University of Illinois Urbana-Champaign


Currently, there are few examples of geometrically nonlinear adaptive civil structures in industry. This has resulted in lightweight, modular, and deployable civil structures not yet actualized and not well-studied in terms of dynamic properties and capturing their performance, establishing a cyclical problem. To break this cycle and to unleash a new era of innovation, there is a need to understand dynamic behavior of these structures, how to control them while being dynamically loaded, and how to monitor their health. This requires reflecting on the system at multi-scale in parallel for research and will be the focus of the education thrust of this proposed work. The research group examines dynamic modeling and control with a focus on structural materials under environmental loading conditions of civil infrastructure, which aligns with the grand challenges faced by the ASCE community. With American civil infrastructure rapidly aging and rehabilitation struggling to keep up, it will be critical in coming years to understand the resiliency of existing civil infrastructure under evolving environmental conditions so that we can intelligently shape the future of advanced structures and use data to predict expected maintenance of existing infrastructure. To provide buildings and bridges that will withstand vibrations due to extreme wind and seismic events, advanced infrastructure capable of adaptation will be a solution. Qualities of advanced infrastructure are directly related to global challenges of raw material usage, transportation and installation costs to rural and remote regions, ease of constructability, and structural stability under increasingly intense and uncertain environmental forces.


Structural dynamics, machine learning, and structural control are studied on various types of deployable structures, such as tensegrity structures and origami. Several large and medium-scale adaptive structures have been built to compare experimental measurements with simulations. For example, the modular tensegrity structure is 4 [m] in length, 1.5 [m] in height, and 1.5 [m] in width, preliminary work on this topology was completed and video was published at The structure will be built in two halves that deploy and connect at mid span. Constructed of four identical k-class 2 modules, the full-tensegrity structure will be a k-class 4, two modules per half. The origami structure is a four-module Miura-Ori topology, two modules wide and 2 modules long, with cell panel dimensions of 0.75 m by 0.5 m with an overall span is 2.6 m. This figure shows the dynamic relaxation model of the structure with diagonal struts with equivalent stiffness to origami panels.

Adaptive structural control

Form-finding simulations are utilized in this research for large-scale slow-actuation structures where inertial effects can be ignored and tested in experimental setups of adaptive structures such as tensegrity structures and origami. Adaptive control employs active elements integrated within the structure that locally lengthen and shorten the structure to cause global shape change. For large-scale structures, this is facilitated by mechanical actuators such as pistons and motors. Small-scale actuation can utilize changes in thermal and chemical fields.

Machine learning

Actuation and control of an adaptive structure is a complex engineering task that is best characterized by model-based learning. Combination with real-time measurements is optimal for fast-converging model-free learning, therefore bridging the gap with model-based priors for model-free reinforcement learning MBMF-RL, a hybrid system is developed in this research. Although the concept of hybrid models has been previously studied, this system specifically utilizes a nearest-neighbor non-gradient optimization algorithm for the model-based (MB) priors and Bayesian optimization for the model-free (MF) algorithm. The MB system determines the current and desired response of the physical tensegrity structure and origami shell in response to the MF actuation control commands to generate the target shell configuration. Bayesian optimization (BO) model-free (MF) algorithm is a gradient-free optimization procedure that given an unknown function finds the global minimum. Using the policy parameters determined for the actuation control, the system is trained through sample policies and storing results and cost data. These training data then initialize the objective function, which utilizes Monte-Carlo to generate the set of cases, and expected improvement (EI) acquisition function yields the next target parameter. The outcome of the BO algorithm is a set of the optimal policy parameters to minimize actuation movements of the tensegrity and origami structure to achieve the desired shape and spring stiffness.

Damage detection using vibrations

For damage detection in a complex system, a statistical tool called second-order blind identification (SOBI) can be utilized to separate the natural frequencies of the system. Natural frequencies have been identified for adaptive structures using ambient vibrations and forced vibrations such that source separation was possible. Analysis of damage location using moving window principal component analysis (MWPCA) for strain measurements utilized a variation limit of 6 sigma and error-domain model falsification (EDMF) for position measurements utilized a variation limit of 2 sigma. The wider variation limit for strain measurements was possible since MWPCA compares variations in values of eigenvectors. Position measurements used for EDMF contain more uncertainty and thus a variation limit of 2σ is acceptable since it is well above sensor resolution.

Dynamic characterization of structural systems

Much about the dynamic behavior of a structure, such as natural frequencies and damping ratio, is embedded in data. Processing of data using statistical signal processing tools can extract the embedded properties of the structure (called modal extraction). The specific method of applying SSI for this study is called a Numerical algorithm for Subspace State Space System IDentification (N4SID). The algorithm and singular value decomposition are performed to separate the natural frequencies of the system for identification. The basic framework of second-order blind identification (SOBI) is the simultaneous diagonalization of two covariance matrices Rx(0) and Rx(p) evaluated at the time-lag zero and p, respectively. Sources identify natural frequencies of the structure using a model-free method. Damage location using natural frequencies is possible when simulations are accurate and structural behavior is repeatable.


Adaptive structural control

Friction effects in the control of adaptive structures are approximated to be those related to static behavior since cables on the tensegrity structure are actuated at low-velocity. When friction behavior is included in tensegrity structure simulations, the sources of differences between simulation and test results are largely limited to those related to joint behavior. Simulated axial forces on the continuous cables were compared with measurements taken from the tensegrity structure. Measurement data was filtered to match the twenty simulated actuation steps. Differences between simulation and measurement for each actuation step have been evaluated for cases with and without friction effects. The simulation with friction effects has a mean difference from measurements of 50% less than the simulation without friction. Additionally, the standard deviations for simulations with friction are on average 40% less than simulations without friction effects.

Machine learning

Modification of control commands through modified versions of RRT*-connect, soft-constraint algorithm with case reuse exhibits biomimetics, the imitation of nature in human-made systems, through progressively reducing future execution time by at least thirty times. For the connected structure, damage mitigation between 36% and 86% was sufficient to satisfy code deflection requirements. The framework using the newly-modified RRT*-connect and the soft-constraint algorithms developed in this paper for mitigation and case reuse have potential to be applied to other active structures and is the subject of current research.

Damage detection using vibrations

Detection of a broken element is successful by observing differences of natural frequencies between healthy and damaged states. Natural frequencies are easily identified for the half-tensegrity structure comparing measured ambient vibrations and simulation. Forced-vibrations area used on the full structure such that second-order blind identification (SOBI) is feasible. Location of a damaged element is successful using nodal position measurements through excluding damage scenarios and using strain measurements to identify elements of significant changes in eigenvector coefficients using principal component analysis. Implementing error-domain model falsification to exclude possible scenarios for location of damaged elements successfully reduced the number of probable cases. An average of 70% of possible scenarios for the half structure and 71% for the full structure are excluded in the process of location. Elements with strain sensors in proximity to true ruptured elements are successfully identified as the most affected by rupture events, thus indicating the location of a ruptured cable. Therefore, the methodology involving error-domain model falsification (EDMF) for damage location is useful for closely-coupled structures that are capable of large shape changes.

Dynamic characterization of structural systems

The finite element model is used to calculate the frequencies for the first three structural modes of the origami pill bug shell structure at different stages of the rolling process. Modal frequencies show a decreasing trend over different stages of the rolling process. The frequency for Mode 1 decreases almost linearly between the unrolled (inactive) and the rolled (active) state of the origami pill bug structure.


Societal Impact – Adaptive structures combines civil engineering, robotics, and architecture. Functionally unique from traditional solutions to operable roofs and lifting bridges, these new concepts for topology of civil structures with adaptive control will provide safety measures for the public in the event of extreme wind and seismic loads. These capabilities can be extended into shape change to address thermal regulation of buildings when climate control demands peak during extreme temperatures. Work using terrestrial surveying path-planning will be useful for future vehicle structural frame safety.

Economic Competitiveness – The need for adaptive structures changing our outlook on infrastructure comes as ASCE has rated US bridges with a “C” grade, which represents a worsening condition from the previous grade in 2017 of “C+”. 7.5% are considered in poor condition, and another 48% in fair condition. The current estimated cost to carry out the repairs needed is $125 billion (ASCE Report Card, 2021).

Environmental Impact – A reduction in volume of material required for infrastructure will be made possible through adaptive control that will reconfigure building components into ideal positions when subjected to more intense loading. This results in less material and minimal power exerted for actuation while adhering to ultimate and serviceability limit states. Since adaptive structures are material-independent, they will provide a novel test bed for sustainable (wood) and alternative building materials (fiber-based materials and composites) that gain points in the USGBC Leadership in Energy and Environmental Design (LEED) program. Through the practice of green building design, work would present an application of traditional and non-traditional recycled materials.

Student involvement - SMARTI lab works with a team of undergraduate students to develop virtual reality and physical manifestations of large-scale adaptive structures. Active in the Worldwide Youth in Science and Engineering summer camp and in Society of Women Engineers (SWE), SMARTI lab champions equitable access to engineering within and beyond the walls of the University of Illinois.

Core competencies

  • adaptive structures
  • machine learning
  • damage detection
  • structural dynamics
  • sensor systems and visualization

Current members

  • Angshuman Baruah, PhD candidate
  • Sagnik Paul, PhD candidate
  • Juan Torres, MS graduate student
  • Kaylee Tucker, MS graduate student

Recent graduate

Heather Gathman


Joe Tom (UIUC), Ryan Beemer (U Mass Dartmouth), Ramez Hajj (UIUC), Nishant Garg (UIUC), Jacob Henschen, Marci Uihlein (UIUC), Eric Shaffer (UIUC), Lesley Sneed (UIC)

Funding agencies

National Science Foundation: Engineering Civil Infrastructure, Discovery Partners Institute, Institute for Sustainability Energy and Environment