Speakers

Talk Abstract: In this presentation we explore an imitation and transfer learning setting for Linear Quadratic Gaussian (LQG) control, where (i) the system dynamics, noise statistics and cost function are unknown and expert data is provided to learn the LQG controller, and (ii) multiple control tasks are performed for the same system but with different LQG costs. We show that the LQG controller can be learned from a set of expert trajectories. Further, the controller can be decomposed as the product of an estimation matrix and a control matrix. This data-based separation principle allows us to transfer the estimation matrix across different LQG tasks, and to reduce the length of the expert trajectories needed to learn the LQG controller.

Bio: Taosha Guo received her M.Phil degree in electronic and computer engineering from the Hong Kong University of Science and Technology, Hong Kong, in 2021. She is currently a Ph.D. candidate in Mechanical Engineering at the University of California, Riverside, under the supervision of Dr. Fabio Pasqualetti. Her main research interests include data-driven control, security of cyber-physical systems, and transfer learning.

Preston Culbertson

Talk Title: Real-time Risk-sensitive Control via Generative Modeling and Stochastic Control Barrier Functions

Talk Abstract: A key source of brittleness for robotic systems is model uncertainty and external disturbances. Existing approaches to robust control either assume a "worst-case" disturbance, or impose step-wise chance constraints (yielding weak trajectory-long safety probabilities). We propose an alternate approach: offline, we train a deep generative model to represent stochastic disturbances, and online use stochastic control barrier functions (CBFs) to provide safety guarantees. We demonstrated our approach by performing aggressive quadrotor flight with an unmodeled slung load. We find combining data-driven disturbance models with stochastic CBFs yields a real-time approach to robust control with rigorous probabilistic safety guarantees.

Bio: Preston Culbertson is a postdoc in the AMBER Lab at Caltech, working with Prof. Aaron Ames to study risk-sensitive robot planning and control. Preston completed his PhD at Stanford University, working under Prof. Mac Schwager on multi-robot manipulation and assembly. Preston's research aims to leverage modern, data-driven perception stacks (in particular, vision-based geometry representations like NeRFs) for difficult tasks like dexterous manipulation, locomotion, and agile drone flight.

Berkay Citmaci

Talk Title: Machine Learning-Based Predictive Control Using On-line Model Linearization: Application to an Experimental Electrochemical Reactor

Talk Abstract: A model predictive control scheme that uses a neural network model as the process model to implement real-time multi-input-multi-output control in an electrochemical reactor for CO2 reduction is demonstrated. The Koopman operator method is used to linearize the LSTM model to reduce the nonlinear optimization step in the MPC to a quadratic programming problem. The performance of the MPC using the linearization of the LSTM model are evaluated with various simulations as well as open-loop and closed-loop experiments. It can drive the two process output states, that are the concentrations of the products to their desired set-points by computing optimal input variables in real-time in closed-loop experiments.

Bio: Berkay Citmaci is a Fulbright alumni, 4th year Chemical Engineering Ph.D. student in Chemical Engineering and a member of Christofides research lab working on automation, modeling, optimization, and control of experimental electrochemical systems.

Jie Feng

Talk Title: Stability-Constrained Reinforcement Learning for Voltage Control

Talk Abstract: DRL's deployment in real-world power systems has been hindered by a lack of explicit stability and safety guarantees. In this talk we propose a stability-constrained reinforcement learning (RL) method for real-time implementation of voltage control, that guarantees system stability both during learning and deployment. The key idea is an explicitly constructed Lyapunov function that leads to a sufficient structural condition for stabilizing policies, i.e., monotonically decreasing policies guarantee stability. Furthermore, we introduce the safe-gradient flow framework to optimize the steady-state performance of voltage control under limited reactive power resources.

Bio: Jie Feng received the B.E. degree in Automation from Zhejiang University, Hangzhou, China, in 2021. He is currently pursuing his Ph.D. degree in Electrical and Computer Engineering at the University of California, San Diego, advised by professor Yuanyuan Shi. His research interests focus on stability-constrained machine learning for power system control.

Asmaa Eldesoukey

Talk Title: Schrodinger's Control and Estimation Paradigm on Graphs with Absorbing States

Talk Abstract: First-passage time statistics grew indispensable in modeling stochastic processes such as neuron firing, bacterial steering via switching, etc. Our work is concerned with regulating first-passage (stopping) time distributions, as target spatial distributions at fixed times may be stringent. In applications such as S/C landing, the landing time may be, ipso facto, random. In the spirit of Schrodinger Bridges, this control problem can be cast as a maximum likelihood problem. To this end, we tackle, in a discrete-time Markov chain setting, the problem of how to find the most likely trajectories ensuing from the control, in the form of transitions, to meet the stopping time marginals at certain nodes.

Bio: Asmaa graduated in 2018 from Zewail University of Science and Technology with an Aerospace Engineering Degree. In 2021, she obtained her Master's in Dynamics and Controls from the University of California, Irvine. Currently. She is a PhD candidate at the same school and is advised by Prof. Tryphon Georgiou. She is interested in stochastic systems and the interplay between information theory and networks.

Bryce L. Ferguson

Talk Title: Collaborative Decision-Making and the k-Strong Price of Anarchy in Common Interest Games

Talk Abstract: With advances in communication and computation technologies, we can consider new collaborative decision-making paradigms that exist somewhere between centralized and distributed control. In this talk, I will discuss this in the context of common interest games in which groups of up to k agents can coordinate their actions in maximizing a common objective function. Studying the equilibria of these systems provides relevant insights into the efficacy of inter-agent collaboration. I will specifically present bounds on how well k-strong Nash equilibria approximate the optimal system welfare, as well as a study on the run-time and transient performance of collaborative agent-based dynamics.

Bio: Bryce L. Ferguson is a PhD candidate in the Electrical and Computer Engineering Department at the University of California, Santa Barbara. Bryce received his BS and MS in Electrical Engineering from the University of California, Santa Barbara in June 2018 and March 2020 respectively. He was a 2022 CPS Rising Star and a finalist for the Best Student Paper Award at the 2020 American Controls Conference. Bryce's research interests focus on game theoretic methods for controlling multi-agent systems.

Paul Serna-Torre

Talk Title: Non-cooperative Game to Control Learned Inverter Dynamics of Distributed Energy Resources

Talk Abstract: We propose a non-cooperative game-based control scheme to control DERs in a microgrid to provide regulation services in support to the upper-level grid. We show that the control scheme we propose enables savings up to 9.3 to 208 times in the DERs objective cost functions and a time-domain response with no oscillations with up to 3 times faster settling times relative to using droop and PI control. Two virtues of the proposed control scheme are that: (i) it employs learned VSI dynamics that reduce the complexity of deriving and computing the full model of DERs with dq-control schemes, and (ii) that it considers the potentially selfish nature of DERs and realistic DER dynamics.

Bio: As a third-year Ph.D. student under the guidance of Dr. Patricia Hidalgo-Gonzalez, I am conducting research on the control of Distributed Energy Resources and least-cost optimization-based planning of the Western Interconnection. Before starting my Ph.D. program, I held positions in the Peru National Electricity Operator, Peru Energy Commission and transformer design companies for a span of 5 years.

Jimmy Ragan

Talk Title: s-FEAST: Safe Fault Estimation via Active Sensing Tree search

Talk Abstract: Autonomous systems must be robust to simultaneous component faults among actuators and sensors. s-FEAST (Safe Fault Estimation via Active Sensing Tree search), is a method that plans control inputs to quickly and safely gain information to estimate failures on board a spacecraft in the presence of process and measurement noise. In this talk we will develop our algorithm for active fault estimation in two parts. First, we extend tree search planning in partially observable settings to belief-based planning for information gathering problems. We then consider probabilistic safety constraints, to ensure safety over the planning horizon. We present theoretical and empirical validation of our method, including real-time hardware experiments.

Bio: Jimmy Ragan is a Ph.D. candidate in Space Engineering in the Graduate Aerospace Laboratory at the California Institute of Technology. His research is into applying autonomous planning methods to fault estimation in domains requiring fast autonomous responses such as deep-space proximity operation or formation flying. He has received B.S. degrees in Aerospace Engineering as well as Astronomy and Physics from the University of Washington.

Will Sharpless

Talk Title: HJ Reachability in the Koopman Space with the Hopf Formula

Talk Abstract: The Hopf formula for Hamilton-Jacobi reachability can solve high-dimensional differential games, producing the set of initial states and the corresponding controller required to reach (or avoid) a target for any bounded disturbance. As a space-parallelizable optimization problem, the Hopf formula avoids the curse of dimensionality but is restricted to linear time-varying systems. To solve high-dimensional nonlinear systems, we pair the Hopf solution with Koopman theory for global linearization. By first lifting a nonlinear reachability problem to a linear space and then solving the Hopf formula, approximate reachable sets and the controller can be computed efficiently, out-performing local linearizations and MPC-based Koopman controllers.

Bio: Will is a Ph.D. student at UCSD working in the SAS lab on applications of HJR algorithms to high-dimensions. His interests revolve around control and optimization in nonlinear, stochastic systems for autonomy in robotics, medicine and economics. As an undergraduate, Will studied applied math and biology at UC Berkeley, during which he discovered a fascination for the theory of nonlinear systems and control that arose in the metabolic networks and cellular ecology.

Charles A. Johnson

Talk Title: Nonlinear data-driven control via Koopman models: the interplay of model approximation and subspace closure in stabilization

Talk Abstract: Koopman system identification techniques can identify nonlinear dynamics by learning a linear model that is higher dimensional than the data. The resulting model has two components. The model approximation component (MAC) and the subspace closure component (SCC). The MAC captures the evolution of the measurements in time. The SCC handles the remaining dynamics in the Koopman model. When designing controllers for a Koopman model, the parameters in the MAC should be fixed and untouched. However, we demonstrate control design strategies that show, that the SCC of the Koopman model is subject to the control design process. This implies that data-driven control strategies should be distinct from sysID strategies that prioritize best fit.

Bio: Charles is a PhD student at UCSB who studies the mathematics of predictive modeling, specifically machine learning and how to model dynamic systems from data. His research uses operator theory to study (1) the output and consequences of data-driven learning, (2) decision-making architecture in biological systems, (3) data-driven methods for learning and controlling biological processes and (4) the computational theory of what makes superbugs resilient to extreme, environmental change.

Fahimeh Arab

Talk Title: Causal Discovery for Control: Navigating Large-Scale Low-Resolution Time-Series with Feedback

Talk Abstract: While correlation is sufficient for prediction, control inputs only affect the system through links that are causal. Traditionally, system identification has resembled Granger causality, relying on temporal order to infer causation from correlation. Yet, contemporaneous effects can arise in data with low sampling resolution, as is prevalent in neuroscience, climate analyses, social sciences, etc. We introduce CaLLTiF, a novel method for causal discovery from low-resolution time-series, particularly in large systems with abundant feedback. I demonstrate CaLLTiF's outstanding performance compared to state-of-the-art using synthetic data with known ground truth and its ability to discover remarkable and meaningful structures from human fMRI.

Bio: Fahimeh holds a BSc and MSc in Electrical Engineering from Amirkabir University of Technology in Iran. Currently, she's pursuing her PhD at the University of California Riverside under the guidance of Professor Erfan Nozari. Her research focuses on deciphering neural mechanisms in the human brain associated with various cognitive processes, employing techniques like system identification and causal discovery.

Skylar X. Wei

Talk Title: Safety-Critical Autonomy in Complex Environment via Data-Driven Set Prediction and Probabilistic Uncertainty Quantification

Talk Abstract: Safety for robot exploration in complex environments is critical. Robots encounter various types of uncertainties, including model uncertainty in system dynamics, and collision uncertainties given dynamic obstacles. By obtaining probabilistic upper bounds on model discrepancies from data, we obtain a maximum tracking tube that upper-bounds the closed-loop tracking deviations. For collision uncertainty, we propose a data-driven predictor and uncertainty estimation pair to predict obstacle-occupied regions. Using adaptive conformal prediction, we adjust the predicted unsafe regions in a distribution-free manner. By recasting safety requirements as state constraints, we provide a provably model predictive planner and is validated on hardware

Bio: Skylar X. Wei is a Ph.D. candidate at the Caltech in Controls and Dynamical System, specializing in safety-critical autonomy in a risk-aware context. She leverages data to make probabilistic unsafe sets predictions, enabling the planning of optimal and safe paths with dynamic obstacles and model discrepancies. She received a B.S. and M.S. degrees in Aerospace Engineering from UCLA in 2018, with a minor in Mathematics. Prior to her Ph.D., she was a GNC engineer at the NAVAIR, China Lake.

Junheng Li

Talk Title: Toward Dynamic Locomotion and Loco-manipulation on Humanoid Robots via Model Predictive Control with Linear Dynamics Models

Talk Abstract: In this presentation, I will share a series of research focused on dynamic humanoid locomotion and loco-manipulation via Model-Predictive-Control (MPC). Specifically, we've proposed a Force-and-moment-based MPC, utilizing a simplified rigid body dynamics model and linear state-space dynamics. The work also features further extension of this MPC including adaptive frequency MPC with offline trajectory optimization for dynamically traversing stepping stone terrains, and multi-contact MPC for adressing humanoid robot carrying and manipulating heavy load. To substantiate our control advancements, we developed high-fidelity simulation framework and in-house small-scale humanoid robot (HECTOR) for experimental validation.

Bio: Junheng Li is a second-year Ph.D. student at the University of Southern California majoring in mechanical engineering. Junheng joined Dynamic Robotics and Control Laboratory (DRCL) in 2020 under the advisement of Professor Quan Nguyen. Junheng's research focuses on bipedal/humanoid locomotion and loco-manipulation control via model-based optimal control schemes such as MPC and Trajectory Optimization. Furthermore, Junheng is also leading the open-source humanoid robot project, HECTOR.

Ulisses Santa Cruz Leal

Talk Title: Assured Perception for Estimation and Control of Autonomous Aircraft Landing

Talk Abstract: This talk focuses on formal verification of a NN-based autonomous landing system. The NN controller guides an aircraft using processed camera images. One significant challenge is the absence of mathematical models linking system states and processed images. We present NNLander-VeriF, a framework for verifying vision-based NN controllers. It uses geometric models of cameras to represent aircraft states and NN inputs. The transformed model, integrated into a NN with manual weights, captures the relationship. The expanded NN enables closed-loop verification, manageable by existing NN model checkers. We assess our framework's effectiveness in verifying a trained NN and discuss ongoing work on a certified vision-based state estimator.

Bio: Ph.D. Candidate at the Electrical Engineering and Computer Science (EECS) Department, University of California, Irvine. Specializing in formal verification of perception-based controlled vehicles under the esteemed guidance of Prof. Yasser Shoukry at the Resilient Cyber-physical Systems Lab. Passionate about safety in autonomous systems with particular focus on aerospace applications. Interests: Control Systems, Machine Learning, Machine Vision, Aerospace Eng., Formal Methods.

Zhichao Liu

Talk Title: Modeling, Control and Impact-Inclusive Motion Planning of Compliant Resilient Aerial Robots

Talk Abstract: In this study, we analyze the effect of compliance through dynamic modeling and demonstrate that the inclusion of passive springs enhances impact resilience. The impact resilience is extensively tested to stabilize the MAV following wall collisions under high-speed and large-angle conditions. Additionally, a new collision-inclusive planning method that aims to prioritize contacts to facilitate aerial robot navigation in cluttered environments is proposed. Our proposed compliant robot and CP planning method can accelerate computation time in offline and online while having shorter trajectory time and larger clearances compared to A* and RRT* planners with velocity constraints.

Bio: Zhichao Liu is a postdoc at UCR. In 2023, he completed his PhD in department of ECE at UCR, supervised by Prof. Konstantinos Karydis. Previously, he obtained a MS degree at PENN and a BE degree at BUPT. Zhichao is interested in perception, navigation and control of aerial and mobile robots under contacts in challenging environments.

Page updated

Report abuse