Danica Kragic

Alfredo Reichlin, Giovanni Luca Marchetti, Hang Yin et al. · stanford

Learning from previously collected datasets of expert data offers the promise of acquiring robotic policies without unsafe and costly online explorations. However, a major challenge is a distributional shift between the states in the training dataset and the ones visited by the learned policy at the test time. While prior works mainly studied the distribution shift caused by the policy during the offline training, the problem of recovering from out-of-distribution states at the deployment time is not very well studied yet. We alleviate the distributional shift at the deployment time by introducing a recovery policy that brings the agent back to the training manifold whenever it steps out of the in-distribution states, e.g., due to an external perturbation. The recovery policy relies on an approximation of the training data density and a learned equivariant mapping that maps visual observations into a latent space in which translations correspond to the robot actions. We demonstrate the effectiveness of the proposed method through several manipulation experiments on a real robotic platform. Our results show that the recovery policy enables the agent to complete tasks while the behavioral cloning alone fails because of the distributional shift problem.

Ali Ghadirzadeh, Petra Poklukar, Karol Arndt et al. · stanford

We present a data-efficient framework for solving sequential decision-making problems which exploits the combination of reinforcement learning (RL) and latent variable generative models. The framework, called GenRL, trains deep policies by introducing an action latent variable such that the feed-forward policy search can be divided into two parts: (i) training a sub-policy that outputs a distribution over the action latent variable given a state of the system, and (ii) unsupervised training of a generative model that outputs a sequence of motor actions conditioned on the latent action variable. GenRL enables safe exploration and alleviates the data-inefficiency problem as it exploits prior knowledge about valid sequences of motor actions. Moreover, we provide a set of measures for evaluation of generative models such that we are able to predict the performance of the RL policy training prior to the actual training on a physical robot. We experimentally determine the characteristics of generative models that have most influence on the performance of the final policy training on two robotics tasks: shooting a hockey puck and throwing a basketball. Furthermore, we empirically demonstrate that GenRL is the only method which can safely and efficiently solve the robotics tasks compared to two state-of-the-art RL methods.

Alberta Longhini, Marco Moletta, Alfredo Reichlin et al. · cmu

We study the problem of learning graph dynamics of deformable objects that generalizes to unknown physical properties. Our key insight is to leverage a latent representation of elastic physical properties of cloth-like deformable objects that can be extracted, for example, from a pulling interaction. In this paper we propose EDO-Net (Elastic Deformable Object - Net), a model of graph dynamics trained on a large variety of samples with different elastic properties that does not rely on ground-truth labels of the properties. EDO-Net jointly learns an adaptation module, and a forward-dynamics module. The former is responsible for extracting a latent representation of the physical properties of the object, while the latter leverages the latent representation to predict future states of cloth-like objects represented as graphs. We evaluate EDO-Net both in simulation and real world, assessing its capabilities of: 1) generalizing to unknown physical properties, 2) transferring the learned representation to new downstream tasks.

Haofei Lu, Hongjia Liu, Yifei Dong et al.

This work addresses sequentially grasping multiple objects with a single dexterous hand without releasing those already held. Most dexterous grasping methods commit all of the hand's degrees of freedom to a single object, underutilizing its dexterity and leaving no redundancy for subsequent grasps. The proposed solution, MoDex, is a diffusion policy that predicts the next gripper pose directly from observations, conditioned on an opposition space and point cloud. The opposition space condition specifies which fingers participate in the current grasp, enabling the gripper to use only a subset of its available degrees of freedom while reserving the remaining degrees of freedom for subsequent grasps. To facilitate sim-to-real transfer, MoDex is trained in two stages: first through imitation learning on expert demonstrations, and subsequently through reinforcement learning fine-tuning, which consistently improves success rates over the pre-trained policy. We evaluate MoDex in simulation on a MuJoCo-based Franka Emika Panda robot equipped with an Allegro Hand and on the corresponding real-world hardware platform. Across both simulation and real-world experiments, MoDex achieves higher success rates than the evaluated learning-based baselines, improving performance by 2.92-17.92% and 6.67-17.78%, respectively. Project page: https://modex2026.github.io/.

Noémie Jaquier, Michael C. Welle, Andrej Gams et al.

Transfer learning is a conceptually-enticing paradigm in pursuit of truly intelligent embodied agents. The core concept -- reusing prior knowledge to learn in and from novel situations -- is successfully leveraged by humans to handle novel situations. In recent years, transfer learning has received renewed interest from the community from different perspectives, including imitation learning, domain adaptation, and transfer of experience from simulation to the real world, among others. In this paper, we unify the concept of transfer learning in robotics and provide the first taxonomy of its kind considering the key concepts of robot, task, and environment. Through a review of the promises and challenges in the field, we identify the need of transferring at different abstraction levels, the need of quantifying the transfer gap and the quality of transfer, as well as the dangers of negative transfer. Via this position paper, we hope to channel the effort of the community towards the most significant roadblocks to realize the full potential of transfer learning in robotics.

Wenjie Yin, Hang Yin, Kim Baraka et al.

We present CycleDance, a dance style transfer system to transform an existing motion clip in one dance style to a motion clip in another dance style while attempting to preserve motion context of the dance. Our method extends an existing CycleGAN architecture for modeling audio sequences and integrates multimodal transformer encoders to account for music context. We adopt sequence length-based curriculum learning to stabilize training. Our approach captures rich and long-term intra-relations between motion frames, which is a common challenge in motion transfer and synthesis work. We further introduce new metrics for gauging transfer strength and content preservation in the context of dance movements. We perform an extensive ablation study as well as a human study including 30 participants with 5 or more years of dance experience. The results demonstrate that CycleDance generates realistic movements with the target style, significantly outperforming the baseline CycleGAN on naturalness, transfer strength, and content preservation.

Giovanni Luca Marchetti, Gustaf Tegnér, Anastasiia Varava et al.

We introduce a general method for learning representations that are equivariant to symmetries of data. Our central idea is to decompose the latent space into an invariant factor and the symmetry group itself. The components semantically correspond to intrinsic data classes and poses respectively. The learner is trained on a loss encouraging equivariance based on supervision from relative symmetry information. The approach is motivated by theoretical results from group theory and guarantees representations that are lossless, interpretable and disentangled. We provide an empirical investigation via experiments involving datasets with a variety of symmetries. Results show that our representations capture the geometry of data and outperform other equivariant representation learning frameworks.

Yuchong Zhang, Bastian Orthmann, Shichen Ji et al.

Robot teleoperation via augmented reality (AR) offers a promising path toward more intuitive human-robot interaction (HRI). We present a head-mounted AR 'puppeteer' system in which users control a physical robot by interacting with its virtual counterpart robot using large language model (LLM)-assisted voice commands and hand-gesture interaction on the Meta Quest 3. In a within-subject user study with 42 participants performing an AR-based robotic pick-and-place pattern-matching task, we empirically compare two interaction conditions: gesture-only (GO) and combined voice+gesture (VG) on performance and user experience (UX). In VG, voice and gesture operate in a sequential role-allocated manner, with voice handling high-level navigation and gesture handling fine manipulation. Our results show that GO currently provides more reliable and efficient control for this time-critical task, while VG introduces additional flexibility but also latency and recognition issues that can increase workload. We additionally analyze how prior robotics expertise differentiates performance and UX across conditions. Based on these findings, we distill a set of design guidelines for AR 'puppeteer' metaphoric robot teleoperation, framing multimodality as an adaptive strategy that must balance efficiency, robustness, and user expertise rather than assuming that additional modalities are universally beneficial.

Gustaf Tegnér, Alfredo Reichlin, Hang Yin et al.

In this work we provide an analysis of the distribution of the post-adaptation parameters of Gradient-Based Meta-Learning (GBML) methods. Previous work has noticed how, for the case of image-classification, this adaptation only takes place on the last layers of the network. We propose the more general notion that parameters are updated over a low-dimensional \emph{subspace} of the same dimensionality as the task-space and show that this holds for regression as well. Furthermore, the induced subspace structure provides a method to estimate the intrinsic dimension of the space of tasks of common few-shot learning datasets.

Alexander Kravberg, Giovanni Luca Marchetti, Vladislav Polianskii et al.

We introduce an algorithm for active function approximation based on nearest neighbor regression. Our Active Nearest Neighbor Regressor (ANNR) relies on the Voronoi-Delaunay framework from computational geometry to subdivide the space into cells with constant estimated function value and select novel query points in a way that takes the geometry of the function graph into account. We consider the recent state-of-the-art active function approximator called DEFER, which is based on incremental rectangular partitioning of the space, as the main baseline. The ANNR addresses a number of limitations that arise from the space subdivision strategy used in DEFER. We provide a computationally efficient implementation of our method, as well as theoretical halting guarantees. Empirical results show that ANNR outperforms the baseline for both closed-form functions and real-world examples, such as gravitational wave parameter inference and exploration of the latent space of a generative model.

Wenjie Yin, Ruibo Tu, Hang Yin et al.

Data-driven and controllable human motion synthesis and prediction are active research areas with various applications in interactive media and social robotics. Challenges remain in these fields for generating diverse motions given past observations and dealing with imperfect poses. This paper introduces MoDiff, an autoregressive probabilistic diffusion model over motion sequences conditioned on control contexts of other modalities. Our model integrates a cross-modal Transformer encoder and a Transformer-based decoder, which are found effective in capturing temporal correlations in motion and control modalities. We also introduce a new data dropout method based on the diffusion forward process to provide richer data representations and robust generation. We demonstrate the superior performance of MoDiff in controllable motion synthesis for locomotion with respect to two baselines and show the benefits of diffusion data dropout for robust synthesis and reconstruction of high-fidelity motion close to recorded data.

Ziqi Zhang, Eunyeong Jin, Miguel Vasco et al.

Predicting human olfactory perception from molecular structure has seen remarkable progress, yet these approaches require explicit chemical structure at inference, which is not available in practical sensing settings. We address this gap by exploring direct electron ionization mass spectrometry (EI-MS), a sensing technique that acquires chemically informative fragmentation fingerprints in seconds, as an alternative input modality for olfactory prediction. We contribute Spectrum-to-Chemical Embedding alignmeNT (SCENT), a multi-modal contrastive learning framework that aligns EI-MS representations with pretrained chemical structure embeddings, while requiring only mass spectra at inference. On the multi-label odor descriptor prediction task, SCENT significantly outperforms MS-only baselines and achieves performance comparable to structure-based models, despite requiring no explicit molecular structure at test time. The learned representations also better approximate continuous human perceptual ratings and generalize to real-world lab-measured spectra, suggesting that cross-modal alignment is an effective strategy for grounding analytical spectra in chemical semantics.

Thomas Lips, Marco Moletta, Michael C. Welle et al.

RGB-based imitation learning requires many demonstrations to generalize to unseen objects or scenes, motivating research into intermediate representations to improve generalization for robotic manipulation. Visual foundation models enable one-shot extraction of keypoints to provide such representation. However, it remains unclear how to integrate them into imitation learning optimally and when they outperform alternative representations. We combine approaches from previous works on keypoint imitation learning (KIL) and investigate several design choices to provide practical guidelines. Using over 2000 real-world rollouts, we also assess the generalization capabilities of KIL to unseen objects and scene variations. KIL achieves a 75% overall success rate across five tasks, significantly outperforming the RGB baseline (47%) and performing on par with S2-diffusion (73%). Finally, we explore the limitations of the foundation models used for keypoint extraction and extend KIL to tasks with multiple object instances. Our results confirm KIL as a data-efficient approach for robot learning, though it does not outperform alternative representations and inherits limitations of the foundation models used for keypoint extraction. All rollout videos, demonstrations, and results are available at https://kil-manipulation.github.io/.

Mohammed Al-Jaff, Giovanni Luca Marchetti, Michael C Welle et al.

Idempotent Generative Networks (IGNs) are deep generative models that also function as local data manifold projectors, mapping arbitrary inputs back onto the manifold. They are trained to act as identity operators on the data and as idempotent operators off the data manifold. However, IGNs suffer from mode collapse, mode dropping, and training instability due to their objectives, which contain adversarial components and can cause the model to cover the data manifold only partially -- an issue shared with generative adversarial networks. We introduce Non-Adversarial Idempotent Generative Networks (NAIGNs) to address these issues. Our loss function combines reconstruction with the non-adversarial generative objective of Implicit Maximum Likelihood Estimation (IMLE). This improves on IGN's ability to restore corrupted data and generate new samples that closely match the data distribution. We moreover demonstrate that NAIGNs implicitly learn the distance field to the data manifold, as well as an energy-based model.

Alejandro García-Castellanos, Giovanni Luca Marchetti, Danica Kragic et al.

Relative representations are an established approach to zero-shot model stitching, consisting of a non-trainable transformation of the latent space of a deep neural network. Based on insights of topological and geometric nature, we propose two improvements to relative representations. First, we introduce a normalization procedure in the relative transformation, resulting in invariance to non-isotropic rescalings and permutations. The latter coincides with the symmetries in parameter space induced by common activation functions. Second, we propose to deploy topological densification when fine-tuning relative representations, a topological regularization loss encouraging clustering within classes. We provide an empirical investigation on a natural language task, where both the proposed variations yield improved performance on zero-shot model stitching.

Wenjie Yin, Qingyuan Yao, Yi Yu et al.

Lyrics often convey information about the songs that are beyond the auditory dimension, enriching the semantic meaning of movements and musical themes. Such insights are important in the dance choreography domain. However, most existing dance synthesis methods mainly focus on music-to-dance generation, without considering the semantic information. To complement it, we introduce JustLMD, a new multimodal dataset of 3D dance motion with music and lyrics. To the best of our knowledge, this is the first dataset with triplet information including dance motion, music, and lyrics. Additionally, we showcase a cross-modal diffusion-based network designed to generate 3D dance motion conditioned on music and lyrics. The proposed JustLMD dataset encompasses 4.6 hours of 3D dance motion in 1867 sequences, accompanied by musical tracks and their corresponding English lyrics.

Ciwan Ceylan, Kambiz Ghoorchian, Danica Kragic

Structural node embeddings, vectors capturing local connectivity information for each node in a graph, have many applications in data mining and machine learning, e.g., network alignment and node classification, clustering and anomaly detection. For the analysis of directed graphs, e.g., transactions graphs, communication networks and social networks, the capability to capture directional information in the structural node embeddings is highly desirable, as is scalability of the embedding extraction method. Most existing methods are nevertheless only designed for undirected graph. Therefore, we present Digraphwave -- a scalable algorithm for extracting structural node embeddings on directed graphs. The Digraphwave embeddings consist of compressed diffusion pattern signatures, which are twice enhanced to increase their discriminate capacity. By proving a lower bound on the heat contained in the local vicinity of a diffusion initialization node, theoretically justified diffusion timescale values are established, and Digraphwave is left with only two easy-to-interpret hyperparameters: the embedding dimension and a neighbourhood resolution specifier. In our experiments, the two embedding enhancements, named transposition and aggregation, are shown to lead to a significant increase in macro F1 score for classifying automorphic identities, with Digraphwave outperforming all other structural embedding baselines. Moreover, Digraphwave either outperforms or matches the performance of all baselines on real graph datasets, displaying a particularly large performance gain in a network alignment task, while also being scalable to graphs with millions of nodes and edges, running up to 30x faster than a previous diffusion pattern based method and with a fraction of the memory consumption.

Luis Armando Pérez Rey, Giovanni Luca Marchetti, Danica Kragic et al.

We introduce Equivariant Isomorphic Networks (EquIN) -- a method for learning representations that are equivariant with respect to general group actions over data. Differently from existing equivariant representation learners, EquIN is suitable for group actions that are not free, i.e., that stabilize data via nontrivial symmetries. EquIN is theoretically grounded in the orbit-stabilizer theorem from group theory. This guarantees that an ideal learner infers isomorphic representations while trained on equivariance alone and thus fully extracts the geometric structure of data. We provide an empirical investigation on image datasets with rotational symmetries and show that taking stabilizers into account improves the quality of the representations.

Yuchong Zhang, Yong Ma, Danica Kragic

This perspective reframes human-robot interaction (HRI) through extended reality (XR), arguing that virtual robots powered by large foundation models (FMs) can serve as cognitively grounded, empathic agents. Unlike physical robots, XR-native agents are unbound by hardware constraints and can be instantiated, adapted, and scaled on demand, while still affording embodiment and co-presence. We synthesize work across XR, HRI, and cognitive AI to show how such agents can support safety-critical scenarios, socially and cognitively empathic interaction across domains, and outreaching physical capabilities with XR and AI integration. We then discuss how multimodal large FMs (e.g., large language model, large vision model, and vision-language model) enable context-aware reasoning, affect-sensitive situations, and long-term adaptation, positioning virtual robots as cognitive and empathic mediators rather than mere simulation assets. At the same time, we highlight challenges and potential risks, including overtrust, cultural and representational bias, privacy concerns around biometric sensing, and data governance and transparency. The paper concludes by outlining a research agenda for human-centered, ethically grounded XR agents - emphasizing multi-layered evaluation frameworks, multi-user ecosystems, mixed virtual-physical embodiment, and societal and ethical design practices to envision XR-based virtual agents powered by FMs as reshaping future HRI into a more efficient and adaptive paradigm.

Alfredo Reichlin, Miguel Vasco, Danica Kragic

Bayesian Neural Networks provide a principled framework for uncertainty quantification by modeling the posterior distribution of network parameters. However, exact posterior inference is computationally intractable, and widely used approximations like the Laplace method struggle with scalability and posterior accuracy in modern deep networks. In this work, we revisit sampling techniques for posterior exploration, proposing a simple variation tailored to efficiently sample from the posterior in over-parameterized networks by leveraging the low-dimensional structure of loss minima. Building on this, we introduce a model that learns a deformation of the parameter space, enabling rapid posterior sampling without requiring iterative methods. Empirical results demonstrate that our approach achieves competitive posterior approximations with improved scalability compared to recent refinement techniques. These contributions provide a practical alternative for Bayesian inference in deep learning.

Alfredo Reichlin, Adriano Pacciarelli, Danica Kragic et al.

Estimating the state of an environment from high-dimensional, multimodal, and noisy observations is a fundamental challenge in reinforcement learning (RL). Traditional approaches rely on probabilistic models to account for the uncertainty, but often require explicit noise assumptions, in turn limiting generalization. In this work, we contribute a novel method to learn a structured latent representation, in which distances between states directly correlate with the minimum number of actions required to transition between them. The proposed metric space formulation provides a geometric interpretation of uncertainty without the need for explicit probabilistic modeling. To achieve this, we introduce a multimodal latent transition model and a sensor fusion mechanism based on inverse distance weighting, allowing for the adaptive integration of multiple sensor modalities without prior knowledge of noise distributions. We empirically validate the approach on a range of multimodal RL tasks, demonstrating improved robustness to sensor noise and superior state estimation compared to baseline methods. Our experiments show enhanced performance of an RL agent via the learned representation, eliminating the need of explicit noise augmentation. The presented results suggest that leveraging transition-aware metric spaces provides a principled and scalable solution for robust state estimation in sequential decision-making.

Alfredo Reichlin, Giovanni Luca Marchetti, Hang Yin et al.

We address the problem of learning representations from observations of a scene involving an agent and an external object the agent interacts with. To this end, we propose a representation learning framework extracting the location in physical space of both the agent and the object from unstructured observations of arbitrary nature. Our framework relies on the actions performed by the agent as the only source of supervision, while assuming that the object is displaced by the agent via unknown dynamics. We provide a theoretical foundation and formally prove that an ideal learner is guaranteed to infer an isometric representation, disentangling the agent from the object and correctly extracting their locations. We evaluate empirically our framework on a variety of scenarios, showing that it outperforms vision-based approaches such as a state-of-the-art keypoint extractor. We moreover demonstrate how the extracted representations enable the agent to solve downstream tasks via reinforcement learning in an efficient manner.

Daniel Palenicek, Florian Vogt, Joe Watson et al.

For reinforcement learning in the real world online exploration is expensive A common practice in robotic reinforcement learning is to incorporate additional data to improve sample efficiency Expert demonstration data is often crucial for solving hard exploration tasks with sparse rewards While prior data is used to augment experience and pretrain models we show that the design of existing algorithms fails to achieve the sample efficiency that is possible in this setting due to a failure to use pretrained policies effectively We propose XQCfD which extends the sample-efficient XQC actor-critic to learn from demonstrations using augmented replay buffers pretrained policies and stationary policy architectures designed to avoid rapidly unlearning the strong initial policy like prior works We show our stationary network architecture enables policy improvement out-of-distribution better than standard network architectures due to its higher entropy predictions XQCfD achieves state of the art performance across a range of complex manipulation tasks with sparse rewards from the popular Adroit Robomimic and MimicGen benchmarks -- notably with a low update-to-data ratio and no ensemble networks

Zehang Weng, Fabian Paus, Anastasiia Varava et al.

Capturing scene dynamics and predicting the future scene state is challenging but essential for robotic manipulation tasks, especially when the scene contains both rigid and deformable objects. In this work, we contribute a simulation environment and generate a novel dataset for task-specific manipulation, involving interactions between rigid objects and a deformable bag. The dataset incorporates a rich variety of scenarios including different object sizes, object numbers and manipulation actions. We approach dynamics learning by proposing an object-centric graph representation and two modules which are Active Prediction Module (APM) and Position Prediction Module (PPM) based on graph neural networks with an encode-process-decode architecture. At the inference stage, we build a two-stage model based on the learned modules for single time step prediction. We combine modules with different prediction horizons into a mixed-horizon model which addresses long-term prediction. In an ablation study, we show the benefits of the two-stage model for single time step prediction and the effectiveness of the mixed-horizon model for long-term prediction tasks. Supplementary material is available at https://github.com/wengzehang/deformable_rigid_interaction_prediction

Ioanna Mitsioni, Joonatan Mänttäri, Yiannis Karayiannidis et al.

Deep Neural Networks (NNs) have been widely utilized in contact-rich manipulation tasks to model the complicated contact dynamics. However, NN-based models are often difficult to decipher which can lead to seemingly inexplicable behaviors and unidentifiable failure cases. In this work, we address the interpretability of NN-based models by introducing the kinodynamic images. We propose a methodology that creates images from the kinematic and dynamic data of a contact-rich manipulation task. Our formulation visually reflects the task's state by encoding its kinodynamic variations and temporal evolution. By using images as the state representation, we enable the application of interpretability modules that were previously limited to vision-based tasks. We use this representation to train Convolution-based Networks and we extract interpretations of the model's decisions with Grad-CAM, a technique that produces visual explanations. Our method is versatile and can be applied to any classification problem using synchronous features in manipulation to visually interpret which parts of the input drive the model's decisions and distinguish its failure modes. We evaluate this approach on two examples of real-world contact-rich manipulation: pushing and cutting, with known and unknown objects. Finally, we demonstrate that our method enables both detailed visual inspections of sequences in a task, as well as high-level evaluations of a model's behavior and tendencies. Data and code for this work are available at https://github.com/imitsioni/interpretable_manipulation.

Zehang Weng, Haofei Lu, Danica Kragic et al.

We introduce DexDiffuser, a novel dexterous grasping method that generates, evaluates, and refines grasps on partial object point clouds. DexDiffuser includes the conditional diffusion-based grasp sampler DexSampler and the dexterous grasp evaluator DexEvaluator. DexSampler generates high-quality grasps conditioned on object point clouds by iterative denoising of randomly sampled grasps. We also introduce two grasp refinement strategies: Evaluator-Guided Diffusion (EGD) and Evaluator-based Sampling Refinement (ESR). The experiment results demonstrate that DexDiffuser consistently outperforms the state-of-the-art multi-finger grasp generation method FFHNet with an, on average, 9.12% and 19.44% higher grasp success rate in simulation and real robot experiments, respectively. Supplementary materials are available at https://yulihn.github.io/DexDiffuser_page/

Tarun Gupta, Wenbo Gong, Chao Ma et al.

Recent advances in foundation models, especially in large multi-modal models and conversational agents, have ignited interest in the potential of generally capable embodied agents. Such agents will require the ability to perform new tasks in many different real-world environments. However, current foundation models fail to accurately model physical interactions and are therefore insufficient for Embodied AI. The study of causality lends itself to the construction of veridical world models, which are crucial for accurately predicting the outcomes of possible interactions. This paper focuses on the prospects of building foundation world models for the upcoming generation of embodied agents and presents a novel viewpoint on the significance of causality within these. We posit that integrating causal considerations is vital to facilitating meaningful physical interactions with the world. Finally, we demystify misconceptions about causality in this context and present our outlook for future research.

Giovanni Luca Marchetti, Christopher Hillar, Danica Kragic et al.

In this work, we formally prove that, under certain conditions, if a neural network is invariant to a finite group then its weights recover the Fourier transform on that group. This provides a mathematical explanation for the emergence of Fourier features -- a ubiquitous phenomenon in both biological and artificial learning systems. The results hold even for non-commutative groups, in which case the Fourier transform encodes all the irreducible unitary group representations. Our findings have consequences for the problem of symmetry discovery. Specifically, we demonstrate that the algebraic structure of an unknown group can be recovered from the weights of a network that is at least approximately invariant within certain bounds. Overall, this work contributes to a foundation for an algebraic learning theory of invariant neural network representations.

Alberta Longhini, Marcel Büsching, Bardienus P. Duisterhof et al.

We introduce Cloth-Splatting, a method for estimating 3D states of cloth from RGB images through a prediction-update framework. Cloth-Splatting leverages an action-conditioned dynamics model for predicting future states and uses 3D Gaussian Splatting to update the predicted states. Our key insight is that coupling a 3D mesh-based representation with Gaussian Splatting allows us to define a differentiable map between the cloth state space and the image space. This enables the use of gradient-based optimization techniques to refine inaccurate state estimates using only RGB supervision. Our experiments demonstrate that Cloth-Splatting not only improves state estimation accuracy over current baselines but also reduces convergence time.

Katharina Friedl, Noémie Jaquier, Jens Lundell et al.

By incorporating physical consistency as inductive bias, deep neural networks display increased generalization capabilities and data efficiency in learning nonlinear dynamic models. However, the complexity of these models generally increases with the system dimensionality, requiring larger datasets, more complex deep networks, and significant computational effort. We propose a novel geometric network architecture to learn physically-consistent reduced-order dynamic parameters that accurately describe the original high-dimensional system behavior. This is achieved by building on recent advances in model-order reduction and by adopting a Riemannian perspective to jointly learn a non-linear structure-preserving latent space and the associated low-dimensional dynamics. Our approach enables accurate long-term predictions of the high-dimensional dynamics of rigid and deformable systems with increased data efficiency by inferring interpretable and physically-plausible reduced Lagrangian models.

Aude Billard, Alin Albu-Schaeffer, Michael Beetz et al.

AI technologies, including deep learning, large-language models have gone from one breakthrough to the other. As a result, we are witnessing growing excitement in robotics at the prospect of leveraging the potential of AI to tackle some of the outstanding barriers to the full deployment of robots in our daily lives. However, action and sensing in the physical world pose greater and different challenges than analysing data in isolation. As the development and application of AI in robotic products advances, it is important to reflect on which technologies, among the vast array of network architectures and learning models now available in the AI field, are most likely to be successfully applied to robots; how they can be adapted to specific robot designs, tasks, environments; which challenges must be overcome. This article offers an assessment of what AI for robotics has achieved since the 1990s and proposes a short- and medium-term research roadmap listing challenges and promises. These range from keeping up-to-date large datasets, representatives of a diversity of tasks robots may have to perform, and of environments they may encounter, to designing AI algorithms tailored specifically to robotics problems but generic enough to apply to a wide range of applications and transfer easily to a variety of robotic platforms. For robots to collaborate effectively with humans, they must predict human behavior without relying on bias-based profiling. Explainability and transparency in AI-driven robot control are not optional but essential for building trust, preventing misuse, and attributing responsibility in accidents. We close on what we view as the primary long-term challenges, that is, to design robots capable of lifelong learning, while guaranteeing safe deployment and usage, and sustainable computational costs.

Quantao Yang, Michael C. Welle, Danica Kragic et al.

Recent advances in skill learning has propelled robot manipulation to new heights by enabling it to learn complex manipulation tasks from a practical number of demonstrations. However, these skills are often limited to the particular action, object, and environment \textit{instances} that are shown in the training data, and have trouble transferring to other instances of the same category. In this work we present an open-vocabulary Spatial-Semantic Diffusion policy (S$^2$-Diffusion) which enables generalization from instance-level training data to category-level, enabling skills to be transferable between instances of the same category. We show that functional aspects of skills can be captured via a promptable semantic module combined with a spatial representation. We further propose leveraging depth estimation networks to allow the use of only a single RGB camera. Our approach is evaluated and compared on a diverse number of robot manipulation tasks, both in simulation and in the real world. Our results show that S$^2$-Diffusion is invariant to changes in category-irrelevant factors as well as enables satisfying performance on other instances within the same category, even if it was not trained on that specific instance. Project website: https://s2-diffusion.github.io.

Farzaneh Taleb, Miguel Vasco, Antônio H. Ribeiro et al.

The human brain encodes stimuli from the environment into representations that form a sensory perception of the world. Despite recent advances in understanding visual and auditory perception, olfactory perception remains an under-explored topic in the machine learning community due to the lack of large-scale datasets annotated with labels of human olfactory perception. In this work, we ask the question of whether pre-trained transformer models of chemical structures encode representations that are aligned with human olfactory perception, i.e., can transformers smell like humans? We demonstrate that representations encoded from transformers pre-trained on general chemical structures are highly aligned with human olfactory perception. We use multiple datasets and different types of perceptual representations to show that the representations encoded by transformer models are able to predict: (i) labels associated with odorants provided by experts; (ii) continuous ratings provided by human participants with respect to pre-defined descriptors; and (iii) similarity ratings between odorants provided by human participants. Finally, we evaluate the extent to which this alignment is associated with physicochemical features of odorants known to be relevant for olfactory decoding.

Aniss Aiman Medbouhi, Giovanni Luca Marchetti, Vladislav Polianskii et al.

Hyperbolic machine learning is an emerging field aimed at representing data with a hierarchical structure. However, there is a lack of tools for evaluation and analysis of the resulting hyperbolic data representations. To this end, we propose Hyperbolic Delaunay Geometric Alignment (HyperDGA) -- a similarity score for comparing datasets in a hyperbolic space. The core idea is counting the edges of the hyperbolic Delaunay graph connecting datapoints across the given sets. We provide an empirical investigation on synthetic and real-life biological data and demonstrate that HyperDGA outperforms the hyperbolic version of classical distances between sets. Furthermore, we showcase the potential of HyperDGA for evaluating latent representations inferred by a Hyperbolic Variational Auto-Encoder.

Wenjie Yin, Yi Yu, Hang Yin et al.

Current training of motion style transfer systems relies on consistency losses across style domains to preserve contents, hindering its scalable application to a large number of domains and private data. Recent image transfer works show the potential of independent training on each domain by leveraging implicit bridging between diffusion models, with the content preservation, however, limited to simple data patterns. We address this by imposing biased sampling in backward diffusion while maintaining the domain independence in the training stage. We construct the bias from the source domain keyframes and apply them as the gradient of content constraints, yielding a framework with keyframe manifold constraint gradients (KMCGs). Our validation demonstrates the success of training separate models to transfer between as many as ten dance motion styles. Comprehensive experiments find a significant improvement in preserving motion contents in comparison to baseline and ablative diffusion-based style transfer models. In addition, we perform a human study for a subjective assessment of the quality of generated dance motions. The results validate the competitiveness of KMCGs.

Donghu Kim, Youngdo Lee, Minho Park et al.

Reinforcement learning (RL) is a core approach for robot control when expert demonstrations are unavailable. On-policy methods such as Proximal Policy Optimization (PPO) are widely used for their stability, but their reliance on narrowly distributed on-policy data limits accurate policy evaluation in high-dimensional state and action spaces. Off-policy methods can overcome this limitation by learning from a broader state-action distribution, yet suffer from slow convergence and instability, as fitting a value function over diverse data requires many gradient updates, causing critic errors to accumulate through bootstrapping. We present FlashSAC, a fast and stable off-policy RL algorithm built on Soft Actor-Critic. Motivated by scaling laws observed in supervised learning, FlashSAC sharply reduces gradient updates while compensating with larger models and higher data throughput. To maintain stability at increased scale, FlashSAC explicitly bounds weight, feature, and gradient norms, curbing critic error accumulation. Across over 60 tasks in 10 simulators, FlashSAC consistently outperforms PPO and strong off-policy baselines in both final performance and training efficiency, with the largest gains on high-dimensional tasks such as dexterous manipulation. In sim-to-real humanoid locomotion, FlashSAC reduces training time from hours to minutes, demonstrating the promise of off-policy RL for sim-to-real transfer.

Nils Ingelhag, Jesper Munkeby, Michael C. Welle et al.

We present a Real-Time Operator Takeover (RTOT) paradigm enabling operators to seamlessly take control of a live visuomotor diffusion policy, guiding the system back into desirable states or reinforcing specific demonstrations. We present new insights in using the Mahalonobis distance to automatically identify undesirable states. Once the operator has intervened and redirected the system, the control is seamlessly returned to the policy, which resumes generating actions until further intervention is required. We demonstrate that incorporating the targeted takeover demonstrations significantly improves policy performance compared to training solely with an equivalent number of, but longer, initial demonstrations. We provide an in-depth analysis of using the Mahalanobis distance to detect out-of-distribution states, illustrating its utility for identifying critical failure points during execution. Supporting materials, including videos of initial and takeover demonstrations and all rice scooping experiments, are available on the project website: https://operator-takeover.github.io/

Daniel Marta, Simon Holk, Miguel Vasco et al.

Preference-based reinforcement learning (PbRL) is a suitable approach for style adaptation of pre-trained robotic behavior: adapting the robot's policy to follow human user preferences while still being able to perform the original task. However, collecting preferences for the adaptation process in robotics is often challenging and time-consuming. In this work we explore the adaptation of pre-trained robots in the low-preference-data regime. We show that, in this regime, recent adaptation approaches suffer from catastrophic reward forgetting (CRF), where the updated reward model overfits to the new preferences, leading the agent to become unable to perform the original task. To mitigate CRF, we propose to enhance the original reward model with a small number of parameters (low-rank matrices) responsible for modeling the preference adaptation. Our evaluation shows that our method can efficiently and effectively adjust robotic behavior to human preferences across simulation benchmark tasks and multiple real-world robotic tasks.

Chao Tang, Jiacheng Xu, Haofei Lu et al.

Building generalist robots capable of performing functional grasping in everyday, open-world environments remains a significant challenge due to the vast diversity of objects and tasks. Existing methods are either constrained to narrow object/task sets or rely on prohibitively large-scale data collection to capture real-world variability. In this work, we present an alternative approach, GraspDreamer, a method that leverages human demonstrations synthesized by visual generative models (VGMs) (e.g., video generation models) to enable zero-shot functional grasping without labor-intensive data collection. The key idea is that VGMs pre-trained on internet-scale human data implicitly encode generalized priors about how humans interact with the physical world, which can be combined with embodiment-specific action optimization to enable functional grasping with minimal effort. Extensive experiments on the public benchmarks with different robot hands demonstrate the superior data efficiency and generalization performance of GraspDreamer compared to previous methods. Real-world evaluations further validate the effectiveness on real robots. Additionally, we showcase that GraspDreamer can (1) be naturally extended to downstream manipulation tasks, and (2) can generate data to support visuomotor policy learning.

Yufei Duan, Hang Yin, Danica Kragic

Diffusion Policies have demonstrated impressive performance in robotic manipulation tasks. However, their long inference time, resulting from an extensive iterative denoising process, and the need to execute an action chunk before the next prediction to maintain consistent actions limit their applicability to latency-critical tasks or simple tasks with a short cycle time. While recent methods explored distillation or alternative policy structures to accelerate inference, these often demand additional training, which can be resource-intensive for large robotic models. In this paper, we introduce a novel approach inspired by the Real-Time Iteration (RTI) Scheme, a method from optimal control that accelerates optimization by leveraging solutions from previous time steps as initial guesses for subsequent iterations. We explore the application of this scheme in diffusion inference and propose a scaling-based method to effectively handle discrete actions, such as grasping, in robotic manipulation. The proposed scheme significantly reduces runtime computational costs without the need for distillation or policy redesign. This enables a seamless integration into many pre-trained diffusion-based models, in particular, to resource-demanding large models. We also provide theoretical conditions for the contractivity which could be useful for estimating the initial denoising step. Quantitative results from extensive simulation experiments show a substantial reduction in inference time, with comparable overall performance compared with Diffusion Policy using full-step denoising. Our project page with additional resources is available at: https://rti-dp.github.io/.

Haofei Lu, Yifei Dong, Zehang Weng et al.

We introduce the sequential multi-object robotic grasp sampling algorithm SeqGrasp that can robustly synthesize stable grasps on diverse objects using the robotic hand's partial Degrees of Freedom (DoF). We use SeqGrasp to construct the large-scale Allegro Hand sequential grasping dataset SeqDataset and use it for training the diffusion-based sequential grasp generator SeqDiffuser. We experimentally evaluate SeqGrasp and SeqDiffuser against the state-of-the-art non-sequential multi-object grasp generation method MultiGrasp in simulation and on a real robot. The experimental results demonstrate that SeqGrasp and SeqDiffuser reach an 8.71%-43.33% higher grasp success rate than MultiGrasp. Furthermore, SeqDiffuser is approximately 1000 times faster at generating grasps than SeqGrasp and MultiGrasp. Project page: https://yulihn.github.io/SeqGrasp/.

Nona Rajabi, Antônio H. Ribeiro, Miguel Vasco et al.

Decoding visual images from brain activity has significant potential for advancing brain-computer interaction and enhancing the understanding of human perception. Recent approaches align the representation spaces of images and brain activity to enable visual decoding. In this paper, we introduce the use of human-aligned image encoders to map brain signals to images. We hypothesize that these models more effectively capture perceptual attributes associated with the rapid visual stimuli presentations commonly used in visual brain data recording experiments. Our empirical results support this hypothesis, demonstrating that this simple modification improves image retrieval accuracy by up to 21% compared to state-of-the-art methods. Comprehensive experiments confirm consistent performance improvements across diverse EEG architectures, image encoders, alignment methods, participants, and brain imaging modalities

Alfredo Reichlin, Miguel Vasco, Hang Yin et al.

We address the problem of learning optimal behavior from sub-optimal datasets for goal-conditioned offline reinforcement learning. To do so, we propose the use of metric learning to approximate the optimal value function for goal-conditioned offline RL problems under sparse rewards, invertible actions and deterministic transitions. We introduce distance monotonicity, a property for representations to recover optimality and propose an optimization objective that leads to such property. We use the proposed value function to guide the learning of a policy in an actor-critic fashion, a method we name MetricRL. Experimentally, we show that our method estimates optimal behaviors from severely sub-optimal offline datasets without suffering from out-of-distribution estimation errors. We demonstrate that MetricRL consistently outperforms prior state-of-the-art goal-conditioned RL methods in learning optimal policies from sub-optimal offline datasets.

Aniss Aiman Medbouhi, Alejandro García-Castellanos, Giovanni Luca Marchetti et al.

We study the problem of constructing Steiner Minimal Trees (SMTs) in hyperbolic space. Exact SMT computation is NP-hard, and existing hyperbolic heuristics such as HyperSteiner are deterministic and often get trapped in locally suboptimal configurations. We introduce Randomized HyperSteiner (RHS), a stochastic Delaunay triangulation heuristic that incorporates randomness into the expansion process and refines candidate trees via Riemannian gradient descent optimization. Experiments on synthetic data sets and a real-world single-cell transcriptomic data show that RHS outperforms Minimum Spanning Tree (MST), Neighbour Joining, and vanilla HyperSteiner (HS). In near-boundary configurations, RHS can achieve a 32% reduction in total length over HS, demonstrating its effectiveness and robustness in diverse data regimes.

Katharina Friedl, Noémie Jaquier, Mika Liao et al.

By embedding physical intuition, network architectures enforce fundamental properties, such as energy conservation laws, leading to plausible predictions. Yet, scaling these models to intrinsically high-dimensional systems remains a significant challenge. This paper introduces Geometric Reduced-order Hamiltonian Neural Network (RO-HNN), a novel physics-inspired neural network that combines the conservation laws of Hamiltonian mechanics with the scalability of model order reduction. RO-HNN is built on two core components: a novel geometrically-constrained symplectic autoencoder that learns a low-dimensional, structure-preserving symplectic submanifold, and a geometric Hamiltonian neural network that models the dynamics on the submanifold. Our experiments demonstrate that RO-HNN provides physically-consistent, stable, and generalizable predictions of complex high-dimensional dynamics, thereby effectively extending the scope of Hamiltonian neural networks to high-dimensional physical systems.

Holly Dinkel, Marcel Büsching, Alberta Longhini et al.

This work presents DLO-Splatting, an algorithm for estimating the 3D shape of Deformable Linear Objects (DLOs) from multi-view RGB images and gripper state information through prediction-update filtering. The DLO-Splatting algorithm uses a position-based dynamics model with shape smoothness and rigidity dampening corrections to predict the object shape. Optimization with a 3D Gaussian Splatting-based rendering loss iteratively renders and refines the prediction to align it with the visual observations in the update step. Initial experiments demonstrate promising results in a knot tying scenario, which is challenging for existing vision-only methods.

Hannah Kuehn, Joseph La Delfa, Miguel Vasco et al.

This paper introduces the concept of coexistence for embodied artificial agents and argues that it is a prerequisite for long-term, in-the-wild interaction with humans. Contemporary embodied artificial agents excel in static, predefined tasks but fall short in dynamic and long-term interactions with humans. On the other hand, humans can adapt and evolve continuously, exploiting the situated knowledge embedded in their environment and other agents, thus contributing to meaningful interactions. We take an interdisciplinary approach at different levels of organization, drawing from biology and design theory, to understand how human and non-human organisms foster entities that coexist within their specific environments. Finally, we propose key research directions for the artificial intelligence community to develop coexisting embodied agents, focusing on the principles, hardware and learning methods responsible for shaping them.

Wenjie Yin, Xuejiao Zhao, Yi Yu et al.

Dance typically involves professional choreography with complex movements that follow a musical rhythm and can also be influenced by lyrical content. The integration of lyrics in addition to the auditory dimension, enriches the foundational tone and makes motion generation more amenable to its semantic meanings. However, existing dance synthesis methods tend to model motions only conditioned on audio signals. In this work, we make two contributions to bridge this gap. First, we propose LM2D, a novel probabilistic architecture that incorporates a multimodal diffusion model with consistency distillation, designed to create dance conditioned on both music and lyrics in one diffusion generation step. Second, we introduce the first 3D dance-motion dataset that encompasses both music and lyrics, obtained with pose estimation technologies. We evaluate our model against music-only baseline models with objective metrics and human evaluations, including dancers and choreographers. The results demonstrate LM2D is able to produce realistic and diverse dance matching both lyrics and music. A video summary can be accessed at: https://youtu.be/4XCgvYookvA.

Reza Dadfar, Sanaz Sabzevari, Mårten Björkman et al.

Language-based fashion image editing allows users to try out variations of desired garments through provided text prompts. Inspired by research on manipulating latent representations in StyleCLIP and HairCLIP, we focus on these latent spaces for editing fashion items of full-body human datasets. Currently, there is a gap in handling fashion image editing due to the complexity of garment shapes and textures and the diversity of human poses. In this paper, we propose an editing optimizer scheme method called Text-Driven Garment Editing Mapper (TD-GEM), aiming to edit fashion items in a disentangled way. To this end, we initially obtain a latent representation of an image through generative adversarial network inversions such as Encoder for Editing (e4e) or Pivotal Tuning Inversion (PTI) for more accurate results. An optimization-based Contrastive Language-Image Pre-training (CLIP) is then utilized to guide the latent representation of a fashion image in the direction of a target attribute expressed in terms of a text prompt. Our TD-GEM manipulates the image accurately according to the target attribute, while other parts of the image are kept untouched. In the experiments, we evaluate TD-GEM on two different attributes (i.e., "color" and "sleeve length"), which effectively generates realistic images compared to the recent manipulation schemes.

Petra Poklukar, Vladislav Polianskii, Anastasia Varava et al.

Advanced representation learning techniques require reliable and general evaluation methods. Recently, several algorithms based on the common idea of geometric and topological analysis of a manifold approximated from the learned data representations have been proposed. In this work, we introduce Delaunay Component Analysis (DCA) - an evaluation algorithm which approximates the data manifold using a more suitable neighbourhood graph called Delaunay graph. This provides a reliable manifold estimation even for challenging geometric arrangements of representations such as clusters with varying shape and density as well as outliers, which is where existing methods often fail. Furthermore, we exploit the nature of Delaunay graphs and introduce a framework for assessing the quality of individual novel data representations. We experimentally validate the proposed DCA method on representations obtained from neural networks trained with contrastive objective, supervised and generative models, and demonstrate various use cases of our extended single point evaluation framework.