, and subject-to-subject variability while maintaining reliable trajectory tracking and assistance.

• Telerehabilitation:
  Rehabilitation for individuals with movement disorders is often a long process, requiring many visits to rehabilitative clinics. To increase access to rehabilitation, the CARE Lab is working to develop control architectures that support more affordable at-home, robot-assisted rehabilitation. Specifically, we aim to lower costs by moving the computational expense from each individual rehabilitative robot to the medical facility. Our telerobotic control designs ensure robust operation under constraints such as communication latency, variable network quality, and limited sensing—aiming for practical pathways to scalable rehabilitation delivery.

2) Data-Based Control

We are interested in “learning that you can trust”—methods that combine data-driven learning with provable stability. In particular, we focus on approaches that update the models/controllers online (i.e., in real-time) while preserving a Lyapunov-based stability structure. When applied to rehabilitation systems, these data-based controllers can learn an individual's unique dynamic model, allowing for the controller to be personalized, thereby catering to the individual's specific needs. Example application areas follow:

• AI-Based Control:
  Recent advancements in artificial intelligence (AI) allow for the efficient approximation of complex dynamic functions. In particular, deep neural networks (DNNs) are capable of more efficiently approximating complex functions compared to single hidden-layer neural networks (NNs) that have long been used in control theory. The CARE Lab has contributed to the development of the first real-time training algorithms for DNN-based controllers, which guarantee trajectory tracking even when the DNN weights are initially randomized. A central goal of our work is to combine the strengths of control theory (stability, robustness, safety) with AI/learning (adaptation, function approximation, improved performance from data), resulting in systems that are not only accurate in simulation but also deployable in real-world settings.

• Concurrent Learning (CL):
  CL is a data-based approach that is often used to augment an adaptive training algorithm to improve the learning performance. Specifically, CL collects and stores data in real-time to drive adaptive estimates to converge towards their true values after a finite excitation (FE) condition has been satisfied. That is to say, CL can ensure finite-time learning, unlike traditional adaptive methods. The CARE Lab has integrated CL within a DNN-based control architecture to drive the DNN's weights towards their ideal values. 

3) Networked or Autonomous Systems

Another focus of the CARE Lab is the design of intelligent control structures for autonomous or networked control systems (NCSs), such as autonomous robots and multiagent systems (MASs). Autonomy requires real-time decision making in an often constrained environment. Networked systems are those that use a communication network to transmit signals (sensor readings, control inputs, desired trajectories). Real-world NCSs may be impacted by communication delays, intermittently dropped signals, limited power supplies and computational power, disturbances, and adversarial attacks. We develop network-aware control and estimation frameworks under realistic conditions. Example application areas follow:

• Centralized MASs:
  A centralized MAS contains a central leader agent that generates and transmits control commands to each follower agent using sensor information transmitted from the follower agents. Due to network constraints, the transmitted information to and from the leader are often delayed, resulting in the control input implemented at the follower being delayed (input delay) and the control input itself being based on delayed state information (output delay). We design control frameworks for centralized MASs that are resilient to network constraints, model uncertainties, disturbances, and adversarial attacks.

• Distributed MASs:
  A distributed MAS consists of multiple agents that share their sensor information with neighboring agents to enable collaboration. Within a distributed MAS, each agent designs its own control input. Similar to centralized MASs, network constraints often cause latency (i.e., delays) in the transmitted signals. Thus, each agent within a distributed MAS receives asynchronous information from its neighboring agents (i.e., the data received from each agent corresponds to different times). We develop resilient control frameworks and consensus algorithms for distributed MASs that approximate the true states of each agent in the network, ensure consensus throughout the network (i.e., each agent agrees on the true states of the other agents), and guarantee successful mission completion, despite network constraints, uncertainties, disturbances, and adversarial attacks.