Cyber physical systems (CPS) are instrumental to current and future Department of Defense (DoD) mission needs – unmanned vehicles, weapon systems, and mission platforms are all examples of military-relevant CPS. These systems and platforms integrate cyber and physical subsystems, and the enormous complexity of the resulting CPS has made their engineering design a daunting challenge. An immediate consequence of this complexity is development cycles with prolonged timelines that challenge DoD’s ability to counter emerging threats.
CPS design is a complex endeavor that involves many domains – from cyber (e.g., software, control, computing, and communication) to physical (e.g., structural, mechanical, thermal, etc.) to manufacturing – and upwards of hundreds of domain-specific tools orchestrated by large teams of engineers with extensive domain knowledge and subject matter expertise. Current engineering design processes start with requirements-driven decomposition into discipline-specific design flows that, at their core, are concurrently running sequential decision-making processes that involve generating candidate architectures; evaluating, selecting, and refining options; and integrating the design until requirements are satisfied.
The Information Innovation Office launced new Symbiotic Design for Cyber Physical Systems program and to facilitate teaming, in August 2019. The goal of the program is to develop AI-based approaches to enable correct-by-construction design of military-relevant, cyber-physical systems (CPS), in order to reduce the time from their inception to deployment from years to months, and enhance innovation in design. These approaches would complement and augment existing model-based design technologies, and enable humans and computers to collaborate on correct-by-construction design of CPS.
The goal of the Symbiotic Design for CPS (SDCPS) program is to develop AI-based approaches to enable correct-by-construction design of military-relevant CPS. SDCPS seeks to reduce the time from CPS inception to deployment from years to months, and enhance innovation in design. To accomplish this, SDCPS will address the following three intrinsic challenges:
- Predictability – The soundness of design decisions relies on accurate predictions of performance prior to the implementation of software and physical components. However, accurate predictions require high-fidelity models that are cost- and time-prohibitive to produce. Cost-effective modeling processes produce results with substantial uncertainty.
- Convergence – Design teams are federated according to discipline boundaries. However, separation of concerns within a complex system neglects the interdependence of design decisions, rendering rapid convergence to a viable integrated solution practically impossible.
- Exploration – Limited by time and resources, engineers have to make tradeoffs that constrain the exploration of the design space to the familiar and known-feasible. This leaves vast areas of design space unexplored, which may contain unconventional but highly performant solutions.
The vision of the program is to vastly expand coverage and accelerate exploration of CPS design spaces with the symbiosis of two very different kinds of agents: humans with their uncanny ability to create intuitive associations across design domains, and machines with their ability to recognize statistical patterns from data and navigate vast search spaces for optimal solutions. The program aims to realize this vision by transforming the human-focused model-based design flows used today into a symbiotic process of collaborative discovery by humans and continuously learning AI-based co-designers.
DARPA SDCPS awards
LOGiCS project receives $8.4M DARPA grant in Dec 2020
Learning-Based Oracle-Guided Compositional Symbiotic Design of CPS (LOGiCS), a project led by Prof. Sanjit Seshia with a team that includes Profs. Prabal Dutta, Björn Hartmann, Alberto Sangiovanni-Vincentelli, Claire Tomlin, and Shankar Sastry, as well as alumni Ankur Mehta (EECS Ph.D. ’12, advisor: Kris Pister) and Daniel Fremont (CS Ph.D. ’20, advisor: Sanjit Seshia), has been awarded an $8.4M Defense Advanced Research Projects Agency (DARPA) grant as part of their Symbiotic Design of Cyber-Physical Systems (SDCPS) program.
CPS has applications not only for DARPA missions but also in areas such as agriculture, environmental science, civil engineering, healthcare, and transportation. SDCPS is a four-year program which aims to “develop AI-based approaches that partner with human intelligence to perform ‘correct-by-construction’ design for cyber-physical systems, which integrate computation with physical processes.” LOGiCS takes a novel approach that blends AI and machine learning with guidance from human and computational oracles to perform compositional design of CPS such as autonomous vehicles that operate on the ground, in the air and in water to achieve complex missions.
“Our primary role is to develop algorithms, formalisms and software for use in the design of CPS,” said Seshia. “These techniques allow designers to represent large, complex design spaces; efficiently search those spaces for safe, high-performance designs; and compose multiple components spanning very different domains — structural, mechanical, electrical and computational.”
AI-assisted CPS Design
The project is part of the Symbiotic Design for CPS (SDCPS) program, with a goal to develop AI-based approaches to enable correct-by-construction design of military-relevant CPS. Beyond novel theoretical discoveries we focus our innovation and research efforts to deliver AI-based Co-Designers that are integratable with the dominantly model-based Cyber-Physical System (CPS) design flows and tool suites. Our vision is the reformulation of the conventional engineering process of CPS as a continuously learning, self-improving process of collaborative discovery. Breakthroughs will emerge from the symbiosis of new, AI-based data-driven approaches in design flows to complement human intuitions and classical analytics for synthesizing and validating candidate solutions.
The project is led by the Institute for Software Integrated Systems of Vanderbilt, and includes collaborators from University of Alberta, Canada and University of Szeged, Hungary. Vanderbilt’s Péter Völgyesi (PI) has over two decades of experience with model-based design, design automation and integration platforms. The Institute for Software Integrated Systems has pioneered generations of metaprogrammable tool suites for modeling and model transformation and their use in design automation.
The University of Alberta team, led by Prof. Csaba Szepesvári, has developed several fundamentally novel AI/ML algorithms that led to breakthroughs, such as DeepMind’s AlphaGo. As the lead of the foundation group at Google’s DeepMind, Prof. Szepesvári has a broad perspective on recent advancements in AI that can change the status quo in model-based design automation. Prof. Miklós Maróti, the lead of the mathematics research team at University of Szeged, has foundational work in applying AI methods within mathematics: augmenting SAT solvers with AI-based approximations to solve algebraic problems and proving stability properties of dynamical systems by learning their Lyapunov functions.
Contact-aware robot design reported in July 2021
Adequate biomimicry in robotics necessitates a delicate balance between design and control, an integral part of making our machines more like us. With machines, designing a new robotic manipulator could mean long, manual iteration cycles of designing, fabricating, and evaluating guided by human intuition.
Most robotic hands are designed for general purposes, as it’s very tedious to make task-specific hands. Existing methods battle trade-offs between the complexity of designs critical for contact-rich tasks, and the practical constraints of manufacturing, and contact handling.
This led researchers at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) to create a new method to computationally optimize the shape and control of a robotic manipulator for a specific task. Their system uses software to manipulate the design, simulate the robot doing a task, and then provide an optimization score to assess the design and control.
Such task-driven manipulator optimization has potential for a wide range of applications in manufacturing and warehouse robot systems, where each task needs to be performed repeatedly, but different manipulators would be suitable for individual tasks.
Traditionally, this joint optimization process consists of using simple, more primitive shapes to approximate each component of a robot design. When creating a three-segment robotic finger, for example, it would likely be approximated by three connected cylinders, where the algorithm optimizes the length and radius to achieve the desired design and shape. While this would simplify the optimization problem, oversimplifying the shape would be limiting for more complex designs, and ultimately complex tasks.
To create more involved manipulators, the team’s method used a technique called “cage-based deformation,” which essentially lets the user change or deform the geometry of a shape in real-time.
Using the software, you’d put something that looks like a cage around the robotic finger, for example. The algorithm can automatically change the cage dimensions to make more sophisticated, natural shapes. The different variations of designs still keep their integrity, so they can be easily fabricated.
A simulator was developed by the team to simulate the manipulator design and control on a task, which then provides a performance score. “Using these simulation tools, we don’t need to evaluate the design by manufacturing and testing it in the real world,” says Jie Xu, MIT PhD student and lead author on a new paper about the research. “In contrast to reinforcement learning algorithms that are popular for manipulation, but are data-inefficient, the proposed cage-based representation and the simulator allows for the use of powerful gradient-based methods. We not only find better solutions, but also find them faster. As a result we can quickly score the design, thus significantly shortening the design cycle.” The work is supported by the Toyota Research Institute as well as the DARPA Symbiotic Design for Cyber Physical Systems (SDCPS) program.
DARPA awards on JARVIS in Nov 2021
Charles River Analytics, along with its partners at Raytheon Technologies‘ Collins Aerospace business, has received a $6 million contract from the U.S. Air Force and the Defense Advanced Research Projects Agency for the development of a platform envisioned to combine human expertise and artificial intelligence, Charles River Analytics announced in November 2021. CRA said the team will work on a prototype Joint Adaptive, Robust Visualization and Interaction System for a joint USAF and DARPA program that focuses on cyber-physical systems design.
Cyber-physical systems design is extremely complex, involving dozes of domains, and requiring subject matter expertise from designers and engineers across many disciplines. JARVIS has the potential to allow humans and machines to work collaboratively, mitigating the potential biases of human designers, while still incorporating human insights and creativity into the design process. Part of this vision is that designers will also be able to review and direct the activity of AI co-designers throughout the systems engineering lifecycle, providing critique and guiding design outcomes.
The Air Force and DARPA eye JARVIS to enable human operators to use their insights and creativity throughout the design process while AI algorithms help mitigate human designers’ potential biases.
JARVIS is being developed under the Symbiotic Design for Cyber Physical Systems (SDCPS) program. The goal of SDCPS is to enhance innovation in design through AI-based approaches that dramatically reduce the time from system inception to deployment (from years to months). The team plans for JARVIS to provide the bridge between that AI tech and human experts, and enable engineering teams to create highly performant military-relevant cyber-physical systems on unprecedented timelines.
“What’s interesting about JARVIS and the SDCPS program is that we’re not just using AI to exhaustively explore a design space; we’re also creating tools that help engineering teams discover and understand promising—but possibly very unconventional—AI design outcomes, injecting their own unique expertise to seed and refine these designs as part of a collaborative human-machine process,” said Dr. Ryan Kilgore, Director of Human-Centered Intelligence Systems and Principal Investigator on the JARVIS effort.