!DOCTYPE html> Jon DeCastro | Ph.D. Candidate, Cornell University

Jonathan A. DeCastro

Toyota Research Institute

I am currently a Researcher at the Toyota Research Institute in Cambridge, MA. I received my Ph.D. in Mechanical and Aerospace Engineering at Cornell University as an affiliate of the Verifiable Robotics Research Group.

My research interests are in the construction of provably-correct controllers for dynamical systems, enabling complex tasks to be verified for a wide class of complex robots. My work spans two primary research thrusts: (1) computational approaches to synthesize continuous controllers for nonlinear systems that are verifed to fulfill high-level specifications, and (2) automatic discovery of certificates for guaranteeing controller execution in dynamic environments. The latter is meant to enable human-robot teams in order to guarantee tasks carried out using robots with dynamics. My work draws from a broad spectrum of ideas, including optimization, hybrid systems, formal methods, and control systems.

One Kendall Square
Cambridge, MA 02139
jad455 - at - cornell - dot - edu



J. A. DeCastro, R. Ehlers, M. Rungger, A. Balkan and H. Kress-Gazit, “Automated Generation of Dynamics-Based Runtime Certificates for High-Level Control,” To appear in the Journal of Discrete Event Dynamical Systems: Special Topical Issue on Formal Methods in Control, 2017.   [bibtex] [link]

J. Alonso-Mora, J. A. DeCastro, V. Raman, D. Rus and H. Kress-Gazit, “Reactive Mission and Motion Planning while Avoiding Dynamic Obstacles,” (in submission) 2016.

Journal Articles

J. A. DeCastro and H. Kress-Gazit, “Synthesis of Nonlinear Continuous Controllers for Verifiably-Correct High-Level, Reactive Behaviors,” International Journal of Robotics Research, 34(3):378-394, 2015.   [bibtex] [link]

J. A. DeCastro, “Rate-Based Model Predictive Control of Turbofan Engine Clearance,” AIAA Journal of Propulsion and Power, 23(4):804–813, 2007. [bibtex] [link]

Refereed Conference Publications

J. A. DeCastro and H. Kress-Gazit, “Nonlinear Controller Synthesis and Automatic Workspace Partitioning for Reactive High-Level Behaviors,” 19th ACM International Conference on Hybrid Systems: Computation and Control (HSCC). Vienna, Austria, 2016.   [bibtex]

J. A. DeCastro, J. Alonso-Mora, V. Raman, D. Rus and H. Kress-Gazit, “Collision-Free Reactive Mission and Motion Planning for Multi-Robot Systems,” International Symposium on Robotics Research (ISRR). Sestri Levante, Italy, 2015.   [bibtex]

J. A. DeCastro, V. Raman and H. Kress-Gazit, “Dynamics-Driven Adaptive Abstraction for Reactive High-Level Mission and Motion Planning,” IEEE International Conference on Robotics and Automation (ICRA). Seattle, WA, USA, 2015.   [bibtex]

J. A. DeCastro and H. Kress-Gazit, “Guaranteeing Reactive High-Level Behaviors for Robots with Complex Dynamics,” IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). Tokyo, Japan, 2013.   [bibtex]

J. A. DeCastro, L. Tang, B. Zhang and G. Vachtsevanos, “A Safety Verification Approach to Fault-Tolerant Aircraft Supervisory Control,” AIAA Guidance, Navigation, and Control Conference (GNC). Portland, OR, USA, 2011. [bibtex]

J. A. DeCastro, L. Tang, C. S. Byington and D. E. Culley, “Analysis of Fault-Tolerance and Decentralization Concepts for Distributed Engine Control,” 45th AIAA Joint Propulsion Conference and Exhibit (JPC). Denver, CO, USA, 2009. [bibtex]

Technical Reports

J. A. DeCastro, R. Ehlers, M. Rungger, A. Balkan, P. Tabuada and H. Kress-Gazit, “Dynamics-Based Reactive Synthesis and Automated Revisions for High-Level Robot Control,”.   [arXiv]


Automatic Synthesis High-Level, Reactive Controllers for Nonlinear Systems

Can we automatically synthesize controllers for robots with real-world dynamics that are capable of fulfilling complex tasks in human environments?

To answer this question, I have adopted an approach for constructing a library of atomic controllers for a wide class of nonlinear systems such that, collectively, guarantee the sequence of motions as requested by the high-level controller (usually represented as a finite-state machine). If the computation is successful, such high-level controllers are correct by construction with respect to the library of low-level controllers that are computed for the intended robot.

I have also developed a computational approach that leverages nonlinear systems theory to construct controllers and regions of invariance that respect the constraints of a nonlinear dynamical system in a highly-constrained workspace (e.g. one filled with many obstacles).

Sometimes the high-level controller produces behaviors that cannot be fulfilled at the low level. The synthesis techniques have been extended in order to adapt the finite-state approximate model of the robot, or its discrete abstraction, greatly reducing the conservatism preventing such controllers to be synthesized. Our approach adapts these abstractions only when necessary, thereby eliminating any unneeded computations in synthesis.

Please see: [DeCastro & Kress-Gazit (IROS 2013)] [DeCastro & Kress-Gazit (IJRR 2015)] [DeCastro, Raman & Kress-Gazit (ICRA 2015)] [DeCastro & Kress-Gazit (HSCC 2016)]

Interactive Specification Design for Guaranteeing Tasks Carried Out by Robots with Dynamics

In teams of people, individuals naturally adapt according to the capabilities of other teammates. Can we establish human-robot teams that proceed without failure by giving the user full awareness of the robot's capability to perform a given task?

When composing the specification for a task, a user must specify the desired robot behaviors along with any assumptions on the sensed environment – the latter are often difficult to predict and encode. If such assumptions are too conservative, there may be no satisfying controller for certain types of dynamical systems (e.g. fixed-wing airplanes, quadrotors, cars). In this light, my aim is to inform the user as to the conditions where the environment must cooperate with a certain robot in order to guarantee the success of the task.

Drawing from recent works in assumption mining in formal methods, we take a discrete representation of the physics of a robot, a logical task specification, and the workspace, then automatically generate a set of environment behaviors that must hold in order for there to exist a high-level controller for the task. The resulting revisions are suggested to the user and, if accepted, become certificates that hold at runtime. We have shown that our approach allows high-level controllers to be synthesized for a wide range of complicated tasks on realistic physical systems.

Please see: [DeCastro, et. al.] [DeCastro, et. al. (ISRR 2015)] [DeCastro & Kress-Gazit (HSCC 2016)]

Low-Complexity Synthesis for Multi-Robot Scenarios with Dynamic Obstacles

In multi-robot scenarios operating in human environments, controller synthesis can quickly become intractible due to the exponential growth of complexity with the number of robots and dynamic obstacles present in the environment. We are collaborating with MIT to form a high-level synthesis paradigm in which we use a local planner with collision avoidance guarantees at the low level, and a high-level controller to coordinate the actions of the robots. To avoid complexity, we do not require observation of the global behavior of the dynamic obstacles; instead the controller reasons about behaviors of the dynamic obstacles locally, yet preserves global guarantees of task satisfaction. Through experiments, we show that we can synthesize robot controllers that are lower in complexity than approaches in which all the obstacles are explicitly modeled.

Please see: [DeCastro, et. al. (ISRR 2015)]


Architectures for Modeling and Controlling Aircraft Propulsion Systems

I have been involved in several projects for NASA, including development of the software tool C-MAPSS, a “virtual” engine model accessible to a wide arena of resesarchers.

Another project focused on fault-tolerant control algorithms for aircraft. The objective was to show flight control when one of the acutators has failed, while minimizing the probability of exceeding flight control limits and failure of any other aircraft components. We approached the control design as an optimization problem (solved with an offline dynamic programming algorithm). When demonstrated on a commercial jet simulator, the controller was able to land an airplane when the rudder had been suddenly damaged.

Please see: [DeCastro, et. al. (2011)]

A random assortment of photos.