Subscribe to our RSS feed
Assistant Professor, University of Colorado Boulder
Friday, December 7, 2018, 3 PM
Woodward Hall, Room 147
Talk Abstract: The age of autonomous machines has arrived. Yet, as products of imperfect human engineering designed to make decisions in an uncertain world, the promise of “set-it-and-forget-it” autonomy is still quite far off: autonomous systems will never operate out of the box “exactly right”. For sufficiently rich tasks that constantly push the technological cutting edge, they will encounter unexpected situations that require reasoning beyond their designed/immediate capabilities. As such, an intelligent autonomous system must be able to independently gather, process, and act on imperfect information—and be cognizant of what it can and cannot accomplish, and know when and how to seek help.
Human-machine interaction is thus a key component of autonomous system design, and must naturally connect to existing perception, planning, learning, and reasoning algorithms that enable autonomy. An autonomous robot, for instance, should ideally enable human stakeholders and user to fluidly delegate tasks, assess available information, and contribute meaningful operation improvements—without having to “babysit”, act as “band aids” (when all else fails), or think too hard about what the system is trying to do. As teammates, intelligent autonomous robots should also be able to communicate with human users to leverage their complementary abilities to improve decision making under uncertainty.
This talk will present novel Bayesian approaches to collaborative human-robot reasoning under uncertainty that can be exploited from the outset in autonomous system design. The talk will focus on probabilistic modeling, inference, and optimization techniques for augmenting autonomous optimal state estimation and planning algorithms with “plug-and-play human sensors”, connected via user-friendly semantic natural language chat and free-form map sketching interfaces. Results from collaborative human-robot teaming applications for target search and tracking applications show that these techniques allow human-machine teams to gracefully “cut knots and fill in gaps” for challenging problems—without undermining individual agent roles or ignoring their limitations.
PhD Candidate, Illinois Institute of Technology
Wednesday, September 26, 2018, 2:15 PM
ECE, Room 118
Talk abstract: In the United States, preventable medical errors are considered the third leading cause of death and may cost the economy up to $1 trillion in “lost human potential and contributions” every year. Preventable medical errors are mostly caused by unjustified deviation from applicable medical best practices. The key to reducing preventable medical errors is to assist medical staff to adhere to medical best practice guidelines through executable and verifiably safe medical best practice guideline systems. The talk presents a framework to provide verifiably safe assurance for executable medical best practice guideline systems. The framework contains five major components: (1) formally verify medical guideline statechart models, which are validated by physicians, by transforming them to timed automata, (2) design statechart model patterns to support modeling medical domain important functionalities with statecharts, (3) use runtime monitor to verify safety properties of statechart generated code, (4) model and integrate medical resource demands and availability in medical guideline models, and (5) provide traceback capability to support root causes identification of safety failures. The uniqueness of the framework is that it supports domain experts from different disciplines to participate in the entire development process without requiring the domain experts to have knowledge of other disciplines. The architecture of the framework can be applied to other safety-critical systems in general.
PhD Candidate, The University of British Columbia
Tuesday, July 31, 2018, 2 PM
ECE, Room 118
Talk abstract: Linear parameter-varying (LPV) control is a systematic way for gain-scheduling control of a nonlinear or time-varying system that has dynamic variations in its operating range. However, when the dynamic variations are large, LPV control may give conservative performance. One way to reduce the conservatism is switching LPV (SLPV) control, in which we partition the parameter variation set into subsets, design one local LPV controller for each subset, and switch among those local controllers according to some switching rules. In this talk, we present three contributions to the SLPV control theory. First, we propose a novel approach to designing SLPV controllers with guaranteed stability and performance even when the scheduling parameters cannot be exactly measured. Secondly, we show the effect of switching surfaces on the performance of an SLPV controller and give an algorithm based on particle swarm optimization to optimize the switching surfaces. Finally, this talk also includes a novel approach to designing SLPV controllers that could yield significantly improved local performance in some subsets without much sacrifice of the worst-case performance. This is different from the traditional approach that often leads to similar performance in all the subsets.
On the application side, we address three practical problems using the developed theory. The first one is control of miniaturized optical image stabilizers with product variations. Specifically, multiple parameter-dependent robust (MPDR) controllers are designed to adapt to the product variations, while being robust against the uncertainties in the measurement of the scheduling parameters that characterize the dynamics variation. The second one is air-fuel ratio control of an automotive engine, for which SLPV controllers are designed to address the variation of the dynamics with the engine speed and air mass flow. The last one is control of a floating offshore wind turbine. SLPV controllers are designed for regulating the power and the generator speed as well as reducing the platform motion in a large operating range.
Ph.D. Candidate, Clemson University
Tuesday, July 10, 2018, 2 PM
ECE, Room 118
Talk abstract: The recent emergence of safe, lightweight, and human-friendly robots has opened a new realm for human-robot collaboration (HRC) in collaborative manufacturing. For such robots with the new human-robot interaction (HRI) functionalities to interact closely and effectively with a human coworker, new HRI-based control criteria that integrate both physical and social interaction are demanded. Social human-robot interaction has been demonstrated in robots with affective abilities in education, social services, healthcare, and entertainment. Nonetheless, sHRI should not be limited only to those areas. Human trust in robot and robot anthropomorphic features may have high impacts on HRI. Human to robot trust is one of the key factors in HRI and a prerequisite for effective HRC. Trust characterizes the reliance and tendency of humans in using robots. Factors within a robotic system (e.g., performance, reliability, or attribute), the task, and the surrounding environment can all impact the trust dynamically. Over-reliance or under-reliance might occur due to improper trust, which results in poor team collaboration, and hence higher task load and lower overall task performance. This presentation summarizes intelligent control frameworks for the manipulator robots that integrate both physical and social HRI factors in the collaborative manufacturing.
Dr. Hossein Sartipizadeh
Postdoctoral Fellow, University of Texas at Austin
Visiting Scientist, Aeronautics & Astronautics, University of Washington Seattle
Tuesday, June 12, 2018, 2–3 PM
ECE, Room 118
Talk abstract: In this talk we focus on the sampling-based techniques, including the well-known scenario approach for stochastic/robust control of uncertain dynamical systems under input and state constraints. In the scenario approach, the uncertainty set is replaced by a finite set of random samples. It is proved that any feasible solution of the scenario approach will also be feasible for the original stochastic problem, although a failure risk is involved. The more the number of samples, the less failure risk will be. The main advantages of scenario approach are its generality and tractability: it converts the original stochastic problem to a deterministic convex problem regardless of the probability distribution of the uncertainty. That is, there is no need to know what the distribution is; all we need is to be able to sample from the distribution. However, to achieve a reasonable risk of failure, a large number of samples is typically needed. The scenario approach may consequently result in a computationally expensive problem since the constraints must be checked at each sample.
University of Illinois at Urbana-Champaign
Tuesday, May 15, 2018, 10 AM
ECE, Room 125E
Talk abstract: Cyber-physical systems, which interconnect continuous physical processes and discrete controlling embedded computers by a complex communication network, appear in almost every aspect of daily life: transportation, energy, medical systems, and food production. In this talk, I will present two of my works that seek to address the challenges of developing mathematically rigorous frameworks and computationally feasible algorithms for the synthesis, analysis, and verification of privacy-guarantee and secure cyber-physical systems. In the first part, I will introduce a novel framework for statistical verification of cyber-physical systems via model reduction, together with computationally feasible algorithms. In the second part, I will present the trade-off between privacy and performance in the design of differentially private communication algorithms in multi-agent systems and the optimality of a class of Laplace-noise-adding mechanism in the unbiased estimation of the private data under different statistical and information-theoretic measures.
PhD Candidate, Louisiana State University
Tuesday, May 8, 2018, 10 AM
ECE, Room 125E
Talk abstract: This talk is concerned with the formation control problem of multiple agents modeled as nonholonomic wheeled mobile robots. Both kinematic and dynamic robot models are considered. Solutions are presented for a class of formation problems that include formation, maneuvering, flocking, and target interception. Graph theory and nonlinear systems theory are the key tools used in the design and stability analysis of the proposed control schemes. Simulation and/or experimental results are presented to illustrate the performance of the controllers. In the first part, we present a leader-follower type solution to the formation maneuvering problem. The solution is based on the graph that models the coordination among the robots being a spanning tree. In the second part, we design a distance-based control scheme for the flocking and target interception of the nonholonomic agents under the assumption that only a subset of the agents know the desired flocking velocity or the target’s velocity and relative position. The control law is designed at the kinematic level and is based on the rigidity properties of the graph modeling the sensing/control interactions among the robots. The resulting controllers include distributed observers to estimate the unknown quantities. The theory of interconnected systems is used to analyze the stability of the observer-controller system.
Dr. David Copp
Postdoctoral Associate, Sandia National Laboratories
Friday, March 9, 2018, at 3 PM
Woodward Hall, Room 147
Talk abstract: In this talk, we discuss the use of control and optimization for solving sophisticated engineering problems, with motivating examples in bioengineering and energy systems. Model predictive control is a particularly popular online optimal control approach due to its ability to explicitly handle hard state and input constraints. We introduce an output-feedback approach to model predictive control for discrete-time nonlinear systems. This approach combines state estimation and control into a single min-max optimization; specifically, a criterion that involves finite forward and backward horizons is minimized with respect to control input variables and is maximized with respect to the unknown initial state as well as disturbance and measurement noise variables. Lastly, we discuss the advantages of using this combined optimal estimation and control approach in applications including the coordination of unmanned aerial vehicles, feedback control of an artificial pancreas, and potential applications in power and energy systems.
PhD Candidate, University of Michigan
Thursday, November 2, 2017, 2–3 PM
ECE, Room 118
Talk abstract: The ubiquity of security breaches in modern times necessitates the development of systems that are able to automatically detect and react to attacks, with the goal of preventing the attacker from reaching its goal. In this talk, I will present a stochastic control approach to the design of a such a system. The model is built upon the notion of a dependency graph which describes how the attacker can use its current set of capabilities to perform exploits and gain further capabilities. The defender does not perfectly observe the capabilities of the attacker at any given time and must infer them from noisy security alerts (generated by an intrusion detection system). The resulting problem of choosing defense actions that strike a trade-off between mitigation of the attacker’s progression and minimization of the negative impact to availability is formulated as a partially observable Markov decision process (POMDP). Unfortunately, due to the scale of the defense problem, obtaining an optimal solution is computationally intractable. As a result, we make of use an online solution method, termed the partially observable Monte-Carlo planning (POMCP) algorithm. The algorithm samples future possible scenarios from the current belief in order to select actions, avoiding the state-space explosion problem.