Joschka Boedecker

Shamil Mamedov, Rudolf Reiter, Seyed Mahdi Basiri Azad et al.

Flexible robots may overcome some of the industry's major challenges, such as enabling intrinsically safe human-robot collaboration and achieving a higher payload-to-mass ratio. However, controlling flexible robots is complicated due to their complex dynamics, which include oscillatory behavior and a high-dimensional state space. Nonlinear model predictive control (NMPC) offers an effective means to control such robots, but its significant computational demand often limits its application in real-time scenarios. To enable fast control of flexible robots, we propose a framework for a safe approximation of NMPC using imitation learning and a predictive safety filter. Our framework significantly reduces computation time while incurring a slight loss in performance. Compared to NMPC, our framework shows more than an eightfold improvement in computation time when controlling a three-dimensional flexible robot arm in simulation, all while guaranteeing safety constraints. Notably, our approach outperforms state-of-the-art reinforcement learning methods. The development of fast and safe approximate NMPC holds the potential to accelerate the adoption of flexible robots in industry. The project code is available at: tinyurl.com/anmpc4fr

Erick Rosete-Beas, Oier Mees, Gabriel Kalweit et al.

Everyday tasks of long-horizon and comprising a sequence of multiple implicit subtasks still impose a major challenge in offline robot control. While a number of prior methods aimed to address this setting with variants of imitation and offline reinforcement learning, the learned behavior is typically narrow and often struggles to reach configurable long-horizon goals. As both paradigms have complementary strengths and weaknesses, we propose a novel hierarchical approach that combines the strengths of both methods to learn task-agnostic long-horizon policies from high-dimensional camera observations. Concretely, we combine a low-level policy that learns latent skills via imitation learning and a high-level policy learned from offline reinforcement learning for skill-chaining the latent behavior priors. Experiments in various simulated and real robot control tasks show that our formulation enables producing previously unseen combinations of skills to reach temporally extended goals by "stitching" together latent skills through goal chaining with an order-of-magnitude improvement in performance upon state-of-the-art baselines. We even learn one multi-task visuomotor policy for 25 distinct manipulation tasks in the real world which outperforms both imitation learning and offline reinforcement learning techniques.

Jessica Borja-Diaz, Oier Mees, Gabriel Kalweit et al.

Robots operating in human-centered environments should have the ability to understand how objects function: what can be done with each object, where this interaction may occur, and how the object is used to achieve a goal. To this end, we propose a novel approach that extracts a self-supervised visual affordance model from human teleoperated play data and leverages it to enable efficient policy learning and motion planning. We combine model-based planning with model-free deep reinforcement learning (RL) to learn policies that favor the same object regions favored by people, while requiring minimal robot interactions with the environment. We evaluate our algorithm, Visual Affordance-guided Policy Optimization (VAPO), with both diverse simulation manipulation tasks and real world robot tidy-up experiments to demonstrate the effectiveness of our affordance-guided policies. We find that our policies train 4x faster than the baselines and generalize better to novel objects because our visual affordance model can anticipate their affordance regions.

Hao Zhu, Jasper Hoffmann, Baohe Zhang et al.

We consider the problem of fitting a reinforcement learning (RL) model to some given behavioral data under a multi-armed bandit environment. These models have received much attention in recent years for characterizing human and animal decision making behavior. We provide a generic mathematical optimization problem formulation for the fitting problem of a wide range of RL models that appear frequently in scientific research applications, followed by a detailed theoretical analysis of its convexity properties. Based on the theoretical results, we introduce a novel solution method for the fitting problem of RL models based on convex relaxation and optimization. Our method is then evaluated in several simulated bandit environments to compare with some benchmark methods that appear in the literature. Numerical results indicate that our method achieves comparable performance to the state-of-the-art, while significantly reducing computation time. We also provide an open-source Python package for our proposed method to empower researchers to apply it in the analysis of their datasets directly, without prior knowledge of convex optimization.

Hao Zhu, Joschka Boedecker

We introduce disciplined biconvex programming (DBCP), a modeling framework for specifying and solving biconvex optimization problems. Biconvex optimization problems arise in various applications, including machine learning, signal processing, computational science, and control. Solving a biconvex optimization problem in practice usually resolves to heuristic methods based on alternate convex search (ACS), which iteratively optimizes over one block of variables while keeping the other fixed, so that the resulting subproblems are convex and can be efficiently solved. However, designing and implementing an ACS solver for a specific biconvex optimization problem usually requires significant effort from the user, which can be tedious and error-prone. DBCP extends the principles of disciplined convex programming to biconvex problems, allowing users to specify biconvex optimization problems in a natural way based on a small number of syntax rules. The resulting problem can then be automatically split and transformed into convex subproblems, for which a customized ACS solver is then generated and applied. DBCP allows users to quickly experiment with different biconvex problem formulations, without expertise in convex optimization. We implement DBCP into the open source Python package dbcp, as an extension to the famous domain specific language CVXPY for convex optimization.

Andrea Ghezzi, Jasper Hoffman, Jonathan Frey et al.

This work presents a novel loss function for learning nonlinear Model Predictive Control policies via Imitation Learning. Standard approaches to Imitation Learning neglect information about the expert and generally adopt a loss function based on the distance between expert and learned controls. In this work, we present a loss based on the Q-function directly embedding the performance objectives and constraint satisfaction of the associated Optimal Control Problem (OCP). However, training a Neural Network with the Q-loss requires solving the associated OCP for each new sample. To alleviate the computational burden, we derive a second Q-loss based on the Gauss-Newton approximation of the OCP resulting in a faster training time. We validate our losses against Behavioral Cloning, the standard approach to Imitation Learning, on the control of a nonlinear system with constraints. The final results show that the Q-function-based losses significantly reduce the amount of constraint violations while achieving comparable or better closed-loop costs.

Mariella Dreissig, Dominik Scheuble, Florian Piewak et al.

Autonomous vehicles rely on a variety of sensors to gather information about their surrounding. The vehicle's behavior is planned based on the environment perception, making its reliability crucial for safety reasons. The active LiDAR sensor is able to create an accurate 3D representation of a scene, making it a valuable addition for environment perception for autonomous vehicles. Due to light scattering and occlusion, the LiDAR's performance change under adverse weather conditions like fog, snow or rain. This limitation recently fostered a large body of research on approaches to alleviate the decrease in perception performance. In this survey, we gathered, analyzed, and discussed different aspects on dealing with adverse weather conditions in LiDAR-based environment perception. We address topics such as the availability of appropriate data, raw point cloud processing and denoising, robust perception algorithms and sensor fusion to mitigate adverse weather induced shortcomings. We furthermore identify the most pressing gaps in the current literature and pinpoint promising research directions.

Marieke Wesselkamp, Matthew Chantry, Ewan Pinnington et al.

Most useful weather prediction for the public is near the surface. The processes that are most relevant for near-surface weather prediction are also those that are most interactive and exhibit positive feedback or have key role in energy partitioning. Land surface models (LSMs) consider these processes together with surface heterogeneity and forecast water, carbon and energy fluxes, and coupled with an atmospheric model provide boundary and initial conditions. This numerical parametrization of atmospheric boundaries being computationally expensive, statistical surrogate models are increasingly used to accelerated progress in experimental research. We evaluated the efficiency of three surrogate models in speeding up experimental research by simulating land surface processes, which are integral to forecasting water, carbon, and energy fluxes in coupled atmospheric models. Specifically, we compared the performance of a Long-Short Term Memory (LSTM) encoder-decoder network, extreme gradient boosting, and a feed-forward neural network within a physics-informed multi-objective framework. This framework emulates key states of the ECMWF's Integrated Forecasting System (IFS) land surface scheme, ECLand, across continental and global scales. Our findings indicate that while all models on average demonstrate high accuracy over the forecast period, the LSTM network excels in continental long-range predictions when carefully tuned, the XGB scores consistently high across tasks and the MLP provides an excellent implementation-time-accuracy trade-off. The runtime reduction achieved by the emulators in comparison to the full numerical models are significant, offering a faster, yet reliable alternative for conducting numerical experiments on land surfaces.

Hao Zhu, Brice De La Crompe, Gabriel Kalweit et al.

In advancing the understanding of natural decision-making processes, inverse reinforcement learning (IRL) methods have proven instrumental in reconstructing animal's intentions underlying complex behaviors. Given the recent development of a continuous-time multi-intention IRL framework, there has been persistent inquiry into inferring discrete time-varying rewards with IRL. To address this challenge, we introduce the class of hierarchical inverse Q-learning (HIQL) algorithms. Through an unsupervised learning process, HIQL divides expert trajectories into multiple intention segments, and solves the IRL problem independently for each. Applying HIQL to simulated experiments and several real animal behavior datasets, our approach outperforms current benchmarks in behavior prediction and produces interpretable reward functions. Our results suggest that the intention transition dynamics underlying complex decision-making behavior is better modeled by a step function instead of a smoothly varying function. This advancement holds promise for neuroscience and cognitive science, contributing to a deeper understanding of decision-making and uncovering underlying brain mechanisms.

Branka Mirchevska, Moritz Werling, Joschka Boedecker

Implementing an autonomous vehicle that is able to output feasible, smooth and efficient trajectories is a long-standing challenge. Several approaches have been considered, roughly falling under two categories: rule-based and learning-based approaches. The rule-based approaches, while guaranteeing safety and feasibility, fall short when it comes to long-term planning and generalization. The learning-based approaches are able to account for long-term planning and generalization to unseen situations, but may fail to achieve smoothness, safety and the feasibility which rule-based approaches ensure. Hence, combining the two approaches is an evident step towards yielding the best compromise out of both. We propose a Reinforcement Learning-based approach, which learns target trajectory parameters for fully autonomous driving on highways. The trained agent outputs continuous trajectory parameters based on which a feasible polynomial-based trajectory is generated and executed. We compare the performance of our agent against four other highway driving agents. The experiments are conducted in the Sumo simulator, taking into consideration various realistic, dynamically changing highway scenarios, including surrounding vehicles with different driver behaviors. We demonstrate that our offline trained agent, with randomly collected data, learns to drive smoothly, achieving velocities as close as possible to the desired velocity, while outperforming the other agents. Code, training data and details available at: https://nrgit.informatik.uni-freiburg. de/branka.mirchevska/offline-rl-tp.

Lukas AW Gemein, Robin T Schirrmeister, Joschka Boedecker et al.

The brain's biological age has been considered as a promising candidate for a neurologically significant biomarker. However, recent results based on longitudinal magnetic resonance imaging data have raised questions on its interpretation. A central question is whether an increased biological age of the brain is indicative of brain pathology and if changes in brain age correlate with diagnosed pathology (state hypothesis). Alternatively, could the discrepancy in brain age be a stable characteristic unique to each individual (trait hypothesis)? To address this question, we present a comprehensive study on brain aging based on clinical EEG, which is complementary to previous MRI-based investigations. We apply a state-of-the-art Temporal Convolutional Network (TCN) to the task of age regression. We train on recordings of the Temple University Hospital EEG Corpus (TUEG) explicitly labeled as non-pathological and evaluate on recordings of subjects with non-pathological as well as pathological recordings, both with examinations at a single point in time and repeated examinations over time. Therefore, we created four novel subsets of TUEG that include subjects with multiple recordings: I) all labeled non-pathological; II) all labeled pathological; III) at least one recording labeled non-pathological followed by at least one recording labeled pathological; IV) similar to III) but with opposing transition (first pathological then non-pathological). The results show that our TCN reaches state-of-the-art performance in age decoding with a mean absolute error of 6.6 years. Our extensive analyses demonstrate that the model significantly underestimates the age of non-pathological and pathological subjects (-1 and -5 years, paired t-test, p <= 0.18 and p <= 0.0066). Furthermore, the brain age gap biomarker is not indicative of pathological EEG.

Mariella Dreissig, Florian Piewak, Joschka Boedecker

The confidence calibration of deep learning-based perception models plays a crucial role in their reliability. Especially in the context of autonomous driving, downstream tasks like prediction and planning depend on accurate confidence estimates. In point-wise multiclass classification tasks like sematic segmentation the model has to deal with heavy class imbalances. Due to their underrepresentation, the confidence calibration of classes with smaller instances is challenging but essential, not only for safety reasons. We propose a metric to measure the confidence calibration quality of a semantic segmentation model with respect to individual classes. It is calculated by computing sparsification curves for each class based on the uncertainty estimates. We use the classification calibration metric to evaluate uncertainty estimation methods with respect to their confidence calibration of underrepresented classes. We furthermore suggest a double use for the method to automatically find label problems to improve the quality of hand- or auto-annotated datasets.

Yuan Zhang, Jianhong Wang, Joschka Boedecker

Reinforcement learning (RL) is recognized as lacking generalization and robustness under environmental perturbations, which excessively restricts its application for real-world robotics. Prior work claimed that adding regularization to the value function is equivalent to learning a robust policy with uncertain transitions. Although the regularization-robustness transformation is appealing for its simplicity and efficiency, it is still lacking in continuous control tasks. In this paper, we propose a new regularizer named $\textbf{U}$ncertainty $\textbf{S}$et $\textbf{R}$egularizer (USR), by formulating the uncertainty set on the parameter space of the transition function. In particular, USR is flexible enough to be plugged into any existing RL framework. To deal with unknown uncertainty sets, we further propose a novel adversarial approach to generate them based on the value function. We evaluate USR on the Real-world Reinforcement Learning (RWRL) benchmark, demonstrating improvements in the robust performance for perturbed testing environments.

Mehdi Naouar, Gabriel Kalweit, Ignacio Mastroleo et al.

Cancer detection and classification from gigapixel whole slide images of stained tissue specimens has recently experienced enormous progress in computational histopathology. The limitation of available pixel-wise annotated scans shifted the focus from tumor localization to global slide-level classification on the basis of (weakly-supervised) multiple-instance learning despite the clinical importance of local cancer detection. However, the worse performance of these techniques in comparison to fully supervised methods has limited their usage until now for diagnostic interventions in domains of life-threatening diseases such as cancer. In this work, we put the focus back on tumor localization in form of a patch-level classification task and take up the setting of so-called coarse annotations, which provide greater training supervision while remaining feasible from a clinical standpoint. To this end, we present a novel ensemble method that not only significantly improves the detection accuracy of metastasis on the open CAMELYON16 data set of sentinel lymph nodes of breast cancer patients, but also considerably increases its robustness against noise while training on coarse annotations. Our experiments show that better results can be achieved with our technique making it clinically feasible to use for cancer diagnosis and opening a new avenue for translational and clinical research.

Wenyuan Sheng, Hao Zhu, Joschka Boedecker

Inverse reinforcement learning (IRL) recovers reward functions from observed behavior, yet traditional methods assume a single stationary reward that cannot capture goal switching within an episode. Recent multi-intention IRL methods address this by segmenting trajectories, but model intention transitions as either a memoryless Markov chain or via manual state augmentation with a fixed history window. We propose the Probabilistic Recurrent Intention Switching Model (PRISM), which replaces both mechanisms with a lightweight recurrent network that maps observation history to a per-step intention distribution. We prove that the resulting EM objective decomposes exactly into independent per-intention reward subproblems, each solvable in closed form, yielding an $\mathcal{O}(nK)$ E-step with no variational approximation. We evaluate PRISM on a non-Markovian gridworld, a mouse labyrinth, and BridgeData~V2 robotic manipulation, the first large-scale robotic application of multi-intention IRL. Across all settings PRISM achieves the highest held-out log-likelihood while recovering nameable, temporally coherent intentions from unlabeled demonstrations, suggesting that discrete goal switching is present in both biological and artificial agents.

Yannick Vogt, Mehdi Naouar, Maria Kalweit et al.

Antibodies offer great potential for the treatment of various diseases. However, the discovery of therapeutic antibodies through traditional wet lab methods is expensive and time-consuming. The use of generative models in designing antibodies therefore holds great promise, as it can reduce the time and resources required. Recently, the class of diffusion models has gained considerable traction for their ability to synthesize diverse and high-quality samples. In their basic form, however, they lack mechanisms to optimize for specific properties, such as binding affinity to an antigen. In contrast, the class of offline Reinforcement Learning (RL) methods has demonstrated strong performance in navigating large search spaces, including scenarios where frequent real-world interaction, such as interaction with a wet lab, is impractical. Our novel method, BetterBodies, which combines Variational Autoencoders (VAEs) with RL guided latent diffusion, is able to generate novel sets of antibody CDRH3 sequences from different data distributions. Using the Absolut! simulator, we demonstrate the improved affinity of our novel sequences to the SARS-CoV spike receptor-binding domain. Furthermore, we reflect biophysical properties in the VAE latent space using a contrastive loss and add a novel Q-function based filtering to enhance the affinity of generated sequences. In conclusion, methods such as ours have the potential to have great implications for real-world biological sequence design, where the generation of novel high-affinity binders is a cost-intensive endeavor.

Yannick Vogt, Mehdi Naouar, Maria Kalweit et al.

The field of antibody-based therapeutics has grown significantly in recent years, with targeted antibodies emerging as a potentially effective approach to personalized therapies. Such therapies could be particularly beneficial for complex, highly individual diseases such as cancer. However, progress in this field is often constrained by the extensive search space of amino acid sequences that form the foundation of antibody design. In this study, we introduce a novel reinforcement learning method specifically tailored to address the unique challenges of this domain. We demonstrate that our method can learn the design of high-affinity antibodies against multiple targets in silico, utilizing either online interaction or offline datasets. To the best of our knowledge, our approach is the first of its kind and outperforms existing methods on all tested antigens in the Absolut! database.

Suresh Guttikonda, Jan Achterhold, Haolong Li et al.

In autonomous navigation settings, several quantities can be subject to variations. Terrain properties such as friction coefficients may vary over time depending on the location of the robot. Also, the dynamics of the robot may change due to, e.g., different payloads, changing the system's mass, or wear and tear, changing actuator gains or joint friction. An autonomous agent should thus be able to adapt to such variations. In this paper, we develop a novel probabilistic, terrain- and robot-aware forward dynamics model, termed TRADYN, which is able to adapt to the above-mentioned variations. It builds on recent advances in meta-learning forward dynamics models based on Neural Processes. We evaluate our method in a simulated 2D navigation setting with a unicycle-like robot and different terrain layouts with spatially varying friction coefficients. In our experiments, the proposed model exhibits lower prediction error for the task of long-horizon trajectory prediction, compared to non-adaptive ablation models. We also evaluate our model on the downstream task of navigation planning, which demonstrates improved performance in planning control-efficient paths by taking robot and terrain properties into account.

Jan Ole von Hartz, Tim Welschehold, Abhinav Valada et al.

Task Parametrized Gaussian Mixture Models (TP-GMM) are a sample-efficient method for learning object-centric robot manipulation tasks. However, there are several open challenges to applying TP-GMMs in the wild. In this work, we tackle three crucial challenges synergistically. First, end-effector velocities are non-Euclidean and thus hard to model using standard GMMs. We thus propose to factorize the robot's end-effector velocity into its direction and magnitude, and model them using Riemannian GMMs. Second, we leverage the factorized velocities to segment and sequence skills from complex demonstration trajectories. Through the segmentation, we further align skill trajectories and hence leverage time as a powerful inductive bias. Third, we present a method to automatically detect relevant task parameters per skill from visual observations. Our approach enables learning complex manipulation tasks from just five demonstrations while using only RGB-D observations. Extensive experimental evaluations on RLBench demonstrate that our approach achieves state-of-the-art performance with 20-fold improved sample efficiency. Our policies generalize across different environments, object instances, and object positions, while the learned skills are reusable.

Yuan Zhang, Joschka Boedecker, Chuxuan Li et al.

Model Predictive Control (MPC) is attracting tremendous attention in the autonomous driving task as a powerful control technique. The success of an MPC controller strongly depends on an accurate internal dynamics model. However, the static parameters, usually learned by system identification, often fail to adapt to both internal and external perturbations in real-world scenarios. In this paper, we firstly (1) reformulate the problem as a Partially Observed Markov Decision Process (POMDP) that absorbs the uncertainties into observations and maintains Markov property into hidden states; and (2) learn a recurrent policy continually adapting the parameters of the dynamics model via Recurrent Reinforcement Learning (RRL) for optimal and adaptive control; and (3) finally evaluate the proposed algorithm (referred as $\textit{MPC-RRL}$) in CARLA simulator and leading to robust behaviours under a wide range of perturbations.

Mariella Dreissig, Florian Piewak, Joschka Boedecker

The calibration of deep learning-based perception models plays a crucial role in their reliability. Our work focuses on a class-wise evaluation of several model's confidence performance for LiDAR-based semantic segmentation with the aim of providing insights into the calibration of underrepresented classes. Those classes often include VRUs and are thus of particular interest for safety reasons. With the help of a metric based on sparsification curves we compare the calibration abilities of three semantic segmentation models with different architectural concepts, each in a in deterministic and a probabilistic version. By identifying and describing the dependency between the predictive performance of a class and the respective calibration quality we aim to facilitate the model selection and refinement for safety-critical applications.

Seyed Mahdi B. Azad, Jasper Hoffmann, Iman Nematollahi et al.

Forward-backward (FB) representations provide a powerful framework for learning the successor representation (SR) in continuous spaces by enforcing a low-rank factorization. However, a fundamental spectral mismatch often exists between the high-rank transition dynamics of continuous environments and the low-rank bottleneck of the FB architecture, making accurate low-rank representation learning difficult. In this work, we analyze temporal abstraction as a mechanism to mitigate this mismatch. By characterizing the spectral properties of the transition operator, we show that temporal abstraction acts as a low-pass filter that suppresses high-frequency spectral components. This suppression reduces the effective rank of the induced SR while preserving a formal bound on the resulting value function error. Empirically, we show that this alignment is a key factor for stable FB learning, particularly at high discount factors where bootstrapping becomes error-prone. Our results identify temporal abstraction as a principled mechanism for shaping the spectral structure of the underlying MDP and enabling effective long-horizon representations in continuous control.

Omar Swelam, Lennart Purucker, Jake Robertson et al.

Causal discovery is fundamental for multiple scientific domains, yet extracting causal information from real world data remains a significant challenge. Given the recent success on real data, we investigate whether TabPFN, a transformer-based tabular foundation model pre-trained on synthetic datasets generated from structural causal models, encodes causal information in its internal representations. We develop an adapter framework using a learnable decoder and causal tokens that extract causal signals from TabPFN's frozen embeddings and decode them into adjacency matrices for causal discovery. Our evaluations demonstrate that TabPFN's embeddings contain causal information, outperforming several traditional causal discovery algorithms, with such causal information being concentrated in mid-range layers. These findings establish a new direction for interpretable and adaptable foundation models and demonstrate the potential for leveraging pre-trained tabular models for causal discovery.

Yuan Zhang, Shaohui Yang, Toshiyuki Ohtsuka et al.

Model predictive control (MPC) has played a more crucial role in various robotic control tasks, but its high computational requirements are concerning, especially for nonlinear dynamical models. This paper presents a $\textbf{la}$tent $\textbf{l}$inear $\textbf{q}$uadratic $\textbf{r}$egulator (LaLQR) that maps the state space into a latent space, on which the dynamical model is linear and the cost function is quadratic, allowing the efficient application of LQR. We jointly learn this alternative system by imitating the original MPC. Experiments show LaLQR's superior efficiency and generalization compared to other baselines.

Hao Zhu, Shengchao Yan, Jasper Hoffmann et al.

Discrete latent factor models (DLFMs) are widely used in various domains such as machine learning, economics, neuroscience, psychology, etc. Currently, fitting a DLFM to some dataset relies on a customized solver for individual models, which requires lots of effort to implement and is limited to the targeted specific instance of DLFMs. In this paper, we propose a generic framework based on CVXPY, which allows users to specify and solve the fitting problem of a wide range of DLFMs, including both regression and classification models, within a very short script. Our framework is flexible and inherently supports the integration of regularization terms and constraints on the DLFM parameters and latent factors, such that the users can easily prototype the DLFM structure according to their dataset and application scenario. We introduce our open-source Python implementation and illustrate the framework in several examples.

Aron Distelzweig, Faris Janjoš, Andreas Look et al.

Recent Autonomous Driving (AD) works such as GigaFlow and PufferDrive have unlocked Reinforcement Learning (RL) at scale as a training strategy for driving policies. Yet such policies remain disconnected from established benchmarks, leaving the performance of large-scale RL for driving on standardized evaluations unknown. We present BehaviorBench -- a comprehensive test suite that closes this gap along three axes: Evaluation, Complexity, and Behavior Diversity. In terms of Evaluation, we provide an interface connecting PufferDrive to nuPlan, which, for the first time, enables policies trained via RL at scale to be evaluated on an established planning benchmark for autonomous driving. Complementarily, we offer an evaluation framework that allows planners to be benchmarked directly inside the PufferDrive simulation, at a fraction of the time. Regarding Complexity, we observe that today's standardized benchmarks are so simple that near-perfect scores are achievable by straight lane following with collision checking. We extract a meaningful, interaction-rich split from the Waymo Open Motion Dataset (WOMD) on which strong performance is impossible without multi-agent reasoning. Lastly, we address Behavior Diversity. Existing benchmarks commonly evaluate planners against a single rule-based traffic model, the Intelligent Driver Model (IDM). We provide a diverse suite of interactive traffic agents to stress-test policies under heterogeneous behaviors, beyond just using IDM. Overall, our benchmarking analysis uncovers the following insight: despite learning interactive behaviors in an emergent manner, policies trained via pure self-play under standard reward functions overfit to their training opponents and fail to generalize to other traffic agent behaviors. Building on this observation, we propose a hybrid planner that combines a PPO policy with a rule-based planner.

Jan Ole von Hartz, Lukas Schweizer, Joschka Boedecker et al.

Generative robot policies such as Flow Matching offer flexible, multi-modal policy learning but are sample-inefficient. Although object-centric policies improve sample efficiency, it does not resolve this limitation. In this work, we propose Multi-Stream Generative Policy (MSG), an inference-time composition framework that trains multiple object-centric policies and combines them at inference to improve generalization and sample efficiency. MSG is model-agnostic and inference-only, hence widely applicable to various generative policies and training paradigms. We perform extensive experiments both in simulation and on a real robot, demonstrating that our approach learns high-quality generative policies from as few as five demonstrations, resulting in a 95% reduction in demonstrations, and improves policy performance by 89 percent compared to single-stream approaches. Furthermore, we present comprehensive ablation studies on various composition strategies and provide practical recommendations for deployment. Finally, MSG enables zero-shot object instance transfer. We make our code publicly available at https://msg.cs.uni-freiburg.de.

Branka Mirchevska, Maria Hügle, Gabriel Kalweit et al.

Well-established optimization-based methods can guarantee an optimal trajectory for a short optimization horizon, typically no longer than a few seconds. As a result, choosing the optimal trajectory for this short horizon may still result in a sub-optimal long-term solution. At the same time, the resulting short-term trajectories allow for effective, comfortable and provable safe maneuvers in a dynamic traffic environment. In this work, we address the question of how to ensure an optimal long-term driving strategy, while keeping the benefits of classical trajectory planning. We introduce a Reinforcement Learning based approach that coupled with a trajectory planner, learns an optimal long-term decision-making strategy for driving on highways. By online generating locally optimal maneuvers as actions, we balance between the infinite low-level continuous action space, and the limited flexibility of a fixed number of predefined standard lane-change actions. We evaluated our method on realistic scenarios in the open-source traffic simulator SUMO and were able to achieve better performance than the 4 benchmark approaches we compared against, including a random action selecting agent, greedy agent, high-level, discrete actions agent and an IDM-based SUMO-controlled agent.

Gabriel Kalweit, Maria Huegle, Moritz Werling et al.

Popular Maximum Entropy Inverse Reinforcement Learning approaches require the computation of expected state visitation frequencies for the optimal policy under an estimate of the reward function. This usually requires intermediate value estimation in the inner loop of the algorithm, slowing down convergence considerably. In this work, we introduce a novel class of algorithms that only needs to solve the MDP underlying the demonstrated behavior once to recover the expert policy. This is possible through a formulation that exploits a probabilistic behavior assumption for the demonstrations within the structure of Q-learning. We propose Inverse Action-value Iteration which is able to fully recover an underlying reward of an external agent in closed-form analytically. We further provide an accompanying class of sampling-based variants which do not depend on a model of the environment. We show how to extend this class of algorithms to continuous state-spaces via function approximation and how to estimate a corresponding action-value function, leading to a policy as close as possible to the policy of the external agent, while optionally satisfying a list of predefined hard constraints. We evaluate the resulting algorithms called Inverse Action-value Iteration, Inverse Q-learning and Deep Inverse Q-learning on the Objectworld benchmark, showing a speedup of up to several orders of magnitude compared to (Deep) Max-Entropy algorithms. We further apply Deep Constrained Inverse Q-learning on the task of learning autonomous lane-changes in the open-source simulator SUMO achieving competent driving after training on data corresponding to 30 minutes of demonstrations.

Gabriel Kalweit, Maria Huegle, Moritz Werling et al.

In many real world applications, reinforcement learning agents have to optimize multiple objectives while following certain rules or satisfying a list of constraints. Classical methods based on reward shaping, i.e. a weighted combination of different objectives in the reward signal, or Lagrangian methods, including constraints in the loss function, have no guarantees that the agent satisfies the constraints at all points in time and can lead to undesired behavior. When a discrete policy is extracted from an action-value function, safe actions can be ensured by restricting the action space at maximization, but can lead to sub-optimal solutions among feasible alternatives. In this work, we propose Constrained Q-learning, a novel off-policy reinforcement learning framework restricting the action space directly in the Q-update to learn the optimal Q-function for the induced constrained MDP and the corresponding safe policy. In addition to single-step constraints referring only to the next action, we introduce a formulation for approximate multi-step constraints under the current target policy based on truncated value-functions. We analyze the advantages of Constrained Q-learning in the tabular case and compare Constrained DQN to reward shaping and Lagrangian methods in the application of high-level decision making in autonomous driving, considering constraints for safety, keeping right and comfort. We train our agent in the open-source simulator SUMO and on the real HighD data set.

Lukas Alexander Wilhelm Gemein, Robin Tibor Schirrmeister, Patryk Chrabąszcz et al.

Machine learning (ML) methods have the potential to automate clinical EEG analysis. They can be categorized into feature-based (with handcrafted features), and end-to-end approaches (with learned features). Previous studies on EEG pathology decoding have typically analyzed a limited number of features, decoders, or both. For a I) more elaborate feature-based EEG analysis, and II) in-depth comparisons of both approaches, here we first develop a comprehensive feature-based framework, and then compare this framework to state-of-the-art end-to-end methods. To this aim, we apply the proposed feature-based framework and deep neural networks including an EEG-optimized temporal convolutional network (TCN) to the task of pathological versus non-pathological EEG classification. For a robust comparison, we chose the Temple University Hospital (TUH) Abnormal EEG Corpus (v2.0.0), which contains approximately 3000 EEG recordings. The results demonstrate that the proposed feature-based decoding framework can achieve accuracies on the same level as state-of-the-art deep neural networks. We find accuracies across both approaches in an astonishingly narrow range from 81--86\%. Moreover, visualizations and analyses indicated that both approaches used similar aspects of the data, e.g., delta and theta band power at temporal electrode locations. We argue that the accuracies of current binary EEG pathology decoders could saturate near 90\% due to the imperfect inter-rater agreement of the clinical labels, and that such decoders are already clinically useful, such as in areas where clinical EEG experts are rare. We make the proposed feature-based framework available open source and thus offer a new tool for EEG machine learning research.

Rudolf Reiter, Jasper Hoffmann, Dirk Reinhardt et al.

The fields of MPC and RL consider two successful control techniques for Markov decision processes. Both approaches are derived from similar fundamental principles, and both are widely used in practical applications, including robotics, process control, energy systems, and autonomous driving. Despite their similarities, MPC and RL follow distinct paradigms that emerged from diverse communities and different requirements. Various technical discrepancies, particularly the role of an environment model as part of the algorithm, lead to methodologies with nearly complementary advantages. Due to their orthogonal benefits, research interest in combination methods has recently increased significantly, leading to a large and growing set of complex ideas leveraging MPC and RL. This work illuminates the differences, similarities, and fundamentals that allow for different combination algorithms and categorizes existing work accordingly. Particularly, we focus on the versatile actor-critic RL approach as a basis for our categorization and examine how the online optimization approach of MPC can be used to improve the overall closed-loop performance of a policy.

Moritz Schneider, Robert Krug, Narunas Vaskevicius et al.

Visual Reinforcement Learning (RL) methods often require extensive amounts of data. As opposed to model-free RL, model-based RL (MBRL) offers a potential solution with efficient data utilization through planning. Additionally, RL lacks generalization capabilities for real-world tasks. Prior work has shown that incorporating pre-trained visual representations (PVRs) enhances sample efficiency and generalization. While PVRs have been extensively studied in the context of model-free RL, their potential in MBRL remains largely unexplored. In this paper, we benchmark a set of PVRs on challenging control tasks in a model-based RL setting. We investigate the data efficiency, generalization capabilities, and the impact of different properties of PVRs on the performance of model-based agents. Our results, perhaps surprisingly, reveal that for MBRL current PVRs are not more sample efficient than learning representations from scratch, and that they do not generalize better to out-of-distribution (OOD) settings. To explain this, we analyze the quality of the trained dynamics model. Furthermore, we show that data diversity and network architecture are the most important contributors to OOD generalization performance.

Jan Ole von Hartz, Adrian Röfer, Joschka Boedecker et al.

We present Mixture of Discrete-time Gaussian Processes (MiDiGap), a novel approach for flexible policy representation and imitation learning in robot manipulation. MiDiGap enables learning from as few as five demonstrations using only camera observations and generalizes across a wide range of challenging tasks. It excels at long-horizon behaviors such as making coffee, highly constrained motions such as opening doors, dynamic actions such as scooping with a spatula, and multimodal tasks such as hanging a mug. MiDiGap learns these tasks on a CPU in less than a minute and scales linearly to large datasets. We also develop a rich suite of tools for inference-time steering using evidence such as collision signals and robot kinematic constraints. This steering enables novel generalization capabilities, including obstacle avoidance and cross-embodiment policy transfer. MiDiGap achieves state-of-the-art performance on diverse few-shot manipulation benchmarks. On constrained RLBench tasks, it improves policy success by 76 percentage points and reduces trajectory cost by 67%. On multimodal tasks, it improves policy success by 48 percentage points and increases sample efficiency by a factor of 20. In cross-embodiment transfer, it more than doubles policy success. We make the code publicly available at https://midigap.cs.uni-freiburg.de.

Hao Zhu, Joschka Boedecker

We consider the inverse problem of multi-armed bandits (IMAB) that are widely used in neuroscience and psychology research for behavior modelling. We first show that the IMAB problem is not convex in general, but can be relaxed to a convex problem via variable transformation. Based on this result, we propose a two-step sequential heuristic for (approximately) solving the IMAB problem. We discuss a condition where our method provides global solution to the IMAB problem with certificate, as well as approximations to further save computing time. Numerical experiments indicate that our heuristic method is more robust than directly solving the IMAB problem via repeated local optimization, and can achieve the performance of Monte Carlo methods within a significantly decreased running time. We provide the implementation of our method based on CVXPY, which allows straightforward application by users not well versed in convex optimization.

Seyed Mahdi B. Azad, Zahra Padar, Gabriel Kalweit et al.

In this paper, we propose a novel method for learning reward functions directly from offline demonstrations. Unlike traditional inverse reinforcement learning (IRL), our approach decouples the reward function from the learner's policy, eliminating the adversarial interaction typically required between the two. This results in a more stable and efficient training process. Our reward function, called \textit{SR-Reward}, leverages successor representation (SR) to encode a state based on expected future states' visitation under the demonstration policy and transition dynamics. By utilizing the Bellman equation, SR-Reward can be learned concurrently with most reinforcement learning (RL) algorithms without altering the existing training pipeline. We also introduce a negative sampling strategy to mitigate overestimation errors by reducing rewards for out-of-distribution data, thereby enhancing robustness. This strategy inherently introduces a conservative bias into RL algorithms that employ the learned reward. We evaluate our method on the D4RL benchmark, achieving competitive results compared to offline RL algorithms with access to true rewards and imitation learning (IL) techniques like behavioral cloning. Moreover, our ablation studies on data size and quality reveal the advantages and limitations of SR-Reward as a proxy for true rewards.

Mariella Dreissig, Simon Ruehle, Florian Piewak et al.

Safety-critical applications such as autonomous driving require robust 3D environment perception algorithms capable of handling diverse and ambiguous surroundings. The predictive performance of classification models is heavily influenced by the dataset and the prior knowledge provided by the annotated labels. While labels guide the learning process, they often fail to capture the inherent relationships between classes that are naturally understood by humans. We propose a training strategy for a 3D LiDAR semantic segmentation model that learns structural relationships between classes through abstraction. This is achieved by implicitly modeling these relationships using a learning rule for hierarchical multi-label classification (HMC). Our detailed analysis demonstrates that this training strategy not only improves the model's confidence calibration but also retains additional information useful for downstream tasks such as fusion, prediction, and planning.

Johannes Meyer, Jasper Hoffmann, Felix Schulz et al.

Semantic segmentation of 3D point cloud data often comes with high annotation costs. Active learning automates the process of selecting which data to annotate, reducing the total amount of annotation needed to achieve satisfactory performance. Recent approaches to active learning for 3D point clouds are often based on sophisticated heuristics for both, splitting point clouds into annotatable regions and selecting the most beneficial for further neural network training. In this work, we propose a novel and easy-to-implement strategy to separate the point cloud into annotatable regions. In our approach, we utilize a 2D grid to subdivide the point cloud into columns. To identify the next data to be annotated, we employ a network ensemble to estimate the uncertainty in the network output. We evaluate our method on the S3DIS dataset, the Toronto-3D dataset, and a large-scale urban 3D point cloud of the city of Freiburg, which we labeled in parts manually. The extensive evaluation shows that our method yields performance on par with, or even better than, complex state-of-the-art methods on all datasets. Furthermore, we provide results suggesting that in the context of point clouds the annotated area can be a more meaningful measure for active learning algorithms than the number of annotated points.

Seyed Mahdi Basiri Azad, Joschka Boedecker

Reinforcement learning (RL) in sparse-reward environments remains a significant challenge due to the lack of informative feedback. We propose a simple yet effective method that uses a small number of successful demonstrations to initialize the value function of an RL agent. By precomputing value estimates from offline demonstrations and using them as targets for early learning, our approach provides the agent with a useful prior over promising actions. The agent then refines these estimates through standard online interaction. This hybrid offline-to-online paradigm significantly reduces the exploration burden and improves sample efficiency in sparse-reward settings. Experiments on benchmark tasks demonstrate that our method accelerates convergence and outperforms standard baselines, even with minimal or suboptimal demonstration data.

Moritz Schneider, Robert Krug, Narunas Vaskevicius et al.

Empowerment, an information-theoretic measure of an agent's potential influence on its environment, has emerged as a powerful intrinsic motivation and exploration framework for reinforcement learning (RL). Besides for unsupervised RL and skill learning algorithms, the specific use of empowerment as a pre-training signal has received limited attention in the literature. We show that empowerment can be used as a pre-training signal for data-efficient downstream task adaptation. For this we extend the traditional notion of empowerment by introducing discounted empowerment, which balances the agent's control over the environment across short- and long-term horizons. Leveraging this formulation, we propose a novel pre-training paradigm that initializes policies to maximize discounted empowerment, enabling agents to acquire a robust understanding of environmental dynamics. We analyze empowerment-based pre-training for various existing RL algorithms and empirically demonstrate its potential as a general-purpose initialization strategy: empowerment-maximizing policies with long horizons are data-efficient and effective, leading to improved adaptability in downstream tasks. Our findings pave the way for future research to scale this framework to high-dimensional and complex tasks, further advancing the field of RL.

Hao Zhu, Yuan Zhang, Joschka Boedecker

We consider the inverse reinforcement learning (IRL) problem, where an unknown reward function of some Markov decision process is estimated based on observed expert demonstrations. In most existing approaches, IRL is formulated and solved as a nonconvex optimization problem, posing challenges in scenarios where robustness and reproducibility are critical. We discuss a convex formulation of the IRL problem (CIRL) initially proposed by Ng and Russel, and reformulate the problem such that the domain-specific language CVXPY can be applied directly to specify and solve the convex problem. We also extend the CIRL problem to scenarios where the expert policy is not given analytically but by trajectory as state-action pairs, which can be strongly inconsistent with optimality, by augmenting some of the constraints. Theoretical analysis and practical implementation for hyperparameter auto-selection are introduced. This note helps the users to easily apply CIRL for their problems, without background knowledge on convex optimization.

Yuan Zhang, Jasper Hoffmann, Joschka Boedecker

Learning-based techniques have become popular in both model predictive control (MPC) and reinforcement learning (RL). Probabilistic ensemble (PE) models offer a promising approach for modelling system dynamics, showcasing the ability to capture uncertainty and scalability in high-dimensional control scenarios. However, PE models are susceptible to mode collapse, resulting in non-robust control when faced with environments slightly different from the training set. In this paper, we introduce the $\textbf{u}$ncertainty-$\textbf{d}$riven rob$\textbf{u}$st $\textbf{c}$ontrol (UDUC) loss as an alternative objective for training PE models, drawing inspiration from contrastive learning. We analyze the robustness of UDUC loss through the lens of robust optimization and evaluate its performance on the challenging Real-world Reinforcement Learning (RWRL) benchmark, which involves significant environmental mismatches between the training and testing environments.

Jan Ole von Hartz, Eugenio Chisari, Tim Welschehold et al.

In policy learning for robotic manipulation, sample efficiency is of paramount importance. Thus, learning and extracting more compact representations from camera observations is a promising avenue. However, current methods often assume full observability of the scene and struggle with scale invariance. In many tasks and settings, this assumption does not hold as objects in the scene are often occluded or lie outside the field of view of the camera, rendering the camera observation ambiguous with regard to their location. To tackle this problem, we present BASK, a Bayesian approach to tracking scale-invariant keypoints over time. Our approach successfully resolves inherent ambiguities in images, enabling keypoint tracking on symmetrical objects and occluded and out-of-view objects. We employ our method to learn challenging multi-object robot manipulation tasks from wrist camera observations and demonstrate superior utility for policy learning compared to other representation learning techniques. Furthermore, we show outstanding robustness towards disturbances such as clutter, occlusions, and noisy depth measurements, as well as generalization to unseen objects both in simulation and real-world robotic experiments.

Nicolai Dorka, Tim Welschehold, Joschka Boedecker et al.

Accurate value estimates are important for off-policy reinforcement learning. Algorithms based on temporal difference learning typically are prone to an over- or underestimation bias building up over time. In this paper, we propose a general method called Adaptively Calibrated Critics (ACC) that uses the most recent high variance but unbiased on-policy rollouts to alleviate the bias of the low variance temporal difference targets. We apply ACC to Truncated Quantile Critics, which is an algorithm for continuous control that allows regulation of the bias with a hyperparameter tuned per environment. The resulting algorithm adaptively adjusts the parameter during training rendering hyperparameter search unnecessary and sets a new state of the art on the OpenAI gym continuous control benchmark among all algorithms that do not tune hyperparameters for each environment. ACC further achieves improved results on different tasks from the Meta-World robot benchmark. Additionally, we demonstrate the generality of ACC by applying it to TD3 and showing an improved performance also in this setting.

Eugenio Chisari, Tim Welschehold, Joschka Boedecker et al.

Learning to solve complex manipulation tasks from visual observations is a dominant challenge for real-world robot learning. Although deep reinforcement learning algorithms have recently demonstrated impressive results in this context, they still require an impractical amount of time-consuming trial-and-error iterations. In this work, we consider the promising alternative paradigm of interactive learning in which a human teacher provides feedback to the policy during execution, as opposed to imitation learning where a pre-collected dataset of perfect demonstrations is used. Our proposed CEILing (Corrective and Evaluative Interactive Learning) framework combines both corrective and evaluative feedback from the teacher to train a stochastic policy in an asynchronous manner, and employs a dedicated mechanism to trade off human corrections with the robot's own experience. We present results obtained with our framework in extensive simulation and real-world experiments to demonstrate that CEILing can effectively solve complex robot manipulation tasks directly from raw images in less than one hour of real-world training.

Alireza Ranjbar, Ngo Anh Vien, Hanna Ziesche et al.

While classic control theory offers state of the art solutions in many problem scenarios, it is often desired to improve beyond the structure of such solutions and surpass their limitations. To this end, residual policy learning (RPL) offers a formulation to improve existing controllers with reinforcement learning (RL) by learning an additive "residual" to the output of a given controller. However, the applicability of such an approach highly depends on the structure of the controller. Often, internal feedback signals of the controller limit an RL algorithm to adequately change the policy and, hence, learn the task. We propose a new formulation that addresses these limitations by also modifying the feedback signals to the controller with an RL policy and show superior performance of our approach on a contact-rich peg-insertion task under position and orientation uncertainty. In addition, we use a recent Cartesian impedance control architecture as the control framework which can be available to us as a black-box while assuming no knowledge about its input/output structure, and show the difficulties of standard RPL. Furthermore, we introduce an adaptive curriculum for the given task to gradually increase the task difficulty in terms of position and orientation uncertainty. A video showing the results can be found at https://youtu.be/SAZm_Krze7U .

Maria Kalweit, Gabriel Kalweit, Moritz Werling et al.

Challenging problems of deep reinforcement learning systems with regard to the application on real systems are their adaptivity to changing environments and their efficiency w.r.t. computational resources and data. In the application of learning lane-change behavior for autonomous driving, agents have to deal with a varying number of surrounding vehicles. Furthermore, the number of required transitions imposes a bottleneck, since test drivers cannot perform an arbitrary amount of lane changes in the real world. In the off-policy setting, additional information on solving the task can be gained by observing actions from others. While in the classical RL setup this knowledge remains unused, we use other drivers as surrogates to learn the agent's value function more efficiently. We propose Surrogate Q-learning that deals with the aforementioned problems and reduces the required driving time drastically. We further propose an efficient implementation based on a permutation-equivariant deep neural network architecture of the Q-function to estimate action-values for a variable number of vehicles in sensor range. We show that the architecture leads to a novel replay sampling technique we call Scene-centric Experience Replay and evaluate the performance of Surrogate Q-learning and Scene-centric Experience Replay in the open traffic simulator SUMO. Additionally, we show that our methods enhance real-world applicability of RL systems by learning policies on the real highD dataset.

Maria Hügle, Gabriel Kalweit, Thomas Huegle et al.

Clinical data from electronic medical records, registries or trials provide a large source of information to apply machine learning methods in order to foster precision medicine, e.g. by finding new disease phenotypes or performing individual disease prediction. However, to take full advantage of deep learning methods on clinical data, architectures are necessary that 1) are robust with respect to missing and wrong values, and 2) can deal with highly variable-sized lists and long-term dependencies of individual diagnosis, procedures, measurements and medication prescriptions. In this work, we elaborate limitations of fully-connected neural networks and classical machine learning methods in this context and propose AdaptiveNet, a novel recurrent neural network architecture, which can deal with multiple lists of different events, alleviating the aforementioned limitations. We employ the architecture to the problem of disease progression prediction in rheumatoid arthritis using the Swiss Clinical Quality Management registry, which contains over 10.000 patients and more than 65.000 patient visits. Our proposed approach leads to more compact representations and outperforms the classical baselines.

Maria Huegle, Gabriel Kalweit, Moritz Werling et al.

The common pipeline in autonomous driving systems is highly modular and includes a perception component which extracts lists of surrounding objects and passes these lists to a high-level decision component. In this case, leveraging the benefits of deep reinforcement learning for high-level decision making requires special architectures to deal with multiple variable-length sequences of different object types, such as vehicles, lanes or traffic signs. At the same time, the architecture has to be able to cover interactions between traffic participants in order to find the optimal action to be taken. In this work, we propose the novel Deep Scenes architecture, that can learn complex interaction-aware scene representations based on extensions of either 1) Deep Sets or 2) Graph Convolutional Networks. We present the Graph-Q and DeepScene-Q off-policy reinforcement learning algorithms, both outperforming state-of-the-art methods in evaluations with the publicly available traffic simulator SUMO.

Gabriel Kalweit, Maria Huegle, Joschka Boedecker

In the past few years, off-policy reinforcement learning methods have shown promising results in their application for robot control. Deep Q-learning, however, still suffers from poor data-efficiency and is susceptible to stochasticity in the environment or reward functions which is limiting with regard to real-world applications. We alleviate these problems by proposing two novel off-policy Temporal-Difference formulations: (1) Truncated Q-functions which represent the return for the first n steps of a target-policy rollout w.r.t. the full action-value and (2) Shifted Q-functions, acting as the farsighted return after this truncated rollout. This decomposition allows us to optimize both parts with their individual learning rates, achieving significant learning speedup. We prove that the combination of these short- and long-term predictions is a representation of the full return, leading to the Composite Q-learning algorithm. We show the efficacy of Composite Q-learning in the tabular case and compare Deep Composite Q-learning with TD3 and TD3(Delta), which we introduce as an off-policy variant of TD(Delta). Moreover, we show that Composite TD3 outperforms TD3 as well as state-of-the-art compositional Q-learning approaches significantly in terms of data-efficiency in multiple simulated robot tasks and that Composite Q-learning is robust to stochastic environments and reward functions.