Sebastian Trimpe

Behnam Khojasteh, Friedrich Solowjow, Sebastian Trimpe et al.

Machine learning and deep learning have been used extensively to classify physical surfaces through images and time-series contact data. However, these methods rely on human expertise and entail the time-consuming processes of data and parameter tuning. To overcome these challenges, we propose an easily implemented framework that can directly handle heterogeneous data sources for classification tasks. Our data-versus-data approach automatically quantifies distinctive differences in distributions in a high-dimensional space via kernel two-sample testing between two sets extracted from multimodal data (e.g., images, sounds, haptic signals). We demonstrate the effectiveness of our technique by benchmarking against expertly engineered classifiers for visual-audio-haptic surface recognition due to the industrial relevance, difficulty, and competitive baselines of this application; ablation studies confirm the utility of key components of our pipeline. As shown in our open-source code, we achieve 97.2% accuracy on a standard multi-user dataset with 108 surface classes, outperforming the state-of-the-art machine-learning algorithm by 6% on a more difficult version of the task. The fact that our classifier obtains this performance with minimal data processing in the standard algorithm setting reinforces the powerful nature of kernel methods for learning to recognize complex patterns.

Dominik Baumann, Friedrich Solowjow, Karl Henrik Johansson et al.

In networked control systems, communication is a shared and therefore scarce resource. Event-triggered control (ETC) can achieve high performance control with a significantly reduced amount of samples compared to classical, periodic control schemes. However, ETC methods usually rely on the availability of an accurate dynamics model, which is oftentimes not readily available. In this paper, we propose a novel event-triggered pulse control strategy that learns dynamics models if necessary. In addition to adapting to changing dynamics, the method also represents a suitable replacement for the integral part typically used in periodic control.

Henrik Hose, Johannes Köhler, Melanie N. Zeilinger et al.

Model predictive control (MPC) achieves stability and constraint satisfaction for general nonlinear systems, but requires computationally expensive online optimization. This paper studies approximations of such MPC controllers via neural networks (NNs) to achieve fast online evaluation. We propose safety augmentation that yields deterministic guarantees for convergence and constraint satisfaction despite approximation inaccuracies. We approximate the entire input sequence of the MPC with NNs, which allows us to verify online if it is a feasible solution to the MPC problem. We replace the NN solution by a safe candidate based on standard MPC techniques whenever it is infeasible or has worse cost. Our method requires a single evaluation of the NN and forward integration of the input sequence online, which is fast to compute on resource-constrained systems. The proposed control framework is illustrated using two numerical non-linear MPC benchmarks of different complexity, demonstrating computational speedups that are orders of magnitude higher than online optimization. In the examples, we achieve deterministic safety through the safety-augmented NNs, where a naive NN implementation fails.

Julia Berger, Bernd Frauenknecht, Sebastian Trimpe et al.

Model-Based Reinforcement Learning distinguishes between physical dynamics models operating on proprioceptive inputs and latent dynamics models operating on high-dimensional image observations. A prominent latent approach is the Recurrent State Space Model used in the Dreamer family. While epistemic uncertainty quantification to inform exploration and mitigate model exploitation is well established for physical dynamics models, its transfer to latent dynamics models has received limited scrutiny. We empirically demonstrate that latent transitions are biased toward well-represented regions of latent space, exhibiting an attractor behavior that can deviate from true environment dynamics. As a result, discrepancies in environment dynamics may not manifest in latent space, undermining the reliability of epistemic uncertainty estimates. Because these attractors often lie in high-reward regions, latent rollouts systematically overestimate predicted rewards. Our findings highlight key limitations of epistemic uncertainty estimation in latent dynamics models and motivate more critical evaluation of this method.

Shamil Mamedov, A. René Geist, Jan Swevers et al.

Accurately predicting deformable linear object (DLO) dynamics is challenging, especially when the task requires a model that is both human-interpretable and computationally efficient. In this work, we draw inspiration from the pseudo-rigid body method (PRB) and model a DLO as a serial chain of rigid bodies whose internal state is unrolled through time by a dynamics network. This dynamics network is trained jointly with a physics-informed encoder that maps observed motion variables to the DLO's hidden state. To encourage the state to acquire a physically meaningful representation, we leverage the forward kinematics of the PRB model as a decoder. We demonstrate in robot experiments that the proposed DLO dynamics model provides physically interpretable predictions from partial observations while being on par with black-box models regarding prediction accuracy. The project code is available at: http://tinyurl.com/prb-networks

Christian Fiedler, Michael Herty, Michael Rom et al.

Kernel methods, being supported by a well-developed theory and coming with efficient algorithms, are among the most popular and successful machine learning techniques. From a mathematical point of view, these methods rest on the concept of kernels and function spaces generated by kernels, so called reproducing kernel Hilbert spaces. Motivated by recent developments of learning approaches in the context of interacting particle systems, we investigate kernel methods acting on data with many measurement variables. We show the rigorous mean field limit of kernels and provide a detailed analysis of the limiting reproducing kernel Hilbert space. Furthermore, several examples of kernels, that allow a rigorous mean field limit, are presented.

Paul Brunzema, Alexander von Rohr, Sebastian Trimpe

Changing conditions or environments can cause system dynamics to vary over time. To ensure optimal control performance, controllers should adapt to these changes. When the underlying cause and time of change is unknown, we need to rely on online data for this adaptation. In this paper, we will use time-varying Bayesian optimization (TVBO) to tune controllers online in changing environments using appropriate prior knowledge on the control objective and its changes. Two properties are characteristic of many online controller tuning problems: First, they exhibit incremental and lasting changes in the objective due to changes to the system dynamics, e.g., through wear and tear. Second, the optimization problem is convex in the tuning parameters. Current TVBO methods do not explicitly account for these properties, resulting in poor tuning performance and many unstable controllers through over-exploration of the parameter space. We propose a novel TVBO forgetting strategy using Uncertainty-Injection (UI), which incorporates the assumption of incremental and lasting changes. The control objective is modeled as a spatio-temporal Gaussian process (GP) with UI through a Wiener process in the temporal domain. Further, we explicitly model the convexity assumptions in the spatial dimension through GP models with linear inequality constraints. In numerical experiments, we show that our model outperforms the state-of-the-art method in TVBO, exhibiting reduced regret and fewer unstable parameter configurations.

Andres Felipe Posada-Moreno, Nikita Surya, Sebastian Trimpe

Convolutional neural networks (CNNs) are increasingly being used in critical systems, where robustness and alignment are crucial. In this context, the field of explainable artificial intelligence has proposed the generation of high-level explanations of the prediction process of CNNs through concept extraction. While these methods can detect whether or not a concept is present in an image, they are unable to determine its location. What is more, a fair comparison of such approaches is difficult due to a lack of proper validation procedures. To address these issues, we propose a novel method for automatic concept extraction and localization based on representations obtained through pixel-wise aggregations of CNN activation maps. Further, we introduce a process for the validation of concept-extraction techniques based on synthetic datasets with pixel-wise annotations of their main components, reducing the need for human intervention. Extensive experimentation on both synthetic and real-world datasets demonstrates that our method outperforms state-of-the-art alternatives.

Mona Buisson-Fenet, Valery Morgenthaler, Sebastian Trimpe et al.

Identifying dynamical systems from experimental data is a notably difficult task. Prior knowledge generally helps, but the extent of this knowledge varies with the application, and customized models are often needed. Neural ordinary differential equations can be written as a flexible framework for system identification and can incorporate a broad spectrum of physical insight, giving physical interpretability to the resulting latent space. In the case of partial observations, however, the data points cannot directly be mapped to the latent state of the ODE. Hence, we propose to design recognition models, in particular inspired by nonlinear observer theory, to link the partial observations to the latent state. We demonstrate the performance of the proposed approach on numerical simulations and on an experimental dataset from a robotic exoskeleton.

Paul Brunzema, Alexander von Rohr, Friedrich Solowjow et al.

We consider the problem of sequentially optimizing a time-varying objective function using time-varying Bayesian optimization (TVBO). Current approaches to TVBO require prior knowledge of a constant rate of change to cope with stale data arising from time variations. However, in practice, the rate of change is usually unknown. We propose an event-triggered algorithm, ET-GP-UCB, that treats the optimization problem as static until it detects changes in the objective function and then resets the dataset. This allows the algorithm to adapt online to realized temporal changes without the need for exact prior knowledge. The event trigger is based on probabilistic uniform error bounds used in Gaussian process regression. We derive regret bounds for adaptive resets without exact prior knowledge of the temporal changes and show in numerical experiments that ET-GP-UCB outperforms competing GP-UCB algorithms on both synthetic and real-world data. The results demonstrate that ET-GP-UCB is readily applicable without extensive hyperparameter tuning.

Pierre-François Massiani, Mona Buisson-Fenet, Friedrich Solowjow et al.

Distinguishability and, by extension, observability are key properties of dynamical systems. Establishing these properties is challenging, especially when no analytical model is available and they are to be inferred directly from measurement data. The presence of noise further complicates this analysis, as standard notions of distinguishability are tailored to deterministic systems. We build on distributional distinguishability, which extends the deterministic notion by comparing distributions of outputs of stochastic systems. We first show that both concepts are equivalent for a class of systems that includes linear systems. We then present a method to assess and quantify distributional distinguishability from output data. Specifically, our quantification measures how much data is required to tell apart two initial states, inducing a continuous spectrum of distinguishability. We propose a statistical test to determine a threshold above which two states can be considered distinguishable with high confidence. We illustrate these tools by computing distinguishability maps over the state space in simulation, then leverage the test to compare sensor configurations on hardware.

Katharina Ensinger, Nicholas Tagliapietra, Sebastian Ziesche et al.

Physical systems can often be described via a continuous-time dynamical system. In practice, the true system is often unknown and has to be learned from measurement data. Since data is typically collected in discrete time, e.g. by sensors, most methods in Gaussian process (GP) dynamics model learning are trained on one-step ahead predictions. This can become problematic in several scenarios, e.g. if measurements are provided at irregularly-sampled time steps or physical system properties have to be conserved. Thus, we aim for a GP model of the true continuous-time dynamics. Higher-order numerical integrators provide the necessary tools to address this problem by discretizing the dynamics function with arbitrary accuracy. Many higher-order integrators require dynamics evaluations at intermediate time steps making exact GP inference intractable. In previous work, this problem is often tackled by approximating the GP posterior with variational inference. However, exact GP inference is preferable in many scenarios, e.g. due to its mathematical guarantees. In order to make direct inference tractable, we propose to leverage multistep and Taylor integrators. We demonstrate how to derive flexible inference schemes for these types of integrators. Further, we derive tailored sampling schemes that allow to draw consistent dynamics functions from the learned posterior. This is crucial to sample consistent predictions from the dynamics model. We demonstrate empirically and theoretically that our approach yields an accurate representation of the continuous-time system.

Andrés Felipe Posada-Moreno, Lukas Kreisköther, Tassilo Glander et al.

Convolutional Neural Networks (CNN) have become a common choice for industrial quality control, as well as other critical applications in the Industry 4.0. When these CNNs behave in ways unexpected to human users or developers, severe consequences can arise, such as economic losses or an increased risk to human life. Concept extraction techniques can be applied to increase the reliability and transparency of CNNs through generating global explanations for trained neural network models. The decisive features of image datasets in quality control often depend on the feature's scale; for example, the size of a hole or an edge. However, existing concept extraction methods do not correctly represent scale, which leads to problems interpreting these models as we show herein. To address this issue, we introduce the Scale-Preserving Automatic Concept Extraction (SPACE) algorithm, as a state-of-the-art alternative concept extraction technique for CNNs, focused on industrial applications. SPACE is specifically designed to overcome the aforementioned problems by avoiding scale changes throughout the concept extraction process. SPACE proposes an approach based on square slices of input images, which are selected and then tiled before being clustered into concepts. Our method provides explanations of the models' decision-making process in the form of human-understandable concepts. We evaluate SPACE on three image classification datasets in the context of industrial quality control. Through experimental results, we illustrate how SPACE outperforms other methods and provides actionable insights on the decision mechanisms of CNNs. Finally, code for the implementation of SPACE is provided.

Katharina Ensinger, Sebastian Ziesche, Barbara Rakitsch et al.

Modeling an unknown dynamical system is crucial in order to predict the future behavior of the system. A standard approach is training recurrent models on measurement data. While these models typically provide exact short-term predictions, accumulating errors yield deteriorated long-term behavior. In contrast, models with reliable long-term predictions can often be obtained, either by training a robust but less detailed model, or by leveraging physics-based simulations. In both cases, inaccuracies in the models yield a lack of short-time details. Thus, different models with contrastive properties on different time horizons are available. This observation immediately raises the question: Can we obtain predictions that combine the best of both worlds? Inspired by sensor fusion tasks, we interpret the problem in the frequency domain and leverage classical methods from signal processing, in particular complementary filters. This filtering technique combines two signals by applying a high-pass filter to one signal, and low-pass filtering the other. Essentially, the high-pass filter extracts high-frequencies, whereas the low-pass filter extracts low frequencies. Applying this concept to dynamics model learning enables the construction of models that yield accurate long- and short-term predictions. Here, we propose two methods, one being purely learning-based and the other one being a hybrid model that requires an additional physics-based simulator.

Bernd Frauenknecht, Lukas Kesper, Daniel Mayfrank et al.

Predictive safety filters (PSFs) leverage model predictive control to enforce constraint satisfaction during deep reinforcement learning (RL) exploration, yet their reliance on first-principles models or Gaussian processes limits scalability and broader applicability. Meanwhile, model-based RL (MBRL) methods routinely employ probabilistic ensemble (PE) neural networks to capture complex, high-dimensional dynamics from data with minimal prior knowledge. However, existing attempts to integrate PEs into PSFs lack rigorous uncertainty quantification. We introduce the Uncertainty-Aware Predictive Safety Filter (UPSi), a PSF that provides rigorous safety predictions using PE dynamics models by formulating future outcomes as reachable sets. UPSi introduces an explicit certainty constraint that prevents model exploitation and integrates seamlessly into common MBRL frameworks. We evaluate UPSi within Dyna-style MBRL on standard safe RL benchmarks and report substantial improvements in exploration safety over prior neural network PSFs while maintaining performance on par with standard MBRL. UPSi bridges the gap between the scalability and generality of modern MBRL and the safety guarantees of predictive safety filters.

Artur Eisele, Bernd Frauenknecht, Friedrich Solowjow et al.

Safety remains an open problem in reinforcement learning (RL), especially during training. While safety filters are promising to address safe exploration, they are generally poorly suited for high-dimensional systems with unknown dynamics. We propose Dyna-style Safety Augmented Reinforcement Learning (Dyna-SAuR), a novel algorithm that learns both a scalable safety filter and a control policy using a learned uncertainty-aware dynamics model, while requiring minimal domain knowledge. The filter avoids failures and high uncertainty regions. Thus, better models expand the set of safe and certain states, reducing filter conservatism. We present the effectiveness of Dyna-SAuR on goal-reaching CartPole as well as MuJoCo Walker, reducing failures compared to state-of-the-art methods by 2 orders of magnitude.

Alexander Gräfe, Ding Huo, Johannes Berger et al.

Transformer models are rapidly becoming a cornerstone of modern Internet of Things (IoT) applications, yet their computational and memory demands far exceed the capabilities of a single typical ultra-low-power IoT device. We present CATS, a framework for distributed transformer inference on ultra-low-power wireless devices, enabling multiple devices to collaboratively execute models far larger than what a single device can sustain. At its core, CATS is a communication-aware distributed transformer inference scheme co-designed across transformer partitioning, wireless communication and training. It employs SomeGather, a new pruned communication primitive that selectively broadcasts activation columns to reduce communication bandwidth and RAM usage without sacrificing model accuracy. Building on SomeGather, we design a partitioning method that exploits this primitive for efficient model parallelism. To cope with unreliable wireless communication, CATS employs message-dropout during training, which mimics packet losses and yields models that are robust to message loss during inference. In real-world experiments, we show that CATS brings distributed transformer inference to ultra-low-power wireless devices for the first time, with deployments on up to 16 devices that collaboratively execute transformer models up to 14 times larger than what a single device can run.

Katharina Ensinger, Sebastian Ziesche, Sebastian Trimpe

Dynamics model learning deals with the task of inferring unknown dynamics from measurement data and predicting the future behavior of the system. A typical approach to address this problem is to train recurrent models. However, predictions with these models are often not physically meaningful. Further, they suffer from deteriorated behavior over time due to accumulating errors. Often, simulators building on first principles are available being physically meaningful by design. However, modeling simplifications typically cause inaccuracies in these models. Consequently, hybrid modeling is an emerging trend that aims to combine the best of both worlds. In this paper, we propose a new approach to hybrid modeling, where we inform the latent states of a learned model via a black-box simulator. This allows to control the predictions via the simulator preventing them from accumulating errors. This is especially challenging since, in contrast to previous approaches, access to the simulator's latent states is not available. We tackle the task by leveraging observers, a well-known concept from control theory, inferring unknown latent states from observations and dynamics over time. In our learning-based setting, we jointly learn the dynamics and an observer that infers the latent states via the simulator. Thus, the simulator constantly corrects the latent states, compensating for modeling mismatch caused by learning. To maintain flexibility, we train an RNN-based residuum for the latent states that cannot be informed by the simulator.

Alexander von Rohr, Friedrich Solowjow, Sebastian Trimpe

Robust controllers ensure stability in feedback loops designed under uncertainty but at the cost of performance. Model uncertainty in time-invariant systems can be reduced by recently proposed learning-based methods, which improve the performance of robust controllers using data. However, in practice, many systems also exhibit uncertainty in the form of changes over time, e.g., due to weight shifts or wear and tear, leading to decreased performance or instability of the learning-based controller. We propose an event-triggered learning algorithm that decides when to learn in the face of uncertainty in the LQR problem with rare or slow changes. Our key idea is to switch between robust and learned controllers. For learning, we first approximate the optimal length of the learning phase via Monte-Carlo estimations using a probabilistic model. We then design a statistical test for uncertain systems based on the moment-generating function of the LQR cost. The test detects changes in the system under control and triggers re-learning when control performance deteriorates due to system changes. We demonstrate improved performance over a robust controller baseline in a numerical example.

Sebastian Trimpe

An event-based state estimation approach for reducing communication in a networked control system is proposed. Multiple distributed sensor-actuator-agents observe a dynamic process and sporadically exchange their measurements and inputs over a bus network. Based on these data, each agent estimates the full state of the dynamic system, which may exhibit arbitrary inter-agent couplings. Local event-based protocols ensure that data is transmitted only when necessary to meet a desired estimation accuracy. This event-based scheme is shown to mimic a centralized Luenberger observer design up to guaranteed bounds, and stability is proven in the sense of bounded estimation errors for bounded disturbances. The stability result extends to the distributed control system that results when the local state estimates are used for distributed feedback control. Simulation results highlight the benefit of the event-based approach over classical periodic ones in reducing communication requirements.

Guner Dilsad Er, Sebastian Trimpe, Michael Muehlebach

Algorithms increasingly operate within complex physical, social, and engineering systems where they are exposed to disturbances, noise, and interconnections with other dynamical systems. This article extends known convergence guarantees of an algorithm operating in isolation (i.e., without disturbances) and systematically derives stability bounds and convergence rates in the presence of such disturbances. By leveraging converse Lyapunov theorems, we derive key inequalities that quantify the impact of disturbances. We further demonstrate how our result can be utilized to assess the effects of disturbances on algorithmic performance in a wide variety of applications, including communication constraints in distributed learning, sensitivity in machine learning generalization, and intentional noise injection for privacy. This underpins the role of our result as a unifying tool for algorithm analysis in the presence of noise, disturbances, and interconnections with other dynamical systems.

Leon Greiser, Ozan Demir, Benjamin Hartmann et al.

Complicated first principles modelling and controller synthesis can be prohibitively slow and expensive for high-mix, low-volume products such as hydraulic excavators. Instead, in a data-driven approach, recorded trajectories from the real system can be used to train local model networks (LMNs), for which feedforward controllers are derived via feedback linearization. However, previous works required LMNs without zero dynamics for feedback linearization, which restricts the model structure and thus modelling capacity of LMNs. In this paper, we overcome this restriction by providing a criterion for when feedback linearization of LMNs with zero dynamics yields a valid controller. As a criterion we propose the bounded-input bounded-output stability of the resulting controller. In two additional contributions, we extend this approach to consider measured disturbance signals and multiple inputs and outputs. We illustrate the effectiveness of our contributions in a hydraulic excavator control application with hardware experiments. To this end, we train LMNs from recorded, noisy data and derive feedforward controllers used as part of a leveling assistance system on the excavator. In our experiments, incorporating disturbance signals and multiple inputs and outputs enhances tracking performance of the learned controller. A video of our experiments is available at https://youtu.be/lrrWBx2ASaE.

Christian Fiedler, Michael Herty, Sebastian Trimpe

In many applications of machine learning, a large number of variables are considered. Motivated by machine learning of interacting particle systems, we consider the situation when the number of input variables goes to infinity. First, we continue the recent investigation of the mean field limit of kernels and their reproducing kernel Hilbert spaces, completing the existing theory. Next, we provide results relevant for approximation with such kernels in the mean field limit, including a representer theorem. Finally, we use these kernels in the context of statistical learning in the mean field limit, focusing on Support Vector Machines. In particular, we show mean field convergence of empirical and infinite-sample solutions as well as the convergence of the corresponding risks. On the one hand, our results establish rigorous mean field limits in the context of kernel methods, providing new theoretical tools and insights for large-scale problems. On the other hand, our setting corresponds to a new form of limit of learning problems, which seems to have not been investigated yet in the statistical learning theory literature.

Bernd Frauenknecht, Tobias Ehlgen, Sebastian Trimpe

Advanced vehicle control is a fundamental building block in the development of autonomous driving systems. Reinforcement learning (RL) promises to achieve control performance superior to classical approaches while keeping computational demands low during deployment. However, standard RL approaches like soft-actor critic (SAC) require extensive amounts of training data to be collected and are thus impractical for real-world application. To address this issue, we apply recently developed data-efficient deep RL methods to vehicle trajectory control. Our investigation focuses on three methods, so far unexplored for vehicle control: randomized ensemble double Q-learning (REDQ), probabilistic ensembles with trajectory sampling and model predictive path integral optimizer (PETS-MPPI), and model-based policy optimization (MBPO). We find that in the case of trajectory control, the standard model-based RL formulation used in approaches like PETS-MPPI and MBPO is not suitable. We, therefore, propose a new formulation that splits dynamics prediction and vehicle localization. Our benchmark study on the CARLA simulator reveals that the three identified data-efficient deep RL approaches learn control strategies on a par with or better than SAC, yet reduce the required number of environment interactions by more than one order of magnitude.

Alexander Gräfe, Sebastian Trimpe

Mixed-precision training has emerged as an indispensable tool for enhancing the efficiency of neural network training in recent years. Concurrently, JAX has grown in popularity as a versatile machine learning toolbox. However, it currently lacks robust support for mixed-precision training. We propose MPX, a mixed-precision training toolbox for JAX that simplifies and accelerates the training of large-scale neural networks while preserving model accuracy. MPX seamlessly integrates with popular toolboxes such as Equinox and Flax, allowing users to convert full-precision pipelines to mixed-precision versions with minimal modifications. By casting both inputs and outputs to half precision, and introducing a dynamic loss-scaling mechanism, MPX alleviates issues like gradient underflow and overflow that commonly arise in half precision computations. Its design inherits critical features from JAX's type-promotion behavior, ensuring that operations take place in the correct precision and allowing for selective enforcement of full precision where needed (e.g., sums, means, or softmax). MPX further provides wrappers for automatic creation and management of mixed-precision gradients and optimizers, enabling straightforward integration into existing JAX training pipelines. MPX's source code, documentation, and usage examples are available at github.com/Data-Science-in-Mechanical-Engineering/mixed_precision_for_JAX .

Sebastian Weichwald, Søren Wengel Mogensen, Tabitha Edith Lee et al.

Questions in causality, control, and reinforcement learning go beyond the classical machine learning task of prediction under i.i.d. observations. Instead, these fields consider the problem of learning how to actively perturb a system to achieve a certain effect on a response variable. Arguably, they have complementary views on the problem: In control, one usually aims to first identify the system by excitation strategies to then apply model-based design techniques to control the system. In (non-model-based) reinforcement learning, one directly optimizes a reward. In causality, one focus is on identifiability of causal structure. We believe that combining the different views might create synergies and this competition is meant as a first step toward such synergies. The participants had access to observational and (offline) interventional data generated by dynamical systems. Track CHEM considers an open-loop problem in which a single impulse at the beginning of the dynamics can be set, while Track ROBO considers a closed-loop problem in which control variables can be set at each time step. The goal in both tracks is to infer controls that drive the system to a desired state. Code is open-sourced ( https://github.com/LearningByDoingCompetition/learningbydoing-comp ) to reproduce the winning solutions of the competition and to facilitate trying out new methods on the competition tasks.

Pascal Reinhold, Alexander Gräfe, Sebastian Trimpe

Drone light shows are redefining aerial entertainment, yet their widespread adoption is bottlenecked by labor-intensive, manual animation. While generative AI promises an automated alternative, current frameworks fail to provide photorealism with fluid, dynamic motion. To address this limitation, we introduce SWAN, an end-to-end pipeline that synthesizes photorealistic, large-scale, and collision-free drone choreographies directly from text prompts. SWAN converts text into realistic reference videos and translates these pixel-space dynamics into physical swarm kinematics using a novel, adaptive point-tracking algorithm. Unlike existing trackers, this method maintains spatial coherence through severe occlusions and rapid topological shifts. A dedicated planner then allocates these trajectories to individual drones, while a subsequent safety filter ensures collision-free execution. We demonstrate scalability by safely orchestrating simulated 2,000-drone formations and validate physical feasibility on a dense real-world swarm of 49 quadcopters, operating everything entirely on standard consumer hardware. Combined, this work demonstrates how generative AI can be leveraged to automate multi-robot choreography design, providing an accessible new framework for drone light shows.

Abdullah Tokmak, Christian Fiedler, Melanie N. Zeilinger et al.

Safety guarantees are vital in many control applications, such as robotics. Model predictive control (MPC) provides a constructive framework for controlling safety-critical systems, but is limited by its computational complexity. We address this problem by presenting a novel algorithm that automatically computes an explicit approximation to nonlinear MPC schemes while retaining closed-loop guarantees. Specifically, the problem can be reduced to a function approximation problem, which we then tackle by proposing ALKIA-X, the Adaptive and Localized Kernel Interpolation Algorithm with eXtrapolated reproducing kernel Hilbert space norm. ALKIA-X is a non-iterative algorithm that ensures numerically well-conditioned computations, a fast-to-evaluate approximating function, and the guaranteed satisfaction of any desired bound on the approximation error. Hence, ALKIA-X automatically computes an explicit function that approximates the MPC, yielding a controller suitable for safety-critical systems and high sampling rates. We apply ALKIA-X to approximate two nonlinear MPC schemes, demonstrating reduced computational demand and applicability to realistic problems.

Henrik Hose, Alexander Gräfe, Sebastian Trimpe

Model Predictive Control (MPC) is a method to control nonlinear systems with guaranteed stability and constraint satisfaction but suffers from high computation times. Approximate MPC (AMPC) with neural networks (NNs) has emerged to address this limitation, enabling deployment on resource-constrained embedded systems. However, when tuning AMPCs for real-world systems, large datasets need to be regenerated and the NN needs to be retrained at every tuning step. This work introduces a novel, parameter-adaptive AMPC architecture capable of online tuning without recomputing large datasets and retraining. By incorporating local sensitivities of nonlinear programs, the proposed method not only mimics optimal MPC inputs but also adjusts to known changes in physical parameters of the model using linear predictions while still guaranteeing stability. We showcase the effectiveness of parameter-adaptive AMPC by controlling the swing-ups of two different real cartpole systems with a severely resource-constrained microcontroller (MCU). We use the same NN across both system instances that have different parameters. This work not only represents the first experimental demonstration of AMPC for fast-moving systems on low-cost MCUs to the best of our knowledge, but also showcases generalization across system instances and variations through our parameter-adaptation method. Taken together, these contributions represent a marked step toward the practical application of AMPC in real-world systems.

Paul Brunzema, Mikkel Jordahn, John Willes et al.

Gaussian Processes (GPs) are widely seen as the state-of-the-art surrogate models for Bayesian optimization (BO) due to their ability to model uncertainty and their performance on tasks where correlations are easily captured (such as those defined by Euclidean metrics) and their ability to be efficiently updated online. However, the performance of GPs depends on the choice of kernel, and kernel selection for complex correlation structures is often difficult or must be made bespoke. While Bayesian neural networks (BNNs) are a promising direction for higher capacity surrogate models, they have so far seen limited use due to poor performance on some problem types. In this paper, we propose an approach which shows competitive performance on many problem types, including some that BNNs typically struggle with. We build on variational Bayesian last layers (VBLLs), and connect training of these models to exact conditioning in GPs. We exploit this connection to develop an efficient online training algorithm that interleaves conditioning and optimization. Our findings suggest that VBLL networks significantly outperform GPs and other BNN architectures on tasks with complex input correlations, and match the performance of well-tuned GPs on established benchmark tasks.

Alexander von Rohr, David Stenger, Dominik Scheurenberg et al.

Controller tuning is crucial for closed-loop performance but often involves manual adjustments. Although Bayesian optimization (BO) has been established as a data-efficient method for automated tuning, applying it to large and high-dimensional search spaces remains challenging. We extend a recently proposed local variant of BO to include crash constraints, where the controller can only be successfully evaluated in an a-priori unknown feasible region. We demonstrate the efficiency of the proposed method through simulations and hardware experiments. Our findings showcase the potential of local BO to enhance controller performance and reduce the time and resources necessary for tuning.

Martin Ziegler, Andres Felipe Posada-Moreno, Friedrich Solowjow et al.

Foundation models have demonstrated remarkable generalization, data efficiency, and robustness properties across various domains. In this paper, we explore the feasibility of foundation models for applications in the control domain. The success of these models is enabled by large-scale pretaining on Internet-scale datasets. These are available in fields like natural language processing and computer vision, but do not exist for dynamical systems. We address this challenge by pretraining a transformer-based foundation model exclusively on synthetic data and propose to sample dynamics functions from a reproducing kernel Hilbert space. Our pretrained model generalizes for prediction tasks across different dynamical systems, which we validate in simulation and hardware experiments, including cart-pole and Furuta pendulum setups. Additionally, the model can be fine-tuned effectively to new systems to increase performance even further. Our results demonstrate the feasibility of foundation models for dynamical systems that outperform specialist models in terms of generalization, data efficiency, and robustness.

Guner Dilsad Er, Sebastian Trimpe, Michael Muehlebach

We consider a distributed learning problem, where agents minimize a global objective function by exchanging information over a network. Our approach has two distinct features: (i) It substantially reduces communication by triggering communication only when necessary, and (ii) it is agnostic to the data-distribution among the different agents. We therefore guarantee convergence even if the local data-distributions of the agents are arbitrarily distinct. We analyze the convergence rate of the algorithm both in convex and nonconvex settings and derive accelerated convergence rates for the convex case. We also characterize the effect of communication failures and demonstrate that our algorithm is robust to these. The article concludes by presenting numerical results from distributed learning tasks on the MNIST and CIFAR-10 datasets. The experiments underline communication savings of 35% or more due to the event-based communication strategy, show resilience towards heterogeneous data-distributions, and highlight that our approach outperforms common baselines such as FedAvg, FedProx, SCAFFOLD and FedADMM.

Shiming He, Alexander von Rohr, Dominik Baumann et al.

How can robots learn and adapt to new tasks and situations with little data? Systematic exploration and simulation are crucial tools for efficient robot learning. We present a novel black-box policy search algorithm focused on data-efficient policy improvements. The algorithm learns directly on the robot and treats simulation as an additional information source to speed up the learning process. At the core of the algorithm, a probabilistic model learns the dependence of the policy parameters and the robot learning objective not only by performing experiments on the robot, but also by leveraging data from a simulator. This substantially reduces interaction time with the robot. Using this model, we can guarantee improvements with high probability for each policy update, thereby facilitating fast, goal-oriented learning. We evaluate our algorithm on simulated fine-tuning tasks and demonstrate the data-efficiency of the proposed dual-information source optimization algorithm. In a real robot learning experiment, we show fast and successful task learning on a robot manipulator with the aid of an imperfect simulator.

Bernd Frauenknecht, Devdutt Subhasish, Friedrich Solowjow et al.

Model-based reinforcement learning (MBRL) seeks to enhance data efficiency by learning a model of the environment and generating synthetic rollouts from it. However, accumulated model errors during these rollouts can distort the data distribution, negatively impacting policy learning and hindering long-term planning. Thus, the accumulation of model errors is a key bottleneck in current MBRL methods. We propose Infoprop, a model-based rollout mechanism that separates aleatoric from epistemic model uncertainty and reduces the influence of the latter on the data distribution. Further, Infoprop keeps track of accumulated model errors along a model rollout and provides termination criteria to limit data corruption. We demonstrate the capabilities of Infoprop in the Infoprop-Dyna algorithm, reporting state-of-the-art performance in Dyna-style MBRL on common MuJoCo benchmark tasks while substantially increasing rollout length and data quality.

Alexander Gräfe, Fabian Mager, Marco Zimmerling et al.

As Machine Learning (ML) becomes integral to Cyber-Physical Systems (CPS), there is growing interest in shifting training from traditional cloud-based to on-device processing (TinyML), for example, due to privacy and latency concerns. However, CPS often comprise ultra-low-power microcontrollers, whose limited compute resources make training challenging. This paper presents RockNet, a new TinyML method tailored for ultra-low-power hardware that achieves state-of-the-art accuracy in timeseries classification, such as fault or malware detection, without requiring offline pretraining. By leveraging that CPS consist of multiple devices, we design a distributed learning method that integrates ML and wireless communication. RockNet leverages all devices for distributed training of specialized compute efficient classifiers that need minimal communication overhead for parallelization. Combined with tailored and efficient wireless multi-hop communication protocols, our approach overcomes the communication bottleneck that often occurs in distributed learning. Hardware experiments on a testbed with 20 ultra-low-power devices demonstrate RockNet's effectiveness. It successfully learns timeseries classification tasks from scratch, surpassing the accuracy of the latest approach for neural network microcontroller training by up to 2x. RockNet's distributed ML architecture reduces memory, latency and energy consumption per device by up to 90 % when scaling from one central device to 20 devices. Our results show that a tight integration of distributed ML, distributed computing, and communication enables, for the first time, training on ultra-low-power hardware with state-of-the-art accuracy.

Pierre-François Massiani, Alexander von Rohr, Lukas Haverbeck et al.

Despite the many recent advances in reinforcement learning (RL), the question of learning policies that robustly satisfy state constraints under unknown disturbances remains open. In this paper, we offer a new perspective on achieving robust safety by analyzing the interplay between two well-established techniques in model-free RL: entropy regularization, and constraints penalization. We reveal empirically that entropy regularization in constrained RL inherently biases learning toward maximizing the number of future viable actions, thereby promoting constraints satisfaction robust to action noise. Furthermore, we show that by relaxing strict safety constraints through penalties, the constrained RL problem can be approximated arbitrarily closely by an unconstrained one and thus solved using standard model-free RL. This reformulation preserves both safety and optimality while empirically improving resilience to disturbances. Our results indicate that the connection between entropy regularization and robustness is a promising avenue for further empirical and theoretical investigation, as it enables robust safety in RL through simple reward shaping.

Pierre-François Massiani, Christian Fiedler, Lukas Haverbeck et al.

We propose a framework for hypothesis testing on conditional probability distributions, which we then use to construct statistical tests of functionals of conditional distributions. These tests identify the inputs where the functionals differ with high probability, and include tests of conditional moments or two-sample tests. Our key idea is to transform confidence bounds of a learning method into a test of conditional expectations. We instantiate this principle for kernel ridge regression (KRR) with subgaussian noise. An intermediate data embedding then enables more general tests -- including conditional two-sample tests -- via kernel mean embeddings of distributions. To have guarantees in this setting, we generalize existing pointwise-in-time or time-uniform confidence bounds for KRR to previously-inaccessible yet essential cases such as infinite-dimensional outputs with non-trace-class kernels. These bounds also circumvent the need for independent data, allowing for instance online sampling. To make our tests readily applicable in practice, we introduce bootstrapping schemes leveraging the parametric form of testing thresholds identified in theory to avoid tuning inaccessible parameters. We illustrate the tests on examples, including one in process monitoring and comparison of dynamical systems. Overall, our results establish a comprehensive foundation for conditional testing on functionals, from theoretical guarantees to an algorithmic implementation, and advance the state of the art on confidence bounds for vector-valued least squares estimation.

Antonia Holzapfel, Andres Felipe Posada-Moreno, Sebastian Trimpe

Convolutional neural networks (CNNs) for time series classification (TSC) are being increasingly used in applications ranging from quality prediction to medical diagnosis. The black box nature of these models makes understanding their prediction process difficult. This issue is crucial because CNNs are prone to learning shortcuts and biases, compromising their robustness and alignment with human expectations. To assess whether such mechanisms are being used and the associated risk, it is essential to provide model explanations that reflect the inner workings of the model. Concept Extraction (CE) methods offer such explanations, but have mostly been developed for the image domain so far, leaving a gap in the time series domain. In this work, we present a CE and localization method tailored to the time series domain, based on the ideas of CE methods for images. We propose the novel method ECLAD-ts, which provides post-hoc global explanations based on how the models encode subsets of the input at different levels of abstraction. For this, concepts are produced by clustering timestep-wise aggregations of CNN activation maps, and their importance is computed based on their impact on the prediction process. We evaluate our method on synthetic and natural datasets. Furthermore, we assess the advantages and limitations of CE in time series through empirical results. Our results show that ECLAD-ts effectively explains models by leveraging their internal representations, providing useful insights about their prediction process.

Emma Cramer, Lukas Jäschke, Sebastian Trimpe

Robotic systems are increasingly employed for industrial automation, with contact-rich tasks like polishing requiring dexterity and compliant behaviour. These tasks are difficult to model, making classical control challenging. Deep reinforcement learning (RL) offers a promising solution by enabling the learning of models and control policies directly from data. However, its application to real-world problems is limited by data inefficiency and unsafe exploration. Adaptive hybrid RL methods blend classical control and RL adaptively, combining the strengths of both: structure from control and learning from RL. This has led to improvements in data efficiency and exploration safety. However, their potential for hardware applications remains underexplored, with no evaluations on physical systems to date. Such evaluations are critical to fully assess the practicality and effectiveness of these methods in real-world settings. This work presents an experimental demonstration of the hybrid RL algorithm CHEQ for robotic polishing with variable impedance, a task requiring precise force and velocity tracking. In simulation, we show that variable impedance enhances polishing performance. We compare standalone RL with adaptive hybrid RL, demonstrating that CHEQ achieves effective learning while adhering to safety constraints. On hardware, CHEQ achieves effective polishing behaviour, requiring only eight hours of training and incurring just five failures. These results highlight the potential of adaptive hybrid RL for real-world, contact-rich tasks trained directly on hardware.

Emma Cramer, Bernd Frauenknecht, Ramil Sabirov et al.

Combining Reinforcement Learning (RL) with a prior controller can yield the best out of two worlds: RL can solve complex nonlinear problems, while the control prior ensures safer exploration and speeds up training. Prior work largely blends both components with a fixed weight, neglecting that the RL agent's performance varies with the training progress and across regions in the state space. Therefore, we advocate for an adaptive strategy that dynamically adjusts the weighting based on the RL agent's current capabilities. We propose a new adaptive hybrid RL algorithm, Contextualized Hybrid Ensemble Q-learning (CHEQ). CHEQ combines three key ingredients: (i) a time-invariant formulation of the adaptive hybrid RL problem treating the adaptive weight as a context variable, (ii) a weight adaption mechanism based on the parametric uncertainty of a critic ensemble, and (iii) ensemble-based acceleration for data-efficient RL. Evaluating CHEQ on a car racing task reveals substantially stronger data efficiency, exploration safety, and transferability to unknown scenarios than state-of-the-art adaptive hybrid RL methods.

Julian Dierkes, Emma Cramer, Holger H. Hoos et al.

There has been significant progress in deep reinforcement learning (RL) in recent years. Nevertheless, finding suitable hyperparameter configurations and reward functions remains challenging even for experts, and performance heavily relies on these design choices. Also, most RL research is conducted on known benchmarks where knowledge about these choices already exists. However, novel practical applications often pose complex tasks for which no prior knowledge about good hyperparameters and reward functions is available, thus necessitating their derivation from scratch. Prior work has examined automatically tuning either hyperparameters or reward functions individually. We demonstrate empirically that an RL algorithm's hyperparameter configurations and reward function are often mutually dependent, meaning neither can be fully optimised without appropriate values for the other. We then propose a methodology for the combined optimisation of hyperparameters and the reward function. Furthermore, we include a variance penalty as an optimisation objective to improve the stability of learned policies. We conducted extensive experiments using Proximal Policy Optimisation and Soft Actor-Critic on four environments. Our results show that combined optimisation significantly improves over baseline performance in half of the environments and achieves competitive performance in the others, with only a minor increase in computational costs. This suggests that combined optimisation should be best practice.

Pierre-François Massiani, Sebastian Trimpe, Friedrich Solowjow

The consistency of a learning method is usually established under the assumption that the observations are a realization of an independent and identically distributed (i.i.d.) or mixing process. Yet, kernel methods such as support vector machines (SVMs), Gaussian processes, or conditional kernel mean embeddings (CKMEs) all give excellent performance under sampling schemes that are obviously non-i.i.d., such as when data comes from a dynamical system. We propose the new notion of empirical weak convergence (EWC) as a general assumption explaining such phenomena for kernel methods. It assumes the existence of a random asymptotic data distribution and is a strict weakening of previous assumptions in the field. Our main results then establish consistency of SVMs, kernel mean embeddings, and general Hilbert-space valued empirical expectations with EWC data. Our analysis holds for both finite- and infinite-dimensional outputs, as we extend classical results of statistical learning to the latter case. In particular, it is also applicable to CKMEs. Overall, our results open new classes of processes to statistical learning and can serve as a foundation for a theory of learning beyond i.i.d. and mixing.

Christian Fiedler, Johanna Menn, Lukas Kreisköther et al.

Optimizing an unknown function under safety constraints is a central task in robotics, biomedical engineering, and many other disciplines, and increasingly safe Bayesian Optimization (BO) is used for this. Due to the safety critical nature of these applications, it is of utmost importance that theoretical safety guarantees for these algorithms translate into the real world. In this work, we investigate three safety-related issues of the popular class of SafeOpt-type algorithms. First, these algorithms critically rely on frequentist uncertainty bounds for Gaussian Process (GP) regression, but concrete implementations typically utilize heuristics that invalidate all safety guarantees. We provide a detailed analysis of this problem and introduce Real-\b{eta}-SafeOpt, a variant of the SafeOpt algorithm that leverages recent GP bounds and thus retains all theoretical guarantees. Second, we identify assuming an upper bound on the reproducing kernel Hilbert space (RKHS) norm of the target function, a key technical assumption in SafeOpt-like algorithms, as a central obstacle to real-world usage. To overcome this challenge, we introduce the Lipschitz-only Safe Bayesian Optimization (LoSBO) algorithm, which guarantees safety without an assumption on the RKHS bound, and empirically show that this algorithm is not only safe, but also exhibits superior performance compared to the state-of-the-art on several function classes. Third, SafeOpt and derived algorithms rely on a discrete search space, making them difficult to apply to higher-dimensional problems. To widen the applicability of these algorithms, we introduce Lipschitz-only GP-UCB (LoS-GP-UCB), a variant of LoSBO applicable to moderately high-dimensional problems, while retaining safety.

Sarah Müller, Alexander von Rohr, Sebastian Trimpe

Reinforcement learning (RL) aims to find an optimal policy by interaction with an environment. Consequently, learning complex behavior requires a vast number of samples, which can be prohibitive in practice. Nevertheless, instead of systematically reasoning and actively choosing informative samples, policy gradients for local search are often obtained from random perturbations. These random samples yield high variance estimates and hence are sub-optimal in terms of sample complexity. Actively selecting informative samples is at the core of Bayesian optimization, which constructs a probabilistic surrogate of the objective from past samples to reason about informative subsequent ones. In this paper, we propose to join both worlds. We develop an algorithm utilizing a probabilistic model of the objective function and its gradient. Based on the model, the algorithm decides where to query a noisy zeroth-order oracle to improve the gradient estimates. The resulting algorithm is a novel type of policy search method, which we compare to existing black-box algorithms. The comparison reveals improved sample complexity and reduced variance in extensive empirical evaluations on synthetic objectives. Further, we highlight the benefits of active sampling on popular RL benchmarks.

Dominik Baumann, Alonso Marco, Matteo Turchetta et al.

When learning policies for robotic systems from data, safety is a major concern, as violation of safety constraints may cause hardware damage. SafeOpt is an efficient Bayesian optimization (BO) algorithm that can learn policies while guaranteeing safety with high probability. However, its search space is limited to an initially given safe region. We extend this method by exploring outside the initial safe area while still guaranteeing safety with high probability. This is achieved by learning a set of initial conditions from which we can recover safely using a learned backup controller in case of a potential failure. We derive conditions for guaranteed convergence to the global optimum and validate GoSafe in hardware experiments.

Pierre-François Massiani, Steve Heim, Friedrich Solowjow et al.

Safety constraints and optimality are important, but sometimes conflicting criteria for controllers. Although these criteria are often solved separately with different tools to maintain formal guarantees, it is also common practice in reinforcement learning to simply modify reward functions by penalizing failures, with the penalty treated as a mere heuristic. We rigorously examine the relationship of both safety and optimality to penalties, and formalize sufficient conditions for safe value functions (SVFs): value functions that are both optimal for a given task, and enforce safety constraints. We reveal this structure by examining when rewards preserve viability under optimal control, and show that there always exists a finite penalty that induces a safe value function. This penalty is not unique, but upper-unbounded: larger penalties do not harm optimality. Although it is often not possible to compute the minimum required penalty, we reveal clear structure of how the penalty, rewards, discount factor, and dynamics interact. This insight suggests practical, theory-guided heuristics to design reward functions for control problems where safety is important.

Alexander von Rohr, Matthias Neumann-Brosig, Sebastian Trimpe

Probabilistic models such as Gaussian processes (GPs) are powerful tools to learn unknown dynamical systems from data for subsequent use in control design. While learning-based control has the potential to yield superior performance in demanding applications, robustness to uncertainty remains an important challenge. Since Bayesian methods quantify uncertainty of the learning results, it is natural to incorporate these uncertainties into a robust design. In contrast to most state-of-the-art approaches that consider worst-case estimates, we leverage the learning method's posterior distribution in the controller synthesis. The result is a more informed and, thus, more efficient trade-off between performance and robustness. We present a novel controller synthesis for linearized GP dynamics that yields robust controllers with respect to a probabilistic stability margin. The formulation is based on a recently proposed algorithm for linear quadratic control synthesis, which we extend by giving probabilistic robustness guarantees in the form of credibility bounds for the system's stability.Comparisons to existing methods based on worst-case and certainty-equivalence designs reveal superior performance and robustness properties of the proposed method.

Christian Fiedler, Carsten W. Scherer, Sebastian Trimpe

The combination of machine learning with control offers many opportunities, in particular for robust control. However, due to strong safety and reliability requirements in many real-world applications, providing rigorous statistical and control-theoretic guarantees is of utmost importance, yet difficult to achieve for learning-based control schemes. We present a general framework for learning-enhanced robust control that allows for systematic integration of prior engineering knowledge, is fully compatible with modern robust control and still comes with rigorous and practically meaningful guarantees. Building on the established Linear Fractional Representation and Integral Quadratic Constraints framework, we integrate Gaussian Process Regression as a learning component and state-of-the-art robust controller synthesis. In a concrete robust control example, our approach is demonstrated to yield improved performance with more data, while guarantees are maintained throughout.

Christian Fiedler, Carsten W. Scherer, Sebastian Trimpe

Gaussian Process Regression is a popular nonparametric regression method based on Bayesian principles that provides uncertainty estimates for its predictions. However, these estimates are of a Bayesian nature, whereas for some important applications, like learning-based control with safety guarantees, frequentist uncertainty bounds are required. Although such rigorous bounds are available for Gaussian Processes, they are too conservative to be useful in applications. This often leads practitioners to replacing these bounds by heuristics, thus breaking all theoretical guarantees. To address this problem, we introduce new uncertainty bounds that are rigorous, yet practically useful at the same time. In particular, the bounds can be explicitly evaluated and are much less conservative than state of the art results. Furthermore, we show that certain model misspecifications lead to only graceful degradation. We demonstrate these advantages and the usefulness of our results for learning-based control with numerical examples.