We present a hierarchical, control theory inspired method for variational inference (VI) for neural stochastic differential equations (SDEs).While VI for neural SDEs is a promising avenue for uncertaintyaware reasoning in time-series, it is computationally challenging due to the iterative nature of maximizing the ELBO.In this work, we propose to decompose the control term into linear and residual non-linear components and derive an optimal control term for linear SDEs, using stochastic optimal control.Modeling the non-linear component by a neural network, we show how to efficiently train neural SDEs without sacrificing their expressive power.Since the linear part of the control term is optimal and does not need to be learned, the training is initialized at a lower cost and we observe faster convergence.* MO acknowledges funding

Precise identification of dynamic models in robotics is essential to support dynamic simulations, control design, friction compensation, output torque estimation, etc. A longstanding challenge remains in the development and identification of friction models for robotic joints, given the numerous physical phenomena affecting the underlying friction dynamics which result into nonlinear characteristics and hysteresis behaviour in particular. These phenomena proof difficult to be modelled and captured accurately using physical analogies alone. This has motivated researchers to shift from physics-based to data-driven models. Currently, these methods are still limited in their ability to generalize effectively to typical industrial robot deployement, characterized by high- and low-velocity operations and frequent direction reversals. Empirical observations motivate the use of dynamic friction models but these remain particulary challenging to establish. To address the current limitations, we propose to account for unidentified dynamics in the robot joints using latent dynamic states. The friction model may then utilize both the dynamic robot state and additional information encoded in the latent state to evaluate the friction torque. We cast this stochastic and partially unsupervised identification problem as a standard probabilistic representation learning problem. In this work both the friction model and latent state dynamics are parametrized as neural networks and are integrated in the conventional lumped parameter dynamic robot model. The complete dynamics model is directly learned from the noisy encoder measurements in the robot joints. We use the Expectation-Maximisation (EM) algorithm to find a Maximum Likelihood Estimate (MLE) of the model parameters. The effectiveness of the proposed method is validated in terms of open-loop prediction accuracy in comparison with baseline methods, using the Kuka KR6 R700 as a test platform.

The increasing complexity of tasks performed by hybrid aerial robotic systems, such as tail-sitters, demands a more integrated approach to their design. Traditional sequential design methods fall short because they separate the control system design from the conceptual design, limiting the poten-tial for discovering coupled solutions. This disjointed process constrains the design space, making it difficult to optimize both the control performance and system dynamics simultaneously. In response to this limitation, there has been growing interest in mission-specific dynamic design procedures, which aim to address specific operational challenges by integrating control and design early in the development process. The multi-disciplinary approach of control co-design (CCD) expands the design space by solving control and system design problems concurrently. The recently introduced DAIMYO framework demonstrated that combining multi-fidelity modelling with a nested CCD approach can tackle the situ-to-reel gap. However, DAIMYO's reliance on Bayesian optimization to account for the computational cost increase of a nested formulation limits its scalability. To address these issues, we propose KUGE, a simultaneous CCD strategy that reduces computational complexity and overcomes dimensionality restrictions through a combined effort of stochastic optimization and Gaussian processes. We validate the effectiveness of KUGE by applying it to the dynamic design of a tail-sitter, showing that it is competitive with the DAIMYO architecture while offering greater computational efficiency.

In the context of web processing machines, the radial geometry of rolls or bobbins, the input material to the machine, must be measured with precision. The control system of the machine requires a measurement of the varying dimension of the bobbins as an input. A web processing machine (WPM) utilizes the rotational speed to regulate the tension of the web within the machine. Consequently, precise radius measurements, for optimal tension control, are of paramount importance. Traditional sensors, e.g., infrared and ultrasonic sensors, lack precision and are unable to detect anomalies. Additionally, their radius measurement is indirect, as these devices essentially function as distance sensors, and thus necessitate a calibration process. Furthermore, as these sensors only measure a single point on the bobbin, the system is vulnerable to localized irregularities and disturbances that could compromise the integrity of the material. This work proposes a novel methodology for measuring the radius of bobbins. A camera system is used in conjunction with Frequency Domain Zero Padding (FDZP) as a sub-pixel technique to ensure high precision. A correction is applied to compensate for the perspective-induced distortion resulting from the camera’s radial viewpoint with respect to the center of the roll. Moreover, the camera possesses the capability to evaluate multiple points on the bobbin, which enables the detection of the aforementioned anomalies at different positions on the bobbin. The system’s validation was conducted in an industrial setting, where it demonstrated a precision that is 6.2 times better than that of conventional sensors. The method ensures real-time performance while remaining low-cost.

Non-prehensile manipulation aims to change how robots handle objects, going beyond conventional grasping, including pushing, sliding, tipping, rolling, and throwing. These manipulation modes however face new challenges in sensing, planning and control. This motivates the development of advanced and dedicated algorithms. This research explores the use of Model Predictive Control to control the motion of objects through pushing and sliding on a practical pusher-slider system. We develop a reactive path planning feedback controller that can navigate dynamic objects in real-time. To that end, we introduce a simplified model with advantageous properties. Differential flatness allows the description of dynamically feasible trajectories in a representation not subject to differential constraints, simplifying motion optimization. A temporal invariance property of the state trajectories allows us to simplify the problem further into a strictly geometric optimization where velocities are imposed in a second stage. We show that the method is sufficiently computationally efficient to support reactive replanning when facing dynamic obstacles. Experimental validation with a KUKA manipulator and vision-based tracking demonstrates the applicability of this method.

We present a novel variational framework for performing inference in (neural) stochastic differential equations (SDEs) driven by Markov-approximate fractional Brownian motion (fBM). SDEs offer a versatile tool for modeling real-world continuous-time dynamic systems with inherent noise and randomness. Combining SDEs with the powerful inference capabilities of variational methods, enables the learning of representative distributions through stochastic gradient descent. However, conventional SDEs typically assume the underlying noise to follow a Brownian motion (BM), which hinders their ability to capture long-term dependencies. In contrast, fractional Brownian motion (fBM) extends BM to encompass non-Markovian dynamics, but existing methods for inferring fBM parameters are either computationally demanding or statistically inefficient. In this paper, building upon the Markov approximation of fBM, we derive the evidence lower bound essential for efficient variational inference of posterior path measures, drawing from the well-established field of stochastic analysis. Additionally, we provide a closed-form expression for optimal approximation coefficients and propose to use neural networks to learn the drift, diffusion and control terms within our variational posterior, leading to the variational training of neural-SDEs. In this framework, we also optimize the Hurst index, governing the nature of our fractional noise. Beyond validation on synthetic data, we contribute a novel architecture for variational latent video prediction,—an approach that, to the best of our knowledge, enables the first variational neural-SDE application to video perception.

Industrial robots are capable of moving at high speed. Each time they come into contact with their environment, e.g. to pick up an object, they decelerate to a near standstill. A solution involving a compliant pneumatic gripper and adapted trajectory plan is presented to initiate contact at a higher speed while remaining within hardware limits. By adding overload clutches in either the robot arm or gripper, tolerance to errors is provided. The key parameters such as gripper compliance and maximum allowed initial impact velocity are identified. Results show that by properly optimizing these parameters, robot picking of rigid objects can be accelerated. The complete high-speed picking solution is experimentally verified. A time reduction of 16% was obtained when making contact at 0.65 m/s.

We present KeyCLD, a framework to learn Lagrangian dynamics from images. Learned keypoints represent semantic landmarks in images and can directly represent state dynamics. We show that interpreting this state as Cartesian coordinates, coupled with explicit holonomic constraints, allows expressing the dynamics with a constrained Lagrangian. KeyCLD is trained unsupervised end-to-end on sequences of images. Our method explicitly models the mass matrix, potential energy and the input matrix, thus allowing energy based control. We demonstrate learning of Lagrangian dynamics from images on the cl_bnmsqnk pendulum, cartpole and acrobot environments. KeyCLD can be learned on these systems, whether they are unactuated, underactuated or fully actuated. Trained models are able to produce long-term video predictions, showing that the dynamics are accurately learned. We compare with Lag-VAE, Lag-caVAE and HGN, and investigate the benefit of the Lagrangian prior and the constraint function. KeyCLD achieves the highest valid prediction time on all benchmarks. Additionally, a very straightforward energy shaping controller is successfully applied on the fully actuated systems.

There is an emerging industrial demand for predictive maintenance algorithms that exhibit high levels of predictive accuracy. Such condition monitoring tools must estimate dynamic quantities, such as Remaining Useful Lifetime (RUL) and the State of Health (SOH), based on a, typically, restricted set of measurements that can be obtained in an operational setting. These quantities exhibit inherent stochasticity and can only be approximately determined a posteriori to system failure. This paper proposes a generic prognostic tool for probabilistic condition monitoring of mechatronic systems, with the aim to improve the probabilistic prediction of condition metrics, specifically RUL and SOH. Therefore we propose to identify a Hidden Markov Model (HMM) from a fully instrumented measurement set, that is only available for a restricted set of run-to-failure experiments, typically gathered in an R&D setting. Although being artificial and retrospectively constructed metrics, we interpret RUL and SOH as physical measurements with the purpose to identify accurate degradation dynamics. Once the degradation model is identified, we practice the mathematical flexibility of the HMM framework to estimate several of the no longer available dynamic quantities of interest in real-time, from the limited set of measurements that are available in an operational setting. This modelling paradigm is known as virtual sensing. Predictive performance and computational efficiency are further improved by domain knowledge based pre-processing of the measurements. We apply our methodology to solenoid valves (SV), a widely used and often critical component in many industrial systems, which display a large variation in useful lifetime. Benchmark results show that the predictive capabilities of the presented methodology compares with prognostic techniques that are more computationally and memory demanding.Note to Practitioners—The motivation for this research is twofold. First there is a pending industrial need for improved diagnostic and prognostic tools. Second there is the observation that lifetime tests usually take place in an R&D setting and that expert labelling of Remaining Useful Lifetime (RUL) or State of Health (SOH) of a component or system is often based on measurement data that is not available in the industrial setting where the prognostic tools are to be deployed in the end. These two observations suggest that there is large potential in methods that can correlate the expert labelling, in particular RUL & SOH signals, with measurement data that is available in the industrial setting. Our approach has been tested in detail on the case of Solenoid Valves, which are widely used in industry and that are often safety critical. Our experiments demonstrate that the method compares with brute force approaches that overpower ours both in terms of computational as well as memory requirements. The method is furthermore generic and there is no reason to assume it would not work for other applications.

Robots are more often deployed in unstructured or unpredictable environments. Particularly collisions at high speed can severely damage the drivetrains and joint bearings of robots. In order to avoid such collisions, path planners exist that adapt the robot's original trajectory online if a collision hazard is detected. These methods require additional sensors such as cameras, are computationally costly and never flawless due to occlusions. Another approach is to incorporate a cost function that promotes collision tolerance while planning the initial trajectory. The resulting impact-aware path plan minimizes the chance of robot hardware damage if a collision would occur. Two algorithms are presented to assess collision tolerance in high-speed robots, taking into account factors such as robot pose, impact direction, and maximum intermittent loading of the gearboxes and bearings. The first algorithm is more general while the second assumes the presence of joint overload clutches that decouple upon impact. These algorithms are applied to plan an impact-aware path for a custom 6-axis series clutched actuated robot that serves as use case. Both for the case with and without clutches, a generic impact-aware plan is presented as well as at least one derived, heuristic alternative. Without clutches, trajectories that are perpendicular to the end effector flange were found to be desirable, as they allow the robot to mitigate the highest collision force without overloading the gearboxes or bearings. On the other hand, with clutches, trajectories that are parallel to the end effector flange were found to be more collision tolerant. The effect of impact direction was also experimentally validated using the custom 6-axis robot. Collisions at velocities up to 1.2 m/s were mitigated through the combination of impact-aware path planning and series clutched actuation.

Adhesive bonding is a joining process used in several industries such as aerospace, automotive, civil construction and manufacturing. Traditionally, the optimization of the parameters for this process is performed by adhesive experts via trial and error which is expensive and time-consuming. Therefore having a process model for optimization purposes is of great interest. In this study, we develop such process model which includes cost, visual quality and joint strength properties for Polyphenylene sulfide bonding use-case. We adopt analytical modeling approaches for those process properties that do not require extensive system knowledge and are not effected by large number of process parameters, namely cost and visual quality. Additionally, we use data-driven genetic programming approach to model the more nonlinear process property, meaning joint strength of the bond. Consequently, we employ a hybrid approach by combining available knowledge and experimental data. The process model can then be implemented for process optimization or to create a digital twin which predicts if the product quality is in scope.

This paper treats a special case of the Imitation from Observations (IfO) problem. IfO is a generalisation of Imitation Learning from state-only demonstrations. Our treatment of IfO considers the case of feature-only demonstrations. This means that the full state is inaccessible for inference, and imitation must occur on the basis of a limited set of features. We refer to this setting as Imitation from Partial Observations (IfPO). This scenario has the advantage of allowing to address a wider variety of demonstrations, as well as solving the problem of heteromorphic student and teacher. We set out for policy learning methods that extract an executable state-feedback policy, directly from those features, which in the literature is known as Behavioural Cloning. In this theoretical work, we formalize the rational inference model of the student decision maker, devoted to imitation, as a controlled Hidden Markov Model. The IfPO problem is then reformulated as a Maximum Likelihood Estimation problem and treated using Expectation-Maximization. We name the resulting fixed point iterations A Posteriori Control Densities. We compare the presented approach to existing methods in the field and identify potential directions for further development, such as an extension to unknown transition and emission models.

In energy-intensive industrial systems, an evaporative cooling process may introduce operational flexibility. Such flexibility refers to a system’s ability to deviate from its scheduled energy consumption. Identifying the flexibility, and therefore, designing control that ensures efficient and reliable operation presents a great challenge due to the inherently complex dynamics of industrial systems. Recently, machine learning (ML) models have attracted attention for identifying flexibility, due to their ability to model complex nonlinear behavior. This article presents ML-based methods that integrate system dynamics into the ML models (e.g., neural networks) for better adherence to physical constraints. We define and evaluate physics-informed long-short term memory networks (PhyLSTMs) and physics-informed neural networks (PhyNN) for the identification of flexibility in the evaporative cooling process. These physics-informed networks approximate the time-dependent relationship between control input and system response while enforcing the dynamics of the process in the neural network architecture. Our proposed PhyLSTM provides less than 2% system response estimation error, converges in less than half iterations compared to a baseline NN, and accurately estimates the defined flexibility metrics. We include a detailed analysis of the impact of training data size on the performance and optimization of our proposed models.

Solenoid valves (SV) are essential components of industrial systems and therefore widely used. As they suffer from high failure rates in the field, fault prognosis of these assets plays a major role for improving their maintenance and reliability. In this work, Bayesian convolutional neural networks are used to predict the remaining useful life (RUL) of SV, by training them on the valve's current signatures. Predictive performance is further improved upon by using salient physical features obtained from an electromechanical model as the network's training input. Results show that our designed network architecture produces well-calibrated uncertainty estimations of the RUL predictive distributions, which is an important concern in prognostic decision-making.

This paper deals with the problem of string stability in a chain of acceleration-controlled vehicles. There exist two different definitions of string stability in literature. It is known that those versions of string stability are impossible to achieve with any linear homogenous controllers if the vehicles only use relative information of few vehicles in front. Previous works have shown that adding absolute velocity into the controller, allows to satisfy the weaker definition of string stability. In this paper, we consider stronger definition of string stability, under using heterogenous controllers in platoon. We prove it is possible to guarantee the most strict definition of string stability using linear heterogenous unidirectional controllers with PD coupling gain between the vehicles.

We propose a series of methods to generate inverse stochastic trajectories given parametric model uncertainties. The methods are applied for the generation of robust quad-copter trajectories. Our aim is to generate unique reference trajectories such that the corresponding uncertain control policies are less sensitive to parameter variations. Such robust trajectories exhibit interesting properties w.r.t. to tracking. The proposed methods can be applied on any system represented by nonlinear dynamics given that an inverse dynamic model is available, be it in a flat interpretation as is the case for quad-copters. Results show a decrease of 78% of parameter induced policy uncertainty when compared with nominal trajectories.

The demand for ever increasing efficiency keeps challenging the design and control of electrical machines. Thermal monitoring is in that perspective an important addition to the electromagnetic aspect. Temperature measurements in electrical machines can be challenging, especially in high-frequency inverter-fed motors. High dv/dt ratios in the stator windings give rise to noise exhibiting high amplitude and frequency. In this paper a multi-channel temperature monitoring system is proposed, implemented and experimentally tested for Resistance Temperature Detectors embedded at various locations in a 5.5 kW inverter-fed induction motor. The system ensures a galvanic isolation between the sensors in the motor and the data-acquisition system. After calibration, the linearity error, common-mode rejection ratio and signal-to-noise ratio are measured to be 0.12 % of full scale, -79.2 dB and 0.125 mV/V respectively. Fiber-Bragg Gratings thermal measurements are performed to confirm the accuracy of the proposed temperature monitoring system. For various operating conditions the temperature measurements have noise amplitudes that remain limited to 0.1 ° C-0.2 °C. The presented temperature monitoring system has the potential to further enhance motor performance when integrated into real-time motor controllers.

Resonant wireless power transfer has been developed and employed for transferring power to electrical loads. In recent research, the induced currents were directly used to exert forces or torques on a movable resonator coil. A high variability is present in the torque profile of a resonance based motoring system. In wireless power transfer, multiple compensation methodologies exist to counteract the reflected impedance of the load and to optimize the power flow, efficiency and/or VA rating of the source. This paper investigates the effect of the capacitor tuning for the two most common compensation methods in wireless power transfer, namely series LC and parallel LCL compensation. As a result of the highly variable reflected impedance, the peak torque does not always coincide with the zero phase angle of the total load impedance or the transmitter current peak. The torque generation capability, namely the ratio of average torque to maximum current, of both methods is largely similar. The LCL method has a higher peak value, which does however not coincide with its efficiency peak, so a trade-off is required. The efficiency of the LCL topology is shown to be significantly larger than the series LC transmitter.