Gerhard Neumann

Markus Walker, Marcel Reith-Braun, Tai Hoang et al.

Sampling-based model predictive control (MPC) is effective for nonlinear systems but often produces non-smooth control inputs due to random sampling. To address this issue, we extend the model predictive path integral (MPPI) framework with deterministic sampling and improvements from cross-entropy method (CEM)--MPC, such as iterative optimization, proposing deterministic sampling MPPI (dsMPPI). This combination leverages the exponential weighting of MPPI alongside the efficiency of deterministic samples. Experiments demonstrate that dsMPPI achieves smoother trajectories compared to state-of-the-art methods.

Maximilian Xiling Li, Onur Celik, Philipp Becker et al.

Learning skills by imitation is a promising concept for the intuitive teaching of robots. A common way to learn such skills is to learn a parametric model by maximizing the likelihood given the demonstrations. Yet, human demonstrations are often multi-modal, i.e., the same task is solved in multiple ways which is a major challenge for most imitation learning methods that are based on such a maximum likelihood (ML) objective. The ML objective forces the model to cover all data, it prevents specialization in the context space and can cause mode-averaging in the behavior space, leading to suboptimal or potentially catastrophic behavior. Here, we alleviate those issues by introducing a curriculum using a weight for each data point, allowing the model to specialize on data it can represent while incentivizing it to cover as much data as possible by an entropy bonus. We extend our algorithm to a Mixture of (linear) Experts (MoE) such that the single components can specialize on local context regions, while the MoE covers all data points. We evaluate our approach in complex simulated and real robot control tasks and show it learns from versatile human demonstrations and significantly outperforms current SOTA methods. A reference implementation can be found at https://github.com/intuitive-robots/ml-cur

Ivan Saraev, Elena Erben, Weida Liao et al.

Light-based advanced manufacturing increasingly requires programmable, closed-loop tools that translate human design intent into executable operations at small length scales. Yet a key bottleneck persists across robotic and manufacturing modalities: turning user intent into machine-readable objectives that are reliably executable. While micro-robotics offers versatile manipulation via optical actuation of fluids, mathematically tractable goal specification remains manual and hard to reuse. Here, we introduce Speak-to-Objective, a modular agentic pipeline that uses a conditioned Large Language Model (LLM) to translate spoken or written commands into fully differentiable objective functions for assembling microparticles in a constraint-aware inverse solver (SLSQP) and on an experimental optofluidic platform. The approach employs a compact loop - perceive -> compose -> propose -> act -> report & learn - that treats the objective as the interface between intent and actuation, separating what to assemble or pattern from how to actuate, while learning from user feedback. The pipeline composes geometry, spacing, and assignment/topology terms to generate robust descriptive objectives that assemble from partial traces and recover after perturbations, as well as explicit objectives for precise placement, all in an actuator-agnostic fashion. Using laser-induced thermoviscous flows as the physical actuation modality, we demonstrate natural-language-programmable, light-based microscale assembly of particle patterns in a microfluidic environment. Beyond its immediate impact on programmable microassembly, and using laser-induced optofluidic actuation as a reduced-complexity experimental platform, our work points toward self-driving, AI-assisted optical manufacturing platforms in which natural language, differentiable objectives, and laser-based actuation are coupled into a reusable digital workflow.

Yumeng Li, Ning Gao, Hanna Ziesche et al. · amazon-science

We present a novel meta-learning approach for 6D pose estimation on unknown objects. In contrast to ``instance-level" and ``category-level" pose estimation methods, our algorithm learns object representation in a category-agnostic way, which endows it with strong generalization capabilities across object categories. Specifically, we employ a neural process-based meta-learning approach to train an encoder to capture texture and geometry of an object in a latent representation, based on very few RGB-D images and ground-truth keypoints. The latent representation is then used by a simultaneously meta-trained decoder to predict the 6D pose of the object in new images. Furthermore, we propose a novel geometry-aware decoder for the keypoint prediction using a Graph Neural Network (GNN), which explicitly takes geometric constraints specific to each object into consideration. To evaluate our algorithm, extensive experiments are conducted on the \linemod dataset, and on our new fully-annotated synthetic datasets generated from Multiple Categories in Multiple Scenes (MCMS). Experimental results demonstrate that our model performs well on unseen objects with very different shapes and appearances. Remarkably, our model also shows robust performance on occluded scenes although trained fully on data without occlusion. To our knowledge, this is the first work exploring \textbf{cross-category level} 6D pose estimation.

Ning Gao, Hanna Ziesche, Ngo Anh Vien et al.

Meta-learning is widely used in few-shot classification and function regression due to its ability to quickly adapt to unseen tasks. However, it has not yet been well explored on regression tasks with high dimensional inputs such as images. This paper makes two main contributions that help understand this barely explored area. \emph{First}, we design two new types of cross-category level vision regression tasks, namely object discovery and pose estimation of unprecedented complexity in the meta-learning domain for computer vision. To this end, we (i) exhaustively evaluate common meta-learning techniques on these tasks, and (ii) quantitatively analyze the effect of various deep learning techniques commonly used in recent meta-learning algorithms in order to strengthen the generalization capability: data augmentation, domain randomization, task augmentation and meta-regularization. Finally, we (iii) provide some insights and practical recommendations for training meta-learning algorithms on vision regression tasks. \emph{Second}, we propose the addition of functional contrastive learning (FCL) over the task representations in Conditional Neural Processes (CNPs) and train in an end-to-end fashion. The experimental results show that the results of prior work are misleading as a consequence of a poor choice of the loss function as well as too small meta-training sets. Specifically, we find that CNPs outperform MAML on most tasks without fine-tuning. Furthermore, we observe that naive task augmentation without a tailored design results in underfitting.

Florian Seligmann, Philipp Becker, Michael Volpp et al.

Bayesian deep learning (BDL) is a promising approach to achieve well-calibrated predictions on distribution-shifted data. Nevertheless, there exists no large-scale survey that evaluates recent SOTA methods on diverse, realistic, and challenging benchmark tasks in a systematic manner. To provide a clear picture of the current state of BDL research, we evaluate modern BDL algorithms on real-world datasets from the WILDS collection containing challenging classification and regression tasks, with a focus on generalization capability and calibration under distribution shift. We compare the algorithms on a wide range of large, convolutional and transformer-based neural network architectures. In particular, we investigate a signed version of the expected calibration error that reveals whether the methods are over- or under-confident, providing further insight into the behavior of the methods. Further, we provide the first systematic evaluation of BDL for fine-tuning large pre-trained models, where training from scratch is prohibitively expensive. Finally, given the recent success of Deep Ensembles, we extend popular single-mode posterior approximations to multiple modes by the use of ensembles. While we find that ensembling single-mode approximations generally improves the generalization capability and calibration of the models by a significant margin, we also identify a failure mode of ensembles when finetuning large transformer-based language models. In this setting, variational inference based approaches such as last-layer Bayes By Backprop outperform other methods in terms of accuracy by a large margin, while modern approximate inference algorithms such as SWAG achieve the best calibration.

Fabian Otto, Onur Celik, Hongyi Zhou et al.

\Episode-based reinforcement learning (ERL) algorithms treat reinforcement learning (RL) as a black-box optimization problem where we learn to select a parameter vector of a controller, often represented as a movement primitive, for a given task descriptor called a context. ERL offers several distinct benefits in comparison to step-based RL. It generates smooth control trajectories, can handle non-Markovian reward definitions, and the resulting exploration in parameter space is well suited for solving sparse reward settings. Yet, the high dimensionality of the movement primitive parameters has so far hampered the effective use of deep RL methods. In this paper, we present a new algorithm for deep ERL. It is based on differentiable trust region layers, a successful on-policy deep RL algorithm. These layers allow us to specify trust regions for the policy update that are solved exactly for each state using convex optimization, which enables policies learning with the high precision required for the ERL. We compare our ERL algorithm to state-of-the-art step-based algorithms in many complex simulated robotic control tasks. In doing so, we investigate different reward formulations - dense, sparse, and non-Markovian. While step-based algorithms perform well only on dense rewards, ERL performs favorably on sparse and non-Markovian rewards. Moreover, our results show that the sparse and the non-Markovian rewards are also often better suited to define the desired behavior, allowing us to obtain considerably higher quality policies compared to step-based RL.

Niklas Freymuth, Philipp Dahlinger, Tobias Würth et al.

Adaptive Mesh Refinement (AMR) enhances the Finite Element Method, an important technique for simulating complex problems in engineering, by dynamically refining mesh regions, enabling a favorable trade-off between computational speed and simulation accuracy. Classical methods for AMR depend on heuristics or expensive error estimators, hindering their use for complex simulations. Recent learning-based AMR methods tackle these issues, but so far scale only to simple toy examples. We formulate AMR as a novel Adaptive Swarm Markov Decision Process in which a mesh is modeled as a system of simple collaborating agents that may split into multiple new agents. This framework allows for a spatial reward formulation that simplifies the credit assignment problem, which we combine with Message Passing Networks to propagate information between neighboring mesh elements. We experimentally validate our approach, Adaptive Swarm Mesh Refinement (ASMR), on challenging refinement tasks. Our approach learns reliable and efficient refinement strategies that can robustly generalize to different domains during inference. Additionally, it achieves a speedup of up to $2$ orders of magnitude compared to uniform refinements in more demanding simulations. We outperform learned baselines and heuristics, achieving a refinement quality that is on par with costly error-based oracle AMR strategies.

Oleg Arenz, Philipp Dahlinger, Zihan Ye et al.

Variational inference with Gaussian mixture models (GMMs) enables learning of highly tractable yet multi-modal approximations of intractable target distributions with up to a few hundred dimensions. The two currently most effective methods for GMM-based variational inference, VIPS and iBayes-GMM, both employ independent natural gradient updates for the individual components and their weights. We show for the first time, that their derived updates are equivalent, although their practical implementations and theoretical guarantees differ. We identify several design choices that distinguish both approaches, namely with respect to sample selection, natural gradient estimation, stepsize adaptation, and whether trust regions are enforced or the number of components adapted. We argue that for both approaches, the quality of the learned approximations can heavily suffer from the respective design choices: By updating the individual components using samples from the mixture model, iBayes-GMM often fails to produce meaningful updates to low-weight components, and by using a zero-order method for estimating the natural gradient, VIPS scales badly to higher-dimensional problems. Furthermore, we show that information-geometric trust-regions (used by VIPS) are effective even when using first-order natural gradient estimates, and often outperform the improved Bayesian learning rule (iBLR) update used by iBayes-GMM. We systematically evaluate the effects of design choices and show that a hybrid approach significantly outperforms both prior works. Along with this work, we publish our highly modular and efficient implementation for natural gradient variational inference with Gaussian mixture models, which supports 432 different combinations of design choices, facilitates the reproduction of all our experiments, and may prove valuable for the practitioner.

Fabian Duffhauss, Tobias Demmler, Gerhard Neumann

Estimating 6D poses of objects is an essential computer vision task. However, most conventional approaches rely on camera data from a single perspective and therefore suffer from occlusions. We overcome this issue with our novel multi-view 6D pose estimation method called MV6D which accurately predicts the 6D poses of all objects in a cluttered scene based on RGB-D images from multiple perspectives. We base our approach on the PVN3D network that uses a single RGB-D image to predict keypoints of the target objects. We extend this approach by using a combined point cloud from multiple views and fusing the images from each view with a DenseFusion layer. In contrast to current multi-view pose detection networks such as CosyPose, our MV6D can learn the fusion of multiple perspectives in an end-to-end manner and does not require multiple prediction stages or subsequent fine tuning of the prediction. Furthermore, we present three novel photorealistic datasets of cluttered scenes with heavy occlusions. All of them contain RGB-D images from multiple perspectives and the ground truth for instance semantic segmentation and 6D pose estimation. MV6D significantly outperforms the state-of-the-art in multi-view 6D pose estimation even in cases where the camera poses are known inaccurately. Furthermore, we show that our approach is robust towards dynamic camera setups and that its accuracy increases incrementally with an increasing number of perspectives.

Hadi Beik-Mohammadi, Søren Hauberg, Georgios Arvanitidis et al.

In recent decades, advancements in motion learning have enabled robots to acquire new skills and adapt to unseen conditions in both structured and unstructured environments. In practice, motion learning methods capture relevant patterns and adjust them to new conditions such as dynamic obstacle avoidance or variable targets. In this paper, we investigate the robot motion learning paradigm from a Riemannian manifold perspective. We argue that Riemannian manifolds may be learned via human demonstrations in which geodesics are natural motion skills. The geodesics are generated using a learned Riemannian metric produced by our novel variational autoencoder (VAE), which is especially intended to recover full-pose end-effector states and joint space configurations. In addition, we propose a technique for facilitating on-the-fly end-effector/multiple-limb obstacle avoidance by reshaping the learned manifold using an obstacle-aware ambient metric. The motion generated using these geodesics may naturally result in multiple-solution tasks that have not been explicitly demonstrated previously. We extensively tested our approach in task space and joint space scenarios using a 7-DoF robotic manipulator. We demonstrate that our method is capable of learning and generating motion skills based on complicated motion patterns demonstrated by a human operator. Additionally, we assess several obstacle avoidance strategies and generate trajectories in multiple-mode settings.

Jonas Linkerhägner, Niklas Freymuth, Paul Maria Scheikl et al.

Physical simulations that accurately model reality are crucial for many engineering disciplines such as mechanical engineering and robotic motion planning. In recent years, learned Graph Network Simulators produced accurate mesh-based simulations while requiring only a fraction of the computational cost of traditional simulators. Yet, the resulting predictors are confined to learning from data generated by existing mesh-based simulators and thus cannot include real world sensory information such as point cloud data. As these predictors have to simulate complex physical systems from only an initial state, they exhibit a high error accumulation for long-term predictions. In this work, we integrate sensory information to ground Graph Network Simulators on real world observations. In particular, we predict the mesh state of deformable objects by utilizing point cloud data. The resulting model allows for accurate predictions over longer time horizons, even under uncertainties in the simulation, such as unknown material properties. Since point clouds are usually not available for every time step, especially in online settings, we employ an imputation-based model. The model can make use of such additional information only when provided, and resorts to a standard Graph Network Simulator, otherwise. We experimentally validate our approach on a suite of prediction tasks for mesh-based interactions between soft and rigid bodies. Our method results in utilization of additional point cloud information to accurately predict stable simulations where existing Graph Network Simulators fail.

Fabian Duffhauss, Sebastian Koch, Hanna Ziesche et al.

Detecting objects and estimating their 6D poses is essential for automated systems to interact safely with the environment. Most 6D pose estimators, however, rely on a single camera frame and suffer from occlusions and ambiguities due to object symmetries. We overcome this issue by presenting a novel symmetry-aware multi-view 6D pose estimator called SyMFM6D. Our approach efficiently fuses the RGB-D frames from multiple perspectives in a deep multi-directional fusion network and predicts predefined keypoints for all objects in the scene simultaneously. Based on the keypoints and an instance semantic segmentation, we efficiently compute the 6D poses by least-squares fitting. To address the ambiguity issues for symmetric objects, we propose a novel training procedure for symmetry-aware keypoint detection including a new objective function. Our SyMFM6D network significantly outperforms the state-of-the-art in both single-view and multi-view 6D pose estimation. We furthermore show the effectiveness of our symmetry-aware training procedure and demonstrate that our approach is robust towards inaccurate camera calibration and dynamic camera setups.

Philipp Becker, Gerhard Neumann

Improved state space models, such as Recurrent State Space Models (RSSMs), are a key factor behind recent advances in model-based reinforcement learning (RL). Yet, despite their empirical success, many of the underlying design choices are not well understood. We show that RSSMs use a suboptimal inference scheme and that models trained using this inference overestimate the aleatoric uncertainty of the ground truth system. We find this overestimation implicitly regularizes RSSMs and allows them to succeed in model-based RL. We postulate that this implicit regularization fulfills the same functionality as explicitly modeling epistemic uncertainty, which is crucial for many other model-based RL approaches. Yet, overestimating aleatoric uncertainty can also impair performance in cases where accurately estimating it matters, e.g., when we have to deal with occlusions, missing observations, or fusing sensor modalities at different frequencies. Moreover, the implicit regularization is a side-effect of the inference scheme and not the result of a rigorous, principled formulation, which renders analyzing or improving RSSMs difficult. Thus, we propose an alternative approach building on well-understood components for modeling aleatoric and epistemic uncertainty, dubbed Variational Recurrent Kalman Network (VRKN). This approach uses Kalman updates for exact smoothing inference in a latent space and Monte Carlo Dropout to model epistemic uncertainty. Due to the Kalman updates, the VRKN can naturally handle missing observations or sensor fusion problems with varying numbers of observations per time step. Our experiments show that using the VRKN instead of the RSSM improves performance in tasks where appropriately capturing aleatoric uncertainty is crucial while matching it in the deterministic standard benchmarks.

Fabian Duffhauss, Ngo Anh Vien, Hanna Ziesche et al.

Sensor fusion can significantly improve the performance of many computer vision tasks. However, traditional fusion approaches are either not data-driven and cannot exploit prior knowledge nor find regularities in a given dataset or they are restricted to a single application. We overcome this shortcoming by presenting a novel deep hierarchical variational autoencoder called FusionVAE that can serve as a basis for many fusion tasks. Our approach is able to generate diverse image samples that are conditioned on multiple noisy, occluded, or only partially visible input images. We derive and optimize a variational lower bound for the conditional log-likelihood of FusionVAE. In order to assess the fusion capabilities of our model thoroughly, we created three novel datasets for image fusion based on popular computer vision datasets. In our experiments, we show that FusionVAE learns a representation of aggregated information that is relevant to fusion tasks. The results demonstrate that our approach outperforms traditional methods significantly. Furthermore, we present the advantages and disadvantages of different design choices.

Ning Gao, Jingyu Zhang, Ruijie Chen et al.

Grasping inhomogeneous objects in real-world applications remains a challenging task due to the unknown physical properties such as mass distribution and coefficient of friction. In this study, we propose a meta-learning algorithm called ConDex, which incorporates Conditional Neural Processes (CNP) with DexNet-2.0 to autonomously discern the underlying physical properties of objects using depth images. ConDex efficiently acquires physical embeddings from limited trials, enabling precise grasping point estimation. Furthermore, ConDex is capable of updating the predicted grasping quality iteratively from new trials in an online fashion. To the best of our knowledge, we are the first who generate two object datasets focusing on inhomogeneous physical properties with varying mass distributions and friction coefficients. Extensive evaluations in simulation demonstrate ConDex's superior performance over DexNet-2.0 and existing meta-learning-based grasping pipelines. Furthermore, ConDex shows robust generalization to previously unseen real-world objects despite training solely in the simulation. The synthetic and real-world datasets will be published as well.

Fabian Otto, Hongyi Zhou, Onur Celik et al.

We introduce a novel deep reinforcement learning (RL) approach called Movement Primitive-based Planning Policy (MP3). By integrating movement primitives (MPs) into the deep RL framework, MP3 enables the generation of smooth trajectories throughout the whole learning process while effectively learning from sparse and non-Markovian rewards. Additionally, MP3 maintains the capability to adapt to changes in the environment during execution. Although many early successes in robot RL have been achieved by combining RL with MPs, these approaches are often limited to learning single stroke-based motions, lacking the ability to adapt to task variations or adjust motions during execution. Building upon our previous work, which introduced an episode-based RL method for the non-linear adaptation of MP parameters to different task variations, this paper extends the approach to incorporating replanning strategies. This allows adaptation of the MP parameters throughout motion execution, addressing the lack of online motion adaptation in stochastic domains requiring feedback. We compared our approach against state-of-the-art deep RL and RL with MPs methods. The results demonstrated improved performance in sophisticated, sparse reward settings and in domains requiring replanning.

Ge Li, Zeqi Jin, Michael Volpp et al.

Movement Primitives (MPs) are a well-known concept to represent and generate modular trajectories. MPs can be broadly categorized into two types: (a) dynamics-based approaches that generate smooth trajectories from any initial state, e. g., Dynamic Movement Primitives (DMPs), and (b) probabilistic approaches that capture higher-order statistics of the motion, e. g., Probabilistic Movement Primitives (ProMPs). To date, however, there is no method that unifies both, i. e. that can generate smooth trajectories from an arbitrary initial state while capturing higher-order statistics. In this paper, we introduce a unified perspective of both approaches by solving the ODE underlying the DMPs. We convert expensive online numerical integration of DMPs into basis functions that can be computed offline. These basis functions can be used to represent trajectories or trajectory distributions similar to ProMPs while maintaining all the properties of dynamical systems. Since we inherit the properties of both methodologies, we call our proposed model Probabilistic Dynamic Movement Primitives (ProDMPs). Additionally, we embed ProDMPs in deep neural network architecture and propose a new cost function for efficient end-to-end learning of higher-order trajectory statistics. To this end, we leverage Bayesian Aggregation for non-linear iterative conditioning on sensory inputs. Our proposed model achieves smooth trajectory generation, goal-attractor convergence, correlation analysis, non-linear conditioning, and online re-planing in one framework.

Vaisakh Shaj, Saleh Gholam Zadeh, Ozan Demir et al.

Intelligent agents use internal world models to reason and make predictions about different courses of their actions at many scales. Devising learning paradigms and architectures that allow machines to learn world models that operate at multiple levels of temporal abstractions while dealing with complex uncertainty predictions is a major technical hurdle. In this work, we propose a probabilistic formalism to learn multi-time scale world models which we call the Multi Time Scale State Space (MTS3) model. Our model uses a computationally efficient inference scheme on multiple time scales for highly accurate long-horizon predictions and uncertainty estimates over several seconds into the future. Our experiments, which focus on action conditional long horizon future predictions, show that MTS3 outperforms recent methods on several system identification benchmarks including complex simulated and real-world dynamical systems. Code is available at this repository: https://github.com/ALRhub/MTS3.

Ning Gao, Ngo Anh Vien, Hanna Ziesche et al.

To enable meaningful robotic manipulation of objects in the real-world, 6D pose estimation is one of the critical aspects. Most existing approaches have difficulties to extend predictions to scenarios where novel object instances are continuously introduced, especially with heavy occlusions. In this work, we propose a few-shot pose estimation (FSPE) approach called SA6D, which uses a self-adaptive segmentation module to identify the novel target object and construct a point cloud model of the target object using only a small number of cluttered reference images. Unlike existing methods, SA6D does not require object-centric reference images or any additional object information, making it a more generalizable and scalable solution across categories. We evaluate SA6D on real-world tabletop object datasets and demonstrate that SA6D outperforms existing FSPE methods, particularly in cluttered scenes with occlusions, while requiring fewer reference images.

Philipp Becker, Sebastian Mossburger, Fabian Otto et al.

Learning self-supervised representations using reconstruction or contrastive losses improves performance and sample complexity of image-based and multimodal reinforcement learning (RL). Here, different self-supervised loss functions have distinct advantages and limitations depending on the information density of the underlying sensor modality. Reconstruction provides strong learning signals but is susceptible to distractions and spurious information. While contrastive approaches can ignore those, they may fail to capture all relevant details and can lead to representation collapse. For multimodal RL, this suggests that different modalities should be treated differently based on the amount of distractions in the signal. We propose Contrastive Reconstructive Aggregated representation Learning (CoRAL), a unified framework enabling us to choose the most appropriate self-supervised loss for each sensor modality and allowing the representation to better focus on relevant aspects. We evaluate CoRAL's benefits on a wide range of tasks with images containing distractions or occlusions, a new locomotion suite, and a challenging manipulation suite with visually realistic distractions. Our results show that learning a multimodal representation by combining contrastive and reconstruction-based losses can significantly improve performance and solve tasks that are out of reach for more naive representation learning approaches and other recent baselines.

Niklas Freymuth, Nicolas Schreiber, Philipp Becker et al.

Humans intuitively solve tasks in versatile ways, varying their behavior in terms of trajectory-based planning and for individual steps. Thus, they can easily generalize and adapt to new and changing environments. Current Imitation Learning algorithms often only consider unimodal expert demonstrations and act in a state-action-based setting, making it difficult for them to imitate human behavior in case of versatile demonstrations. Instead, we combine a mixture of movement primitives with a distribution matching objective to learn versatile behaviors that match the expert's behavior and versatility. To facilitate generalization to novel task configurations, we do not directly match the agent's and expert's trajectory distributions but rather work with concise geometric descriptors which generalize well to unseen task configurations. We empirically validate our method on various robot tasks using versatile human demonstrations and compare to imitation learning algorithms in a state-action setting as well as a trajectory-based setting. We find that the geometric descriptors greatly help in generalizing to new task configurations and that combining them with our distribution-matching objective is crucial for representing and reproducing versatile behavior.

Vaisakh Shaj, Dieter Buchler, Rohit Sonker et al.

Recurrent State-space models (RSSMs) are highly expressive models for learning patterns in time series data and system identification. However, these models assume that the dynamics are fixed and unchanging, which is rarely the case in real-world scenarios. Many control applications often exhibit tasks with similar but not identical dynamics which can be modeled as a latent variable. We introduce the Hidden Parameter Recurrent State Space Models (HiP-RSSMs), a framework that parametrizes a family of related dynamical systems with a low-dimensional set of latent factors. We present a simple and effective way of learning and performing inference over this Gaussian graphical model that avoids approximations like variational inference. We show that HiP-RSSMs outperforms RSSMs and competing multi-task models on several challenging robotic benchmarks both on real-world systems and simulations.

Moritz Reuss, Niels van Duijkeren, Robert Krug et al.

It is well-known that inverse dynamics models can improve tracking performance in robot control. These models need to precisely capture the robot dynamics, which consist of well-understood components, e.g., rigid body dynamics, and effects that remain challenging to capture, e.g., stick-slip friction and mechanical flexibilities. Such effects exhibit hysteresis and partial observability, rendering them, particularly challenging to model. Hence, hybrid models, which combine a physical prior with data-driven approaches are especially well-suited in this setting. We present a novel hybrid model formulation that enables us to identify fully physically consistent inertial parameters of a rigid body dynamics model which is paired with a recurrent neural network architecture, allowing us to capture unmodeled partially observable effects using the network memory. We compare our approach against state-of-the-art inverse dynamics models on a 7 degree of freedom manipulator. Using data sets obtained through an optimal experiment design approach, we study the accuracy of offline torque prediction and generalization capabilities of joint learning methods. In control experiments on the real system, we evaluate the model as a feed-forward term for impedance control and show the feedback gains can be drastically reduced to achieve a given tracking accuracy.

Abdalkarim Mohtasib, Gerhard Neumann, Heriberto Cuayahuitl

Learning robotic tasks in the real world is still highly challenging and effective practical solutions remain to be found. Traditional methods used in this area are imitation learning and reinforcement learning, but they both have limitations when applied to real robots. Combining reinforcement learning with pre-collected demonstrations is a promising approach that can help in learning control policies to solve robotic tasks. In this paper, we propose an algorithm that uses novel techniques to leverage offline expert data using offline and online training to obtain faster convergence and improved performance. The proposed algorithm (AWET) weights the critic losses with a novel agent advantage weight to improve over the expert data. In addition, AWET makes use of an automatic early termination technique to stop and discard policy rollouts that are not similar to expert trajectories -- to prevent drifting far from the expert data. In an ablation study, AWET showed improved and promising performance when compared to state-of-the-art baselines on four standard robotic tasks.

Ning Gao, Bernard Hohmann, Gerhard Neumann

Object-centric representations using slots have shown the advances towards efficient, flexible and interpretable abstraction from low-level perceptual features in a compositional scene. Current approaches randomize the initial state of slots followed by an iterative refinement. As we show in this paper, the random slot initialization significantly affects the accuracy of the final slot prediction. Moreover, current approaches require a predetermined number of slots from prior knowledge of the data, which limits the applicability in the real world. In our work, we initialize the slot representations with clustering algorithms conditioned on the perceptual input features. This requires an additional layer in the architecture to initialize the slots given the identified clusters. We design permutation invariant and permutation equivariant versions of this layer to enable the exchangeable slot representations after clustering. Additionally, we employ mean-shift clustering to automatically identify the number of slots for a given scene. We evaluate our method on object discovery and novel view synthesis tasks with various datasets. The results show that our method outperforms prior works consistently, especially for complex scenes.

Philipp Blättner, Johannes Brand, Gerhard Neumann et al.

Robotic grasping is a fundamental skill required for object manipulation in robotics. Multi-fingered robotic hands, which mimic the structure of the human hand, can potentially perform complex object manipulation. Nevertheless, current techniques for multi-fingered robotic grasping frequently predict only a single grasp for each inference time, limiting computational efficiency and their versatility, i.e. unimodal grasp distribution. This paper proposes a differentiable multi-fingered grasp generation network (DMFC-GraspNet) with three main contributions to address this challenge. Firstly, a novel neural grasp planner is proposed, which predicts a new grasp representation to enable versatile and dense grasp predictions. Secondly, a scene creation and label mapping method is developed for dense labeling of multi-fingered robotic hands, which allows a dense association of ground truth grasps. Thirdly, we propose to train DMFC-GraspNet end-to-end using using a forward-backward automatic differentiation approach with both a supervised loss and a differentiable collision loss and a generalized Q 1 grasp metric loss. The proposed approach is evaluated using the Shadow Dexterous Hand on Mujoco simulation and ablated by different choices of loss functions. The results demonstrate the effectiveness of the proposed approach in predicting versatile and dense grasps, and in advancing the field of multi-fingered robotic grasping.

Maximilian Hüttenrauch, Gerhard Neumann

Most successful stochastic black-box optimizers, such as CMA-ES, use rankings of the individual samples to obtain a new search distribution. Yet, the use of rankings also introduces several issues such as the underlying optimization objective is often unclear, i.e., we do not optimize the expected fitness. Further, while these algorithms typically produce a high-quality mean estimate of the search distribution, the produced samples can have poor quality as these algorithms are ignorant of the regret. Lastly, noisy fitness function evaluations may result in solutions that are highly sub-optimal on expectation. In contrast, stochastic optimizers that are motivated by policy gradients, such as the Model-based Relative Entropy Stochastic Search (MORE) algorithm, directly optimize the expected fitness function without the use of rankings. MORE can be derived by applying natural policy gradients and compatible function approximation, and is using information theoretic constraints to ensure the stability of the policy update. While MORE does not suffer from the listed limitations, it often cannot achieve state of the art performance in comparison to ranking based methods. We improve MORE by decoupling the update of the mean and covariance of the search distribution allowing for more aggressive updates on the mean while keeping the update on the covariance conservative, an improved entropy scheduling technique based on an evolution path which results in faster convergence and a simplified and more effective model learning approach in comparison to the original paper. We compare our algorithm to state of the art black-box optimization algorithms on standard optimization tasks as well as on episodic RL tasks in robotics where it is also crucial to have small regret. We obtain competitive results on benchmark functions and clearly outperform ranking-based methods in terms of regret on the RL tasks.

Yufeng Jin, Jianfei Guo, Xiaogang Jia et al.

Robot learning research is fragmented across policy families, benchmark suites, and real robots; each implementation is entangled with the others in a complex combination matrix, making it an engineering nightmare to port any single element. General-purpose coding agents may occasionally bridge specific setups, but cannot close this gap at scale because they lack the procedural priors and validation practices that characterize robotics research workflows. We propose NAUTILUS, an open-source harness that turns a single user prompt -- for example, "Evaluate policy A with benchmark B" -- into ready-to-use reproduction, evaluation, fine-tuning, and deployment workflows. NAUTILUS provides: plug-and-play agent skill sets with distilled priors from robotics research; typed contracts among policies, simulators/benchmarks, and real-world robots; unified interfaces and execution environments; and a trustworthy agentic coding workflow with explicit, automated validation, and testing at each milestone. NAUTILUS can not only automatically generate the required adapters and containers for existing implementations, but also wrap and onboard new or user-provided policies, simulators/benchmarks, and robots, all connected via a uniform interface. This expands cross-validation coverage without hand-written glue code. Like a nautilus shell that grows by adding chambers, NAUTILUS scales by extending its execution in chambered units, making it a research harness for scalability rather than a hand-curated framework, and aiming to reduce the engineering burden of cross-family reproduction and evaluation in the ever-growing robot learning ecosystem.

Roman Freiberg, Alexander Qualmann, Ngo Anh Vien et al.

Multi-embodiment grasping focuses on developing approaches that exhibit generalist behavior across diverse gripper designs. Existing methods often learn the kinematic structure of the robot implicitly and face challenges due to the difficulty of sourcing the required large-scale data. In this work, we present a data-efficient, flow-based, equivariant grasp synthesis architecture that can handle different gripper types with variable degrees of freedom and successfully exploit the underlying kinematic model, deducing all necessary information solely from the gripper and scene geometry. Unlike previous equivariant grasping methods, we translated all modules from the ground up to JAX and provide a model with batching capabilities over scenes, grippers, and grasps, resulting in smoother learning, improved performance and faster inference time. Our dataset encompasses grippers ranging from humanoid hands to parallel yaw grippers and includes 25,000 scenes and 20 million grasps.

Philipp Dahlinger, Balázs Gyenes, Niklas Freymuth et al.

Graph Network Simulators (GNSs) have emerged as powerful surrogates for complex physics-based simulation, offering inherent differentiability and orders-of-magnitude speedups over traditional solvers. However, GNSs typically assume access to the underlying material parameters, such as stiffness or viscosity, severely limiting their utility in realistic experimental settings. While recent meta-learning approaches address the parameter dependency by inferring properties from mesh trajectories, reconstructing a mesh from an observed scene is challenging. In this work, we introduce Point Cloud Encoding for Accurate Context Handling (PEACH), a novel framework that applies in-context learning on point clouds to adapt a learned simulator to unseen physical properties during inference. Our approach relies on a novel spatio-temporal point cloud sequence encoder, as well as two forms of auxiliary supervision to help improve simulation fidelity. We demonstrate that PEACH is capable of accurate zero-shot sim-to-real transfer on a challenging, dynamic scene. Experiments on simulation scenes show that PEACH even outperforms mesh-based baselines on prediction accuracy, while being much more practical for real-world deployment.

Denis Blessing, Onur Celik, Xiaogang Jia et al.

Imitation learning uses data for training policies to solve complex tasks. However, when the training data is collected from human demonstrators, it often leads to multimodal distributions because of the variability in human actions. Most imitation learning methods rely on a maximum likelihood (ML) objective to learn a parameterized policy, but this can result in suboptimal or unsafe behavior due to the mode-averaging property of the ML objective. In this work, we propose Information Maximizing Curriculum, a curriculum-based approach that assigns a weight to each data point and encourages the model to specialize in the data it can represent, effectively mitigating the mode-averaging problem by allowing the model to ignore data from modes it cannot represent. To cover all modes and thus, enable diverse behavior, we extend our approach to a mixture of experts (MoE) policy, where each mixture component selects its own subset of the training data for learning. A novel, maximum entropy-based objective is proposed to achieve full coverage of the dataset, thereby enabling the policy to encompass all modes within the data distribution. We demonstrate the effectiveness of our approach on complex simulated control tasks using diverse human demonstrations, achieving superior performance compared to state-of-the-art methods.

Pit Henrich, Balázs Gyenes, Paul Maria Scheikl et al.

In deformable object manipulation, we often want to interact with specific segments of an object that are only defined in non-deformed models of the object. We thus require a system that can recognize and locate these segments in sensor data of deformed real world objects. This is normally done using deformable object registration, which is problem specific and complex to tune. Recent methods utilize neural occupancy functions to improve deformable object registration by registering to an object reconstruction. Going one step further, we propose a system that in addition to reconstruction learns segmentation of the reconstructed object. As the resulting output already contains the information about the segments, we can skip the registration process. Tested on a variety of deformable objects in simulation and the real world, we demonstrate that our method learns to robustly find these segments. We also introduce a simple sampling algorithm to generate better training data for occupancy learning.

Claudius Kienle, Benjamin Alt, Onur Celik et al.

High-level robot skills represent an increasingly popular paradigm in robot programming. However, configuring the skills' parameters for a specific task remains a manual and time-consuming endeavor. Existing approaches for learning or optimizing these parameters often require numerous real-world executions or do not work in dynamic environments. To address these challenges, we propose MuTT, a novel encoder-decoder transformer architecture designed to predict environment-aware executions of robot skills by integrating vision, trajectory, and robot skill parameters. Notably, we pioneer the fusion of vision and trajectory, introducing a novel trajectory projection. Furthermore, we illustrate MuTT's efficacy as a predictor when combined with a model-based robot skill optimizer. This approach facilitates the optimization of robot skill parameters for the current environment, without the need for real-world executions during optimization. Designed for compatibility with any representation of robot skills, MuTT demonstrates its versatility across three comprehensive experiments, showcasing superior performance across two different skill representations.

C. F. Maximilian Nagy, Onur Celik, Emiliyan Gospodinov et al.

Action chunking can improve exploration and value estimation in long horizon reinforcement learning, but makes learning substantially harder since the critic must evaluate action sequences rather than single actions, greatly increasing approximation and data efficiency challenges. As a result, existing action chunking methods, primarily designed for the offline and offline-to-online settings, have not achieved strong performance in purely online reinforcement learning. We introduce SEAR, an off policy online reinforcement learning algorithm for action chunking. It exploits the temporal structure of action chunks and operates with a receding horizon, effectively combining the benefits of small and large chunk sizes. SEAR outperforms state of the art online reinforcement learning methods on Metaworld, training with chunk sizes up to 20.

Xiaogang Jia, Atalay Donat, Xi Huang et al.

Designing modern imitation learning (IL) policies requires making numerous decisions, including the selection of feature encoding, architecture, policy representation, and more. As the field rapidly advances, the range of available options continues to grow, creating a vast and largely unexplored design space for IL policies. In this work, we present X-IL, an accessible open-source framework designed to systematically explore this design space. The framework's modular design enables seamless swapping of policy components, such as backbones (e.g., Transformer, Mamba, xLSTM) and policy optimization techniques (e.g., Score-matching, Flow-matching). This flexibility facilitates comprehensive experimentation and has led to the discovery of novel policy configurations that outperform existing methods on recent robot learning benchmarks. Our experiments demonstrate not only significant performance gains but also provide valuable insights into the strengths and weaknesses of various design choices. This study serves as both a practical reference for practitioners and a foundation for guiding future research in imitation learning.

Andreas Boltres, Niklas Freymuth, Benjamin Schichtholz et al.

Routing algorithms are crucial for efficient computer network operations, and in many settings they must be able to react to traffic bursts within milliseconds. Live telemetry data can provide informative signals to routing algorithms, and recent work has trained neural networks to exploit such signals for traffic-aware routing. Yet, aggregating network-wide information is subject to communication delays, and existing neural approaches either assume unrealistic delay-free global states, or restrict routers to purely local telemetry. This leaves their deployability in real-world environments unclear. We cast telemetry-aware routing as a delay-aware closed-loop control problem and introduce a framework that trains and evaluates neural routing algorithms, while explicitly modeling communication and inference delays. On top of this framework, we propose LOGGIA, a scalable graph neural routing algorithm that predicts log-space link weights from attributed topology-and-telemetry graphs. It utilizes a data-driven pre-training stage, followed by on-policy Reinforcement Learning. Across synthetic and real network topologies, and unseen mixed TCP/UDP traffic sequences, LOGGIA consistently outperforms shortest-path baselines, whereas neural baselines fail once realistic delays are enforced. Our experiments further suggest that neural routing algorithms like LOGGIA perform best when deployed fully locally, i.e., observing network states and inferring actions at every router individually, as opposed to centralized decision making.

Tai Hoang, Alessandro Trenta, Alessio Gravina et al.

Learning to simulate complex physical systems from data has emerged as a promising way to overcome the limitations of traditional numerical solvers, which often require prohibitive computational costs for high-fidelity solutions. Recent Graph Neural Simulators (GNSs) accelerate simulations by learning dynamics on graph-structured data, yet often struggle to capture long-range interactions and suffer from error accumulation under autoregressive rollouts. To address these challenges, we propose Information-preserving Graph Neural Simulators (IGNS), a graph-based neural simulator built on the principles of Hamiltonian dynamics. This structure guarantees preservation of information across the graph, while extending to port-Hamiltonian systems allows the model to capture a broader class of dynamics, including non-conservative effects. IGNS further incorporates a warmup phase to initialize global context, geometric encoding to handle irregular meshes, and a multi-step training objective to reduce rollout error. To evaluate these properties systematically, we introduce new benchmarks that target long-range dependencies and challenging external forcing scenarios. Across all tasks, IGNS consistently outperforms state-of-the-art GNSs, achieving higher accuracy and stability under challenging and complex dynamical systems.

Philipp Dahlinger, Philipp Becker, Maximilian Hüttenrauch et al.

Stochastic gradient-based optimization is crucial to optimize neural networks. While popular approaches heuristically adapt the step size and direction by rescaling gradients, a more principled approach to improve optimizers requires second-order information. Such methods precondition the gradient using the objective's Hessian. Yet, computing the Hessian is usually expensive and effectively using second-order information in the stochastic gradient setting is non-trivial. We propose using Information-Theoretic Trust Region Optimization (arTuRO) for improved updates with uncertain second-order information. By modeling the network parameters as a Gaussian distribution and using a Kullback-Leibler divergence-based trust region, our approach takes bounded steps accounting for the objective's curvature and uncertainty in the parameters. Before each update, it solves the trust region problem for an optimal step size, resulting in a more stable and faster optimization process. We approximate the diagonal elements of the Hessian from stochastic gradients using a simple recursive least squares approach, constructing a model of the expected Hessian over time using only first-order information. We show that arTuRO combines the fast convergence of adaptive moment-based optimization with the generalization capabilities of SGD.

Andreas Boltres, Niklas Freymuth, Gerhard Neumann

Telemetry-Aware routing promises to increase efficacy and responsiveness to traffic surges in computer networks. Recent research leverages Machine Learning to deal with the complex dependency between network state and routing, but sacrifices explainability of routing decisions due to the black-box nature of the proposed neural routing modules. We propose \emph{Placer}, a novel algorithm using Message Passing Networks to transform network states into latent node embeddings. These embeddings facilitate quick greedy next-hop routing without directly solving the all-pairs shortest paths problem, and let us visualize how certain network events shape routing decisions.

Philipp Dahlinger, Niklas Freymuth, Tai Hoang et al.

Simulating object deformations is a critical challenge across many scientific domains, including robotics, manufacturing, and structural mechanics. Learned Graph Network Simulators (GNSs) offer a promising alternative to traditional mesh-based physics simulators. Their speed and inherent differentiability make them particularly well suited for applications that require fast and accurate simulations, such as robotic manipulation or manufacturing optimization. However, existing learned simulators typically rely on single-step observations, which limits their ability to exploit temporal context. Without this information, these models fail to infer, e.g., material properties. Further, they rely on auto-regressive rollouts, which quickly accumulate error for long trajectories. We instead frame mesh-based simulation as a trajectory-level meta-learning problem. Using Conditional Neural Processes, our method enables rapid adaptation to new simulation scenarios from limited initial data while capturing their latent simulation properties. We utilize movement primitives to directly predict fast, stable and accurate simulations from a single model call. The resulting approach, Movement-primitive Meta-MeshGraphNet (M3GN), provides higher simulation accuracy at a fraction of the runtime cost compared to state-of-the-art GNSs across several tasks.

Philipp Dahlinger, Niklas Freymuth, Michael Volpp et al.

Simulating dynamic physical interactions is a critical challenge across multiple scientific domains, with applications ranging from robotics to material science. For mesh-based simulations, Graph Network Simulators (GNSs) pose an efficient alternative to traditional physics-based simulators. Their inherent differentiability and speed make them particularly well-suited for inverse design problems. Yet, adapting to new tasks from limited available data is an important aspect for real-world applications that current methods struggle with. We frame mesh-based simulation as a meta-learning problem and use a recent Bayesian meta-learning method to improve GNSs adaptability to new scenarios by leveraging context data and handling uncertainties. Our approach, latent task-specific graph network simulator, uses non-amortized task posterior approximations to sample latent descriptions of unknown system properties. Additionally, we leverage movement primitives for efficient full trajectory prediction, effectively addressing the issue of accumulating errors encountered by previous auto-regressive methods. We validate the effectiveness of our approach through various experiments, performing on par with or better than established baseline methods. Movement primitives further allow us to accommodate various types of context data, as demonstrated through the utilization of point clouds during inference. By combining GNSs with meta-learning, we bring them closer to real-world applicability, particularly in scenarios with smaller datasets.

Xiaogang Jia, Qian Wang, Atalay Donat et al.

This work presents Mamba Imitation Learning (MaIL), a novel imitation learning (IL) architecture that provides an alternative to state-of-the-art (SoTA) Transformer-based policies. MaIL leverages Mamba, a state-space model designed to selectively focus on key features of the data. While Transformers are highly effective in data-rich environments due to their dense attention mechanisms, they can struggle with smaller datasets, often leading to overfitting or suboptimal representation learning. In contrast, Mamba's architecture enhances representation learning efficiency by focusing on key features and reducing model complexity. This approach mitigates overfitting and enhances generalization, even when working with limited data. Extensive evaluations on the LIBERO benchmark demonstrate that MaIL consistently outperforms Transformers on all LIBERO tasks with limited data and matches their performance when the full dataset is available. Additionally, MaIL's effectiveness is validated through its superior performance in three real robot experiments. Our code is available at https://github.com/ALRhub/MaIL.

Fabian Otto, Philipp Becker, Ngo Anh Vien et al.

Trust region methods are a popular tool in reinforcement learning as they yield robust policy updates in continuous and discrete action spaces. However, enforcing such trust regions in deep reinforcement learning is difficult. Hence, many approaches, such as Trust Region Policy Optimization (TRPO) and Proximal Policy Optimization (PPO), are based on approximations. Due to those approximations, they violate the constraints or fail to find the optimal solution within the trust region. Moreover, they are difficult to implement, often lack sufficient exploration, and have been shown to depend on seemingly unrelated implementation choices. In this work, we propose differentiable neural network layers to enforce trust regions for deep Gaussian policies via closed-form projections. Unlike existing methods, those layers formalize trust regions for each state individually and can complement existing reinforcement learning algorithms. We derive trust region projections based on the Kullback-Leibler divergence, the Wasserstein L2 distance, and the Frobenius norm for Gaussian distributions. We empirically demonstrate that those projection layers achieve similar or better results than existing methods while being almost agnostic to specific implementation choices. The code is available at https://git.io/Jthb0.

André Eberhard, Gerhard Neumann, Pascal Friederich

Despite tremendous improvements in tasks such as image classification, object detection, and segmentation, the recognition of visual relationships, commonly modeled as the extraction of a graph from an image, remains a challenging task. We believe that this mainly stems from the fact that there is no canonical way to approach the visual graph recognition task. Most existing solutions are specific to a problem and cannot be transferred between different contexts out-of-the box, even though the conceptual problem remains the same. With broad applicability and simplicity in mind, in this paper we develop a method, \textbf{Gra}ph Recognition via \textbf{S}ubgraph \textbf{P}rediction (\textbf{GraSP}), for recognizing graphs in images. We show across several synthetic benchmarks and one real-world application that our method works with a set of diverse types of graphs and their drawings, and can be transferred between tasks without task-specific modifications, paving the way to a more unified framework for visual graph recognition.

Paul Maria Scheikl, Nicolas Schreiber, Christoph Haas et al.

Policy learning in robot-assisted surgery (RAS) lacks data efficient and versatile methods that exhibit the desired motion quality for delicate surgical interventions. To this end, we introduce Movement Primitive Diffusion (MPD), a novel method for imitation learning (IL) in RAS that focuses on gentle manipulation of deformable objects. The approach combines the versatility of diffusion-based imitation learning (DIL) with the high-quality motion generation capabilities of Probabilistic Dynamic Movement Primitives (ProDMPs). This combination enables MPD to achieve gentle manipulation of deformable objects, while maintaining data efficiency critical for RAS applications where demonstration data is scarce. We evaluate MPD across various simulated and real world robotic tasks on both state and image observations. MPD outperforms state-of-the-art DIL methods in success rate, motion quality, and data efficiency. Project page: https://scheiklp.github.io/movement-primitive-diffusion/

Enes Ulas Dincer, Manuel Zaremski, Alexandra Nick et al.

Robot teleoperation is critical for applications such as remote maintenance, fleet robotics, search and rescue, and data collection for robot learning. Effective teleoperation requires intuitive 3D visualization with reliable depth cues, which conventional screen-based interfaces often fail to provide. We introduce a multi-view VR telepresence system that (1) fuses geometry from three cameras to produce GPU-accelerated point-cloud rendering on standalone VR hardware, and (2) integrates a wrist-mounted RGB stream to provide high-resolution local detail where point-cloud accuracy is limited. Our pipeline supports real-time rendering of approximately 75k points on the Meta Quest 3. A within-subject study was conducted with 31 participants to compare our system to other visualisation modalities, such as RGB streams, a projection of stereo-vision directly in the VR device and point clouds without providing additional RGB information. Across three different teleoperated manipulation tasks, we measured task success, completion time, perceived workload, and usability. Our system achieved the best overall performance, while the Point Cloud modality without RGB also outperforming the RGB streams and OpenTeleVision. These results show that combining global 3D structure with localized high-resolution detail substantially improves telepresence for manipulation and provides a strong foundation for next-generation robot teleoperation systems.

Junhua Chen, Lorenz Richter, Julius Berner et al.

An effective approach for sampling from unnormalized densities is based on the idea of gradually transporting samples from an easy prior to the complicated target distribution. Two popular methods are (1) Sequential Monte Carlo (SMC), where the transport is performed through successive annealed densities via prescribed Markov chains and resampling steps, and (2) recently developed diffusion-based sampling methods, where a learned dynamical transport is used. Despite the common goal, both approaches have different, often complementary, advantages and drawbacks. The resampling steps in SMC allow focusing on promising regions of the space, often leading to robust performance. While the algorithm enjoys asymptotic guarantees, the lack of flexible, learnable transitions can lead to slow convergence. On the other hand, diffusion-based samplers are learned and can potentially better adapt themselves to the target at hand, yet often suffer from training instabilities. In this work, we present a principled framework for combining SMC with diffusion-based samplers by viewing both methods in continuous time and considering measures on path space. This culminates in the new Sequential Controlled Langevin Diffusion (SCLD) sampling method, which is able to utilize the benefits of both methods and reaches improved performance on multiple benchmark problems, in many cases using only 10% of the training budget of previous diffusion-based samplers.

Denis Blessing, Julius Berner, Lorenz Richter et al.

We provide a general framework for learning diffusion bridges that transport prior to target distributions. It includes existing diffusion models for generative modeling, but also underdamped versions with degenerate diffusion matrices, where the noise only acts in certain dimensions. Extending previous findings, our framework allows to rigorously show that score matching in the underdamped case is indeed equivalent to maximizing a lower bound on the likelihood. Motivated by superior convergence properties and compatibility with sophisticated numerical integration schemes of underdamped stochastic processes, we propose \emph{underdamped diffusion bridges}, where a general density evolution is learned rather than prescribed by a fixed noising process. We apply our method to the challenging task of sampling from unnormalized densities without access to samples from the target distribution. Across a diverse range of sampling problems, our approach demonstrates state-of-the-art performance, notably outperforming alternative methods, while requiring significantly fewer discretization steps and no hyperparameter tuning.

Onur Celik, Zechu Li, Denis Blessing et al.

Maximum entropy reinforcement learning (MaxEnt-RL) has become the standard approach to RL due to its beneficial exploration properties. Traditionally, policies are parameterized using Gaussian distributions, which significantly limits their representational capacity. Diffusion-based policies offer a more expressive alternative, yet integrating them into MaxEnt-RL poses challenges-primarily due to the intractability of computing their marginal entropy. To overcome this, we propose Diffusion-Based Maximum Entropy RL (DIME). \emph{DIME} leverages recent advances in approximate inference with diffusion models to derive a lower bound on the maximum entropy objective. Additionally, we propose a policy iteration scheme that provably converges to the optimal diffusion policy. Our method enables the use of expressive diffusion-based policies while retaining the principled exploration benefits of MaxEnt-RL, significantly outperforming other diffusion-based methods on challenging high-dimensional control benchmarks. It is also competitive with state-of-the-art non-diffusion based RL methods while requiring fewer algorithmic design choices and smaller update-to-data ratios, reducing computational complexity.