Intelligent Machines? – Self-organized Nonlinear Dynamics of Machines across Scales

For a variety of reasons, mechanical engineering systems, like aircraft, space structures, vehicles, and the like, have always been said to be designed and operated at the limit of the feasible. Traditionally, complexity in this context has mostly been understood in terms of optimally exploiting physical processes and other variables of design spaces. The resulting machines typically consist of many thousands of interlinked and interacting components, which are often operating in highly sophisticated time-dependent manners. Still, the majority of design paradigms in mechanical engineering are largely dominated by the concepts of static equilibrium or simple periodicity. Since the early days of the industrial revolution, this approach has undoubtedly been highly successful. Often, however, the price was an over-design of machines to cope with a wide envelope of complex loading and response states. It is only rather recently that the wish for further system optimisation, e.g. towards truly light-weight structures, a stronger and stronger individualization of machines, etc, and a growing amount of machine intelligence questions the traditional approach. Findings from research, as well as from business innovations pose the question if innovative future mechanical engineering systems will also be designed and operated on the basis of truly complex dynamics.

Today’s robots rely predominantly on rigid metallic components and electromagnetic motors, which make them heavy, expensive, and ill-suited to unstructured environments. Hydraulically amplified, self-healing, electrostatic (HASEL) actuators are a new versatile class of soft actuators with muscle-like behavior, which may overcome some of the drawbacks of hard robots. HASEL actuators consist of liquid-filled deformable shells and use an electrohydraulic mechanism for actuation. This talk gives a brief overview of the HASEL technology and discusses the physics of electrohydraulic actuation. It will show how geometrical and materials parameters govern the nonlinear quasistatic and dynamic actuation behaviors of HASEL actuators and point towards ways to drastically improve their performance.

This talk addresses several aspects of complex dynamics observed in friction-affected and friction-excited mechanical structures. On the example of automotive brake systems, the multi-scale nature of the self-excited vibrations are studied by means of nonlinear time series analysis, state space reconstruction, scale separation and machine learning. Experimental observations cover load-history effects, multi-stability as well as transitions from regular to irregular dynamics. Challenges for physics-based modelling are discussed and an outlook is given towards physical machine learning methods.

Ising Machines are physical implementations of systems whose collective dynamics lead to asymptotic states that are related to the minimization of the Ising Hamiltonian. As shown in, e.g., [1], many NP hard problems can be mapped onto the Ising Hamiltonian. Hence, if an Ising Machine can be build and scaled, it can act as an analog computer to solve these NP hard problems by relaxing to the ground states of the system. It has also been shown in [2] that the classical Kuramoto model with nearest neighbor interactions maps to the Ising Hamiltonian when adding a term that accounts for a 2$^\textrm{nd}$ harmonic injection locking signal (SHIL). This term leads to a binarization of the oscillators phases. Hence, systems of mutually coupled oscillators where each oscillator receives an additional SHIL signal may be candidates for physical implementations of Ising Machines. We discuss whether systems of mutually coupled electronic oscillators, so called phase-locked loops (PLLs), with an SHIL signal are candidates for such analog computing approaches. More specifically, we discuss the role of oscillator inertia and time-delayed coupling that become relevant when scaling such systems to larger network sizes. We present the results of numerical simulations of a 2$^\textrm{nd}$ order Kuramoto model with time-delayed couplings and SHIL using the maximum cut problem for benchmarking. [1] A. Lucas, "Ising formulations of many NP problems", Frontiers in Physics, Vol. 2 (2014) [2] T. Wang and J. Roychowdhury, "OIM: Oscillator-based Ising Machines for Solving Combinatorial Optimisation Problems", Unconventional Computation and Natural Computation: 18th International Conference, UCNC 2019, Tokyo, Japan (2019)

This talk presents a proof of concept for a multi-robot system that forms emergent space-time patterns. Inspired by the theory of swarmalators, in which synchronization and swarming of agents are mutually coupled, we propose a robot-suitable model for coordination in time and space. The approach is evaluated using small ground robots and drones. Video: https://www.youtube.com/watch?v=YCZREhgCF84

In a so-called overpopulated world, sustainable consumption is of existential importance. However, the expanding spectrum of product choices and their production complexity challenge consumers to make informed and value-sensitive decisions. Recent approaches based on (personalized) psychological manipulation are often intransparent, potentially privacy-invasive and inconsistent with (informational) self-determination. By contrast, responsible consumption based on informed choices currently requires reasoning to an extent that tends to overwhelm human cognitive capacity. As a result, a collective shift towards sustainable consumption remains a grand challenge. I present a novel personal shopping assistant implemented as a smart phone app that supports a value-sensitive design and leverages sustainability awareness, using experts’ knowledge and ‘wisdom of the crowd’ for transparent product information and explainable product ratings. Real-world field experiments in two supermarkets confirm higher sustainability awareness and a bottom-up behavioural shift towards more sustainable consumption. These results encourage novel business models for retailers and producers, ethically aligned with consumer preferences and with higher sustainability.

The ability to efficiently move in complex environments is a fundamental property both for animals and for robots, and the problem of locomotion and movement control is an area in which neuroscience, biomechanics, and robotics can fruitfully interact. In this talk, I will present how biorobots and numerical models can be used to explore the interplay of the four main components underlying animal locomotion, namely central pattern generators (CPGs), reflexes, descending modulation, and the musculoskeletal system. Going from lamprey to human locomotion, I will present a series of models that tend to show that the respective roles of these components have changed during evolution with a dominant role of CPGs in lamprey and salamander locomotion, and a more important role for sensory feedback and descending modulation in human locomotion. I will also present a recent project showing how robotics can provide scientific tools for paleontology. Interesting properties for robot and lower-limb exoskeleton locomotion control will finally be discussed.

We revisit the recently revealed regimes of “nonconventional synchronization” in systems of coupled bi-stable van der Pol oscillators. These regimes are characterized by periodic (or quasiperiodic) almost complete energy exchanges between the coupled oscillators. It is demonstrated that such responses correspond to synchronization of the modulation amplitudes between symmetric and antisymmetric modes of the system, with persistent phase drift. This observation substantially simplifies the treatment, reduces the dynamics to a simple phase cylinder and allows to reveal a rich set of global and local bifurcations of limit cycles and tori in the system. Among other findings, one encounters an unexpected regime of nonstationary asymmetric synchronization.

We explore passive nonlinear targeted energy transfer (TET) in dynamical and acoustical systems based on the synergy of intentional strong nonlinearity, asymmetry, and possible internal scale hierarchy. This is a process where, through predictable design, broadband or narrowband input energy is either irreversibly directed in preferential paths/modes, passively scattered in the frequency/wavenumber domains, dissipated locally, or harvested at a priori designated sites. Interestingly, TET mimics analogous energy cascades occurring often in Nature (e.g., in turbulent flows or granular media), and, as such, benefits from the well-known robust and enhanced dissipative features exhibited by these natural phenomena. Our approach dictates advanced theoretical modelling and analysis accounting for strongly nonlinear effects, but also nonlinear system identification and reduced-order modelling to characterize the experimental realizations that validate the theoretical predictions. We discuss several applications, including implementing intermodal TET, designing, analyzing, characterizing, and experimentally testing non-reciprocal lattice materials incorporating internal hierarchical scales, and employing TET for tuning the bandwidths of nonlinear oscillators. The aim is to translate these approaches to new methods, technologies and devices that exploit and showcase nonlinear TET. This work is funded in part by National Science Foundation (NSF) Emerging Frontiers Research Initiative Grant 1741565. Any expressed opinion, findings, conclusions, or recommendations are those of the presenter and do not necessarily reflect the views of NSF.

Walking animals can effectively perform self-organized and robust locomotion. They can quickly adapt their gaits to deal with damage and unpredictable changing environments. They can even take proactive steps to avoid colliding with an obstacle. Biological studies reveal that the complex locomotion behavior is largely attained by several ingredients including neural control mechanisms with plasticity and memory, sensory feedback and adaptation, body dynamics, and their dynamical interactions with the environment. In this talk, I will present “how we can realize these ingredients inspired by nature for complex machines (robots) so they can become more intelligent like their biological counterparts”.

Recent advancements in artificial intelligence and robotics could create a significant effect on the role of machines in our daily life, namely human-machine interaction. Within Industry 4.0, a paradigm shift in the future of robotics from creating robots doing repetitive and high-precision tasks to intelligent and helpful robots that complement humans’ capabilities, learn new tasks, and adapt to uncertain environments is foreseen. These intelligent robots, do not have just powerful brains to keep everything under control, but also intelligent bodies which could intrinsically, react to the environmental changes and external interactions. Having adaptable body properties (e.g., compliance) with bioinspired actuation and morphological design is one step towards developing machines with mechanical intelligence. Recently, we introduced the hybrid Electric-Pneumatic-Actuator (EPA) as a practical implementation of this idea. Combining control and body intelligence with the EPA system improves the agility, efficiency, and robustness of different gaits in robots. With EPA we can benefit from nonlinear system dynamics adaptation to 1) improve robot interaction with the environment with less sensitivity to sensory issues (e.g., noise, delay, and fault), 2) develop robots with higher efficiency and robustness against perturbations and uncertainties, 3) approach human locomotion behavior with the robots, and 4) finally to be used in assistive devices with more compatible design and control to humans.

Mobile microrobots, which can navigate, sense, and interact with their environment, could potentially revolutionize medicine and environmental remediation. Many self-organizing microrobotic collectives have been developed to overcome inherent limits in actuation, sensing, and manipulation of individual microrobots; however, reconfigurable self-organized collectives with robust transitions between behaviors are rare. Such systems that perform multiple functions are advantageous to operate in complex environments. Here, we present a versatile self-organizing robotic collective system at the microscale capable of on-demand reconfiguration to adapt to and utilize their environments to perform various functions at the air-water interface. Each microrobot is disc-shaped with a magnetic film coating to be oscillated and pulled using applied external magnetic field waveforms and gradients, respectively. Microrobots have magnetic, capillary and hydrodynamic interactions inducing various nonlinear coupled dynamics and dynamic self-assembly modes. Our system exhibits diverse modes ranging from isotropic to anisotropic behaviors and transitions between a globally driven and a novel self-propelling behavior. We show the transition between different modes in experiments and simulations, and demonstrate various functions, using the reconfigurability of our system to navigate, explore, and interact with the environment. Such versatile microrobot collectives with globally driven and self-propelled behaviors have significant potential in future medical and environmental applications.

Propiosensation is often essential for biological locomotion. Humans, f.i., cannot perform coordinated movements when deprived of the capability to sense limb postures via embedded muscle tension sensors. Alternatively, movements can be induced via top-down commands, f.i. by a central pattern generator (CPU). There exist hence a range of control principles that starts at one end with pure top-down control, viz without sensory feedback. In the other extreme, termed 'attractoring', locomotion is generated via local feedback loops, which become defunct without propriosensation. Attractoring implies that the states of body and environment are part of the controlling dynamical systems (via propriosensation), which implies that locomotion arises via self-organizing processes. For several simulated and real-world robots we show that attractoring leads to robust locomotion, with stable limit cycles in the sensory-motor loop corresponding to compliant gaits. Attractoring allows to reduce the complexity of to-down control. Short burst of commands can be used to switch between different gaits, viz to kick the systems from a given attracting limit cycle into the basin of attraction of another attractor.

Active materials and smart structures are characterized by coupling of the mechanical field with non-mechanical fields. In the talk, first, the fundamentals of various active materials are presented. Then, the modeling and numerical simulation of representative applications in the field of smart structures are given.

The reliable operation of power grids, the largest man-made infrastructure networks, underlies almost all aspects of our daily lives. Understanding how the numerous interconnected power generators and consumers in the grids collectively respond to external perturbations, e.g. from the fluctuating renewable power supply and the intermittent consumption behavior, is critical for ensuring the safe operation of power grids. Yet, many aspects of the collective response dynamics of power grids still remain unclear to date. Here we present the discovery of self-organized distributed patterns emerging in the collective response dynamics of fluctuation-driven power grids: first, we identify three regimes of spatio-temporal response patterns with distinctive characteristics depending on the timescale of the fluctuation signal, and second, we reveal a topological determinant that largely characterizes the pattern in the spreading of a sudden perturbation through power grid networks. These findings might help to control, optimize and design power grid systems with enhanced stability and robustness against external fluctuations.

Several independent research groups have investigated the self-adaptive behavior of a clamped-clamped beam with attached slider. Under harmonic base excitation the slider passively moves to a certain position on the beam, going along with a significant increase of amplitude. We are the first to fully explain this complex process theoretically. Our model accounts for nonsmooth contact interactions between beam and slider as well as the beam's hardening nonlinearity. This model agrees with experiments qualitatively and quanitatively, reproducing all identified kinds of operating behavior in an excellent manner. Moreover, we show that the slider’s movement takes places on a slower time scale compared to the beam’s vibration. Exploiting this temporal separability, we explain the different mechanisms yielding transport towards or away from beam’s center.

Eusocial insects (bees, ants, termites) fascinate through collective behaviour, strict self-organisation, and speciation. For many species a seemingly exotic ability to communicate via micro-vibrations has emerged. Here it is hypothesised that structured nests, often of amazingly complex topology, self-ventilation, and thermoregulation properties, may also function as vibro-acoustic communication networks. Examples of complex topology of termite, bee and ant nests will be examined for their features using e.g., (high-resolution) micro-computed tomography. The possibility of using these structures as communication channel for propagating miniscule waves as used in vibrational communication (biotremology) and effects of nonlinearity in signal and system is explored - to understand the nest as a living part of the colony as superorganism and how direct and indirect communication may be entangled in nature on different behavioural scales for increased resilience.

Electronic circuits and the resulting capabilities are the carriers of the current industrial revolution with its tremendous possibilities. What is the potential of integrated circuits that process multimodal information? On the basis of two examples of such circuits from microfluidics and robotics, we discuss their potential and challenges.

We discuss a general methodology to obtain data-driven reduced-order models from numerical and experimental data. These models are obtained as normal forms on attracting spectral submanifolds (SSMs) in the phase space of the underlying physical system. Our main focus is to model essentially nonlinear (or nonlinearizable) behavior that cannot be captured by linear reduced-order modelling techniques. We show several applications of SSM-based reduced-order model identification from data sets arising in solid and fluid mechanics.

Today, the engineering view on nonlinear dynamics is governed by linear and weakly nonlinear time series analysis techniques and modeling approaches grounded in analytical equations. These methods are mostly rooted in time and frequency domain, and are therefore inherently limited to resolving only specific characteristics of the dynamics. Many aspects of large mechanical systems and their dynamics remain challenging, like the multiplicity of components and degrees of freedom, friction, nonlinearities, emergent behavior and rare events, and multi-scale dynamics. This work proposes a novel way of approaching the challenges in mechanical dynamical systems analysis by complementing the classical view of a geometric structure with a functional network perspective. Different approaches to obtaining functional networks from time series data are presented to study the information on the underlying dynamics encoded in the network, such as invariant measures or bifurcation scenarios.

TBD

I will present dynamic neural fields as a computational framework to structure spiking neural network architectures to solve real-world robotic tasks using neuromorphic computing hardware.

Challenges to meet today’s computing needs include the slowing down of technology scaling and the exponential growth of data that need to be processed. A promising alternative approach consists in performing computing based on intrinsic device dynamics, such that each device functionally replaces expensive circuits, leading to adaptive ‘complex computing’. In this talk, I will present a few examples of how material properties can enable device dynamics, and how internal device dynamics can lead to interesting network dynamics and potential applications. These native dynamics at the device and material levels can enable new computing architectures, such as brain- inspired neuromorphic systems, which offer both high energy efficiency and high computing capacity.

Thermal ratchets are a paradigm of statistical physics. They taught us how to harness thermal energy from Brownian motion by keeping the system out of equilibrium. A collection of microswimmers such as bacteria also perform random motion. They can drive a polar tracer particle suspended in an active bath since they constantly push against its surface. I review the properties of the active bath including its colored noise, show the effective Langevin equations of the polar tracer, and also demonstrate the highly nonlinear microrheology of a tracer particle. Biological microswimmers also need to navigate in complex environments such as porous soil or landscapes of external cues. For artificially generated microswimmers it is a challenge to mimic navigation strategies from biology. I show how a smart active particle, which is able to sense its environment, can learn to optimally move in a potential landscape using a reinforcement-learning strategy.

Polymer gels, such as hydrogels, have been widely used in biomedical applications, flexible electronics, and soft machines. Polymer network design and its contribution to the performance of gels have been extensively studied. In this study, the critical influence of the solvent nature on the mechanical properties and performance of soft polymer gels is demonstrated. A polymer gel platform based on poly(ethylene glycol) (PEG) as solvent is reported (PEGgel). Compared to the corresponding hydrogel or ethylene glycol gel, the PEGgel with physically cross-linked poly(hydroxyethyl methacrylate-co-acrylic acid) demonstrates high stretchability and toughness, rapid self-healing, and long-term stability. Depending on the molecular weight and fraction of PEG, the tensile strength of the PEGgels varies from 0.22 to 41.3 MPa, fracture strain from 12% to 4336%, modulus from 0.08 to 352 MPa, and toughness from 2.89 to 56.23 MJ m–3. Finally, rapid self-healing of the PEGgel is demonstrated and a self-healing pneumatic actuator is fabricated by 3D printing. The enhanced mechanical properties of the PEGgel system may be extended to other polymer networks (both chemically and physically cross-linked). Such a simple 3D-printable, self-healing, and tough soft material holds promise for broad applications in wearable electronics, soft actuators, and robotics.

An interesting application of phase-locking of oscillators is the locomotion of multiple-legged robots and animals, for which it is required that the periodic dynamics of individual legs are self-organized on the level of the body into motion patterns with well-defined phase differences. Here, we propose a network of neural oscillators for generating hexapod-type motion patterns. Neurons are coupled both directly via synaptic weights and also indirectly by closing the sensorimotor loop via proprioception. The oscillatory behavior of individual neurons results from homeostatic adaptation. The resulting phase-locked patterns are investigated by numerical simulations within a physical simulation environment for rigid-body dynamics but also on real hexapod-type robots. We find that the hexapod is able to walk using different phase-locked patterns even for network structures with fully symmetric coupling. While within this model the presence of propriosensory feedback is found to widen the oscillatory parameter region for single neurons, on the level of the full system it may also result in body twisting without locomotion.

Legged robots pose one of the most significant challenges in robotics. The dynamic and agile maneuvers of animals cannot be imitated by existing methods that are crafted by humans. A compelling alternative is reinforcement learning, which requires minimal craftsmanship and promotes the natural evolution of a control policy. In this presentation, I will share our latest research efforts in learning locomotion at KAIST, which includes hardware design, simulation, control, and planning. I'll also introduce our new quadruped robot Raibo, which can run over 3 m/s on diverse challenging terrains thanks to learning-based methods.

Adhesive pads are functional systems which evolved as a response to the environment and therefore often exemplify convergent evolution. The functional constraints under which attachment evolved are general requirements. The strategies to solve similar attachment problems often follow similar physical principles, hence, the morphology of attachment devices is affected by physical constraints. This resulted in two main types of attachment devices, hairy and smooth ones, in animals. They differ in morphology and ultrastructure but achieve mechanical adaptation to substrates with different roughness and maximise the actual contact area with them. Species-specific environmental surface conditions are similar in the specific environments unrelated to the group of animals, therefore, micro- and nanostructural adaptations of the attachment systems of different animal groups reveal similar mechanisms. Consequently, similar attachment organs evolved in a convergent manner and different attachment solutions can occur within closely related lineages. This talk shows parallel evolution of attachment structures on micro- and nanoscales at different phylogenetic levels, focuses on insects as the largest animal group on earth, and subsequently zoom into the attachment pads of the stick and leaf insects (Phasmatodea) to explore convergent evolution of attachment pads at even smaller scales. As a primary influence on the forces that can be exerted to the substrate by the animals, attachment has a big influence on the walking dynamics of living animals, as it limits propulsion and defines how many legs can be lifted off the substrate at once. Since convergent events might be potentially interesting for engineers as a kind of optimal solution by nature, the functional morphology of attachment devices can be of concern for biomimetic design of walking machines and understanding walking dynamics.

The capability to physically interact with complex environments is a hallmark of animals and humans, such as insect navigation and human manipulation (of flexible tools, e.g., whips). How they handle neural, muscle, and tool complexities is an interesting question, where robotics and computational neuroscience can mutually benefit from. In this talk, I will present how robots and neuromechanical models that can be used to provide novel hypotheses for insect navigation, human arm swings, and human manipulation of whips. It shows a way from machines for (uncovering) biological intelligence. Another way from biological to machine intelligence will be presented in insect-like walking and wearable robots.