Tamim Asfour

Yahav Avigal, Lars Berscheid, Tamim Asfour et al.

Folding garments reliably and efficiently is a long standing challenge in robotic manipulation due to the complex dynamics and high dimensional configuration space of garments. An intuitive approach is to initially manipulate the garment to a canonical smooth configuration before folding. In this work, we develop SpeedFolding, a reliable and efficient bimanual system, which given user-defined instructions as folding lines, manipulates an initially crumpled garment to (1) a smoothed and (2) a folded configuration. Our primary contribution is a novel neural network architecture that is able to predict pairs of gripper poses to parameterize a diverse set of bimanual action primitives. After learning from 4300 human-annotated and self-supervised actions, the robot is able to fold garments from a random initial configuration in under 120s on average with a success rate of 93%. Real-world experiments show that the system is able to generalize to unseen garments of different color, shape, and stiffness. While prior work achieved 3-6 Folds Per Hour (FPH), SpeedFolding achieves 30-40 FPH.

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.

Leonard Bärmann, Rainer Kartmann, Fabian Peller-Konrad et al.

Natural-language dialog is key for intuitive human-robot interaction. It can be used not only to express humans' intents, but also to communicate instructions for improvement if a robot does not understand a command correctly. Of great importance is to endow robots with the ability to learn from such interaction experience in an incremental way to allow them to improve their behaviors or avoid mistakes in the future. In this paper, we propose a system to achieve incremental learning of complex behavior from natural interaction, and demonstrate its implementation on a humanoid robot. Building on recent advances, we present a system that deploys Large Language Models (LLMs) for high-level orchestration of the robot's behavior, based on the idea of enabling the LLM to generate Python statements in an interactive console to invoke both robot perception and action. The interaction loop is closed by feeding back human instructions, environment observations, and execution results to the LLM, thus informing the generation of the next statement. Specifically, we introduce incremental prompt learning, which enables the system to interactively learn from its mistakes. For that purpose, the LLM can call another LLM responsible for code-level improvements of the current interaction based on human feedback. The improved interaction is then saved in the robot's memory, and thus retrieved on similar requests. We integrate the system in the robot cognitive architecture of the humanoid robot ARMAR-6 and evaluate our methods both quantitatively (in simulation) and qualitatively (in simulation and real-world) by demonstrating generalized incrementally-learned knowledge.

Fabian Peller-Konrad, Rainer Kartmann, Christian R. G. Dreher et al.

Cognitive agents such as humans and robots perceive their environment through an abundance of sensors producing streams of data that need to be processed to generate intelligent behavior. A key question of cognition-enabled and AI-driven robotics is how to organize and manage knowledge efficiently in a cognitive robot control architecture. We argue, that memory is a central active component of such architectures that mediates between semantic and sensorimotor representations, orchestrates the flow of data streams and events between different processes and provides the components of a cognitive architecture with data-driven services for the abstraction of semantics from sensorimotor data, the parametrization of symbolic plans for execution and prediction of action effects. Based on related work, and the experience gained in developing our ARMAR humanoid robot systems, we identified conceptual and technical requirements of a memory system as central component of cognitive robot control architecture that facilitate the realization of high-level cognitive abilities such as explaining, reasoning, prospection, simulation and augmentation. Conceptually, a memory should be active, support multi-modal data representations, associate knowledge, be introspective, and have an inherently episodic structure. Technically, the memory should support a distributed design, be access-efficient and capable of long-term data storage. We introduce the memory system for our cognitive robot control architecture and its implementation in the robot software framework ArmarX. We evaluate the efficiency of the memory system with respect to transfer speeds, compression, reproduction and prediction capabilities.

Jongseok Lee, Ribin Balachandran, Konstantin Kondak et al.

This article presents a novel telepresence system for advancing aerial manipulation in dynamic and unstructured environments. The proposed system not only features a haptic device, but also a virtual reality (VR) interface that provides real-time 3D displays of the robot's workspace as well as a haptic guidance to its remotely located operator. To realize this, multiple sensors namely a LiDAR, cameras and IMUs are utilized. For processing of the acquired sensory data, pose estimation pipelines are devised for industrial objects of both known and unknown geometries. We further propose an active learning pipeline in order to increase the sample efficiency of a pipeline component that relies on Deep Neural Networks (DNNs) based object detection. All these algorithms jointly address various challenges encountered during the execution of perception tasks in industrial scenarios. In the experiments, exhaustive ablation studies are provided to validate the proposed pipelines. Methodologically, these results commonly suggest how an awareness of the algorithms' own failures and uncertainty (`introspection') can be used tackle the encountered problems. Moreover, outdoor experiments are conducted to evaluate the effectiveness of the overall system in enhancing aerial manipulation capabilities. In particular, with flight campaigns over days and nights, from spring to winter, and with different users and locations, we demonstrate over 70 robust executions of pick-and-place, force application and peg-in-hole tasks with the DLR cable-Suspended Aerial Manipulator (SAM). As a result, we show the viability of the proposed system in future industrial applications.

Noémie Jaquier, Leonel Rozo, Miguel González-Duque et al.

Human motion taxonomies serve as high-level hierarchical abstractions that classify how humans move and interact with their environment. They have proven useful to analyse grasps, manipulation skills, and whole-body support poses. Despite substantial efforts devoted to design their hierarchy and underlying categories, their use remains limited. This may be attributed to the lack of computational models that fill the gap between the discrete hierarchical structure of the taxonomy and the high-dimensional heterogeneous data associated to its categories. To overcome this problem, we propose to model taxonomy data via hyperbolic embeddings that capture the associated hierarchical structure. We achieve this by formulating a novel Gaussian process hyperbolic latent variable model that incorporates the taxonomy structure through graph-based priors on the latent space and distance-preserving back constraints. We validate our model on three different human motion taxonomies to learn hyperbolic embeddings that faithfully preserve the original graph structure. We show that our model properly encodes unseen data from existing or new taxonomy categories, and outperforms its Euclidean and VAE-based counterparts. Finally, through proof-of-concept experiments, we show that our model may be used to generate realistic trajectories between the learned embeddings.

Jannik Steinbring, Christian Mandery, Nikolaus Vahrenkamp et al.

We present a new online approach to track human whole-body motion from motion capture data, i.e., positions of labeled markers attached to the human body. Tracking in noisy data can be effectively performed with the aid of well-established recursive state estimation techniques. This allows us to systematically take noise of the marker measurements into account. However, as joint limits imposed by the human body have to be satisfied during estimation, first we transform this constrained estimation problem into an unconstrained one by using periodic functions. Then, we apply the Smart Sampling Kalman Filter to solve this unconstrained estimation problem. The proposed recursive state estimation approach makes the human motion tracking very robust to partial occlusion of markers and avoids any special treatment or reconstruction of the missed markers. A concrete implementation built on the kinematic human reference model of the Master Motor Map framework and a Vicon motion capture system is evaluated. Different captured motions show that our implementation can accurately estimate whole-body human motion in real-time and outperforms existing gradient-based approaches. In addition, we demonstrate its ability to smoothly handle incomplete marker data.

Noémie Jaquier, Leonel Rozo, Tamim Asfour

In the realm of robotics, numerous downstream robotics tasks leverage machine learning methods for processing, modeling, or synthesizing data. Often, this data comprises variables that inherently carry geometric constraints, such as the unit-norm condition of quaternions representing rigid-body orientations or the positive definiteness of stiffness and manipulability ellipsoids. Handling such geometric constraints effectively requires the incorporation of tools from differential geometry into the formulation of machine learning methods. In this context, Riemannian manifolds emerge as a powerful mathematical framework to handle such geometric constraints. Nevertheless, their recent adoption in robot learning has been largely characterized by a mathematically-flawed simplification, hereinafter referred to as the "single tangent space fallacy". This approach involves merely projecting the data of interest onto a single tangent (Euclidean) space, over which an off-the-shelf learning algorithm is applied. This paper provides a theoretical elucidation of various misconceptions surrounding this approach and offers experimental evidence of its shortcomings. Finally, it presents valuable insights to promote best practices when employing Riemannian geometry within robot learning applications.

Marta Lagomarsino, Elena Merlo, Andrea Pupa et al.

Remarkable capabilities have been achieved by robotics and AI, mastering complex tasks and environments. Yet, humans often remain passive observers, fascinated but uncertain how to engage. Robots, in turn, cannot reach their full potential in human-populated environments without effectively modeling human states and intentions and adapting their behavior. To achieve a synergistic human-robot collaboration (HRC), a continuous information flow should be established: humans must intuitively communicate instructions, share expertise, and express needs. In parallel, robots must clearly convey their internal state and forthcoming actions to keep users informed, comfortable, and in control. This review identifies and connects key components enabling intuitive information exchange and skill transfer between humans and robots. We examine the full interaction pipeline: from the human-to-robot communication bridge translating multimodal inputs into robot-understandable representations, through adaptive planning and role allocation, to the control layer and feedback mechanisms to close the loop. Finally, we highlight trends and promising directions toward more adaptive, accessible HRC.

Leonard Bärmann, Chad DeChant, Joana Plewnia et al.

Verbalization of robot experience, i.e., summarization of and question answering about a robot's past, is a crucial ability for improving human-robot interaction. Previous works applied rule-based systems or fine-tuned deep models to verbalize short (several-minute-long) streams of episodic data, limiting generalization and transferability. In our work, we apply large pretrained models to tackle this task with zero or few examples, and specifically focus on verbalizing life-long experiences. For this, we derive a tree-like data structure from episodic memory (EM), with lower levels representing raw perception and proprioception data, and higher levels abstracting events to natural language concepts. Given such a hierarchical representation built from the experience stream, we apply a large language model as an agent to interactively search the EM given a user's query, dynamically expanding (initially collapsed) tree nodes to find the relevant information. The approach keeps computational costs low even when scaling to months of robot experience data. We evaluate our method on simulated household robot data, human egocentric videos, and real-world robot recordings, demonstrating its flexibility and scalability.

Leonard Bärmann, Joana Plewnia, Alex Waibel et al.

Robots must verbalize their past experiences when users ask "Where did you put my keys?" or "Why did the task fail?" Yet maintaining life-long episodic memory (EM) from continuous multimodal perception quickly exceeds storage limits and makes real-time query impractical, calling for selective forgetting that adapts to users' notions of relevance. We present H$^2$-EMV, a framework enabling humanoids to learn what to remember through user interaction. Our approach incrementally constructs hierarchical EM, selectively forgets using language-model-based relevance estimation conditioned on learned natural-language rules, and updates these rules given user feedback about forgotten details. Evaluations on simulated household tasks and 20.5-hour-long real-world recordings from ARMAR-7 demonstrate that H$^2$-EMV maintains question-answering accuracy while reducing memory size by 45% and query-time compute by 35%. Critically, performance improves over time - accuracy increases 70% in second-round queries by adapting to user-specific priorities - demonstrating that learned forgetting enables scalable, personalized EM for long-term human-robot collaboration.

Andrea Testa, Søren Hauberg, Tamim Asfour et al.

The Schrödinger Bridge provides a principled framework for modeling stochastic processes between distributions; however, existing methods are limited by energy-conservation assumptions, which constrains the bridge's shape preventing it from model varying-energy phenomena. To overcome this, we introduce the non-conservative generalized Schrödinger bridge (NCGSB), a novel, energy-varying reformulation based on contact Hamiltonian mechanics. By allowing energy to change over time, the NCGSB provides a broader class of real-world stochastic processes, capturing richer and more faithful intermediate dynamics. By parameterizing the Wasserstein manifold, we lift the bridge problem to a tractable geodesic computation in a finite-dimensional space. Unlike computationally expensive iterative solutions, our contact Wasserstein geodesic (CWG) is naturally implemented via a ResNet architecture and relies on a non-iterative solver with near-linear complexity. Furthermore, CWG supports guided generation by modulating a task-specific distance metric. We validate our framework on tasks including manifold navigation, molecular dynamics predictions, and image generation, demonstrating its practical benefits and versatility.

Markus Grotz, Mohit Shridhar, Tamim Asfour et al.

Bimanual manipulation is challenging due to precise spatial and temporal coordination required between two arms. While there exist several real-world bimanual systems, there is a lack of simulated benchmarks with a large task diversity for systematically studying bimanual capabilities across a wide range of tabletop tasks. This paper addresses the gap by extending RLBench to bimanual manipulation. We open-source our code and benchmark comprising 13 new tasks with 23 unique task variations, each requiring a high degree of coordination and adaptability. To kickstart the benchmark, we extended several state-of-the art methods to bimanual manipulation and also present a language-conditioned behavioral cloning agent -- PerAct2, which enables the learning and execution of bimanual 6-DoF manipulation tasks. Our novel network architecture efficiently integrates language processing with action prediction, allowing robots to understand and perform complex bimanual tasks in response to user-specified goals. Project website with code is available at: http://bimanual.github.io

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

Max Braun, Noémie Jaquier, Leonel Rozo et al.

We introduce Riemannian Flow Matching Policies (RFMP), a novel model for learning and synthesizing robot visuomotor policies. RFMP leverages the efficient training and inference capabilities of flow matching methods. By design, RFMP inherits the strengths of flow matching: the ability to encode high-dimensional multimodal distributions, commonly encountered in robotic tasks, and a very simple and fast inference process. We demonstrate the applicability of RFMP to both state-based and vision-conditioned robot motion policies. Notably, as the robot state resides on a Riemannian manifold, RFMP inherently incorporates geometric awareness, which is crucial for realistic robotic tasks. To evaluate RFMP, we conduct two proof-of-concept experiments, comparing its performance against Diffusion Policies. Although both approaches successfully learn the considered tasks, our results show that RFMP provides smoother action trajectories with significantly lower inference times.

Timo Birr, Christoph Pohl, Abdelrahman Younes et al.

Recent advances in task planning leverage Large Language Models (LLMs) to improve generalizability by combining such models with classical planning algorithms to address their inherent limitations in reasoning capabilities. However, these approaches face the challenge of dynamically capturing the initial state of the task planning problem. To alleviate this issue, we propose AutoGPT+P, a system that combines an affordance-based scene representation with a planning system. Affordances encompass the action possibilities of an agent on the environment and objects present in it. Thus, deriving the planning domain from an affordance-based scene representation allows symbolic planning with arbitrary objects. AutoGPT+P leverages this representation to derive and execute a plan for a task specified by the user in natural language. In addition to solving planning tasks under a closed-world assumption, AutoGPT+P can also handle planning with incomplete information, e. g., tasks with missing objects by exploring the scene, suggesting alternatives, or providing a partial plan. The affordance-based scene representation combines object detection with an automatically generated object-affordance-mapping using ChatGPT. The core planning tool extends existing work by automatically correcting semantic and syntactic errors. Our approach achieves a success rate of 98%, surpassing the current 81% success rate of the current state-of-the-art LLM-based planning method SayCan on the SayCan instruction set. Furthermore, we evaluated our approach on our newly created dataset with 150 scenarios covering a wide range of complex tasks with missing objects, achieving a success rate of 79% on our dataset. The dataset and the code are publicly available at https://git.h2t.iar.kit.edu/birr/autogpt-p-standalone.

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.

Andrea Testa, Søren Hauberg, Tamim Asfour et al.

Accurately modeling and predicting complex dynamical systems, particularly those involving force exchange and dissipation, is crucial for applications ranging from fluid dynamics to robotics, but presents significant challenges due to the intricate interplay of geometric constraints and energy transfer. This paper introduces Geometric Contact Flows (GFC), a novel framework leveraging Riemannian and Contact geometry as inductive biases to learn such systems. GCF constructs a latent contact Hamiltonian model encoding desirable properties like stability or energy conservation. An ensemble of contactomorphisms then adapts this model to the target dynamics while preserving these properties. This ensemble allows for uncertainty-aware geodesics that attract the system's behavior toward the data support, enabling robust generalization and adaptation to unseen scenarios. Experiments on learning dynamics for physical systems and for controlling robots on interaction tasks demonstrate the effectiveness of our approach.

Luis Augenstein, Noémie Jaquier, Tamim Asfour et al.

Probabilistic Latent Variable Models (LVMs) excel at modeling complex, high-dimensional data through lower-dimensional representations. Recent advances show that equipping these latent representations with a Riemannian metric unlocks geometry-aware distances and shortest paths that comply with the underlying data structure. This paper focuses on hyperbolic embeddings, a particularly suitable choice for modeling hierarchical relationships. Previous approaches relying on hyperbolic geodesics for interpolating the latent space often generate paths crossing low-data regions, leading to highly uncertain predictions. Instead, we propose augmenting the hyperbolic manifold with a pullback metric to account for distortions introduced by the LVM's nonlinear mapping and provide a complete development for pullback metrics of Gaussian Process LVMs (GPLVMs). Our experiments demonstrate that geodesics on the pullback metric not only respect the geometry of the hyperbolic latent space but also align with the underlying data distribution, significantly reducing uncertainty in predictions.

Xi Huang, Hongyi Zhou, Ge Li et al.

We propose MoRe-ERL, a framework that combines Episodic Reinforcement Learning (ERL) and residual learning, which refines preplanned reference trajectories into safe, feasible, and efficient task-specific trajectories. This framework is general enough to incorporate into arbitrary ERL methods and motion generators seamlessly. MoRe-ERL identifies trajectory segments requiring modification while preserving critical task-related maneuvers. Then it generates smooth residual adjustments using B-Spline-based movement primitives to ensure adaptability to dynamic task contexts and smoothness in trajectory refinement. Experimental results demonstrate that residual learning significantly outperforms training from scratch using ERL methods, achieving superior sample efficiency and task performance. Hardware evaluations further validate the framework, showing that policies trained in simulation can be directly deployed in real-world systems, exhibiting a minimal sim-to-real gap.

Harisankar Babu, Philipp Schillinger, Tamim Asfour

We introduce TAPAS (Task-based Adaptation and Planning using AgentS), a multi-agent framework that integrates Large Language Models (LLMs) with symbolic planning to solve complex tasks without the need for manually defined environment models. TAPAS employs specialized LLM-based agents that collaboratively generate and adapt domain models, initial states, and goal specifications as needed using structured tool-calling mechanisms. Through this tool-based interaction, downstream agents can request modifications from upstream agents, enabling adaptation to novel attributes and constraints without manual domain redefinition. A ReAct (Reason+Act)-style execution agent, coupled with natural language plan translation, bridges the gap between dynamically generated plans and real-world robot capabilities. TAPAS demonstrates strong performance in benchmark planning domains and in the VirtualHome simulated real-world environment.

Niklas Hemken, Florian Jacob, Fabian Peller-Konrad et al.

Humanoid robots will be able to assist humans in their daily life, in particular due to their versatile action capabilities. However, while these robots need a certain degree of autonomy to learn and explore, they also should respect various constraints, for access control and beyond. We explore the novel field of incorporating privacy, security, and access control constraints with robot task planning approaches. We report preliminary results on the classical symbolic approach, deep-learned neural networks, and modern ideas using large language models as knowledge base. From analyzing their trade-offs, we conclude that a hybrid approach is necessary, and thereby present a new use case for the emerging field of neuro-symbolic artificial intelligence.

Christian Dreher, Patrick Dormanns, Andre Meixner et al.

Temporal task structure is fundamental for bimanual manipulation: a robot must not only know that one action precedes or overlaps another, but also when each action should occur and how long it should take. While symbolic temporal relations enable high-level reasoning about task structure and alternative execution sequences, concrete timing parameters are equally essential for coordinating two hands at the execution level. Existing approaches address these two levels in isolation, leaving a gap between high-level task planning and low-level movement synchronization. This work presents an approach for learning both symbolic and subsymbolic temporal task constraints from human demonstrations and deriving executable, temporally parametrized plans for bimanual manipulation. Our contributions are (i) a 3-dimensional representation of timings between two actions with methods based on multivariate Gaussian Mixture Models to represent temporal relationships between actions on a subsymbolic level, (ii) a method based on the Davis-Putnam-Logemann-Loveland (DPLL) algorithm that finds and ranks all contradiction-free assignments of Allen relations to action pairs, representing different modes of a task, and (iii) an optimization-based planning system that combines the identified symbolic and subsymbolic temporal task constraints to derive temporally parametrized plans for robot execution. We evaluate our approach on several datasets, demonstrating that our method generates temporally parametrized plans closer to human demonstrations than the most characteristic demonstration baseline.

Luis Augenstein, Noémie Jaquier, Tamim Asfour et al.

Human-like motion generation for robots often draws inspiration from biomechanical studies, which often categorize complex human motions into hierarchical taxonomies. While these taxonomies provide rich structural information about how movements relate to one another, this information is frequently overlooked in motion generation models, leading to a disconnect between the generated motions and their underlying hierarchical structure. This paper introduces the \ac{gphdm}, a novel approach that learns latent representations preserving both the hierarchical structure of motions and their temporal dynamics to ensure physical consistency. Our model achieves this by extending the dynamics prior of the Gaussian Process Dynamical Model (GPDM) to the hyperbolic manifold and integrating it with taxonomy-aware inductive biases. Building on this geometry- and taxonomy-aware frameworks, we propose three novel mechanisms for generating motions that are both taxonomically-structured and physically-consistent: two probabilistic recursive approaches and a method based on pullback-metric geodesics. Experiments on generating realistic motion sequences on the hand grasping taxonomy show that the proposed GPHDM faithfully encodes the underlying taxonomy and temporal dynamics, and generates novel physically-consistent trajectories.

Noah Geiger, Tamim Asfour, Neville Hogan et al.

Learning methods excel at motion generation in the information domain but are not primarily designed for physical interaction in the energy domain. Impedance Control shapes physical interaction but requires task-aware tuning by selecting feasible impedance parameters. We present Diffusion-Based Impedance Learning, a framework that combines both domains. A Transformer-based Diffusion Model with cross-attention to external wrenches reconstructs a simulated Zero-Force Trajectory (sZFT). This captures both translational and rotational task-space behavior. For rotations, we introduce a novel SLERP-based quaternion noise scheduler that ensures geometric consistency. The reconstructed sZFT is then passed to an energy-based estimator that updates stiffness and damping parameters. A directional rule is applied that reduces impedance along non task axes while preserving rigidity along task directions. Training data were collected for a parkour scenario and robotic-assisted therapy tasks using teleoperation with Apple Vision Pro. With only tens of thousands of samples, the model achieved sub-millimeter positional accuracy and sub-degree rotational accuracy. Its compact model size enabled real-time torque control and autonomous stiffness adaptation on a KUKA LBR iiwa robot. The controller achieved smooth parkour traversal within force and velocity limits and 30/30 success rates for cylindrical, square, and star peg insertions without any peg-specific demonstrations in the training data set. All code for the Transformer-based Diffusion Model, the robot controller, and the Apple Vision Pro telemanipulation framework is publicly available. These results mark an important step towards Physical AI, fusing model-based control for physical interaction with learning-based methods for trajectory generation.

Tu Anh Dinh, Carlos Mullov, Leonard Bärmann et al.

With the rapid development of Large Language Models (LLMs), it is crucial to have benchmarks which can evaluate the ability of LLMs on different domains. One common use of LLMs is performing tasks on scientific topics, such as writing algorithms, querying databases or giving mathematical proofs. Inspired by the way university students are evaluated on such tasks, in this paper, we propose SciEx - a benchmark consisting of university computer science exam questions, to evaluate LLMs ability on solving scientific tasks. SciEx is (1) multilingual, containing both English and German exams, and (2) multi-modal, containing questions that involve images, and (3) contains various types of freeform questions with different difficulty levels, due to the nature of university exams. We evaluate the performance of various state-of-the-art LLMs on our new benchmark. Since SciEx questions are freeform, it is not straightforward to evaluate LLM performance. Therefore, we provide human expert grading of the LLM outputs on SciEx. We show that the free-form exams in SciEx remain challenging for the current LLMs, where the best LLM only achieves 59.4\% exam grade on average. We also provide detailed comparisons between LLM performance and student performance on SciEx. To enable future evaluation of new LLMs, we propose using LLM-as-a-judge to grade the LLM answers on SciEx. Our experiments show that, although they do not perform perfectly on solving the exams, LLMs are decent as graders, achieving 0.948 Pearson correlation with expert grading.

Noémie Jaquier, Viacheslav Borovitskiy, Andrei Smolensky et al.

Bayesian optimization is a data-efficient technique which can be used for control parameter tuning, parametric policy adaptation, and structure design in robotics. Many of these problems require optimization of functions defined on non-Euclidean domains like spheres, rotation groups, or spaces of positive-definite matrices. To do so, one must place a Gaussian process prior, or equivalently define a kernel, on the space of interest. Effective kernels typically reflect the geometry of the spaces they are defined on, but designing them is generally non-trivial. Recent work on the Riemannian Matérn kernels, based on stochastic partial differential equations and spectral theory of the Laplace-Beltrami operator, offers promising avenues towards constructing such geometry-aware kernels. In this paper, we study techniques for implementing these kernels on manifolds of interest in robotics, demonstrate their performance on a set of artificial benchmark functions, and illustrate geometry-aware Bayesian optimization for a variety of robotic applications, covering orientation control, manipulability optimization, and motion planning, while showing its improved performance.

You Zhou, Jianfeng Gao, Tamim Asfour

One challenge of motion generation using robot learning from demonstration techniques is that human demonstrations follow a distribution with multiple modes for one task query. Previous approaches fail to capture all modes or tend to average modes of the demonstrations and thus generate invalid trajectories. The other difficulty is the small number of demonstrations that cannot cover the entire working space. To overcome this problem, a motion generation model with extrapolation ability is needed. Previous works restrict task queries as local frames and learn representations in local frames. We propose a model to solve both problems. For multiple modes, we suggest to learn local latent representations of motion trajectories with a density estimation method based on real-valued non-volume preserving (RealNVP) transformations that provides a set of powerful, stably invertible, and learnable transformations. To improve the extrapolation ability, we propose to shift the attention of the robot from one local frame to another during the task execution. In experiments, we consider the docking problem used also in previous works where a trajectory has to be generated to connect two dockers without collision. We increase complexity of the task and show that the proposed method outperforms other approaches. In addition, we evaluate the approach in real robot experiments.

Ahmet E. Tekden, Aykut Erdem, Erkut Erdem et al.

Pushing is an essential non-prehensile manipulation skill used for tasks ranging from pre-grasp manipulation to scene rearrangement, reasoning about object relations in the scene, and thus pushing actions have been widely studied in robotics. The effective use of pushing actions often requires an understanding of the dynamics of the manipulated objects and adaptation to the discrepancies between prediction and reality. For this reason, effect prediction and parameter estimation with pushing actions have been heavily investigated in the literature. However, current approaches are limited because they either model systems with a fixed number of objects or use image-based representations whose outputs are not very interpretable and quickly accumulate errors. In this paper, we propose a graph neural network based framework for effect prediction and parameter estimation of pushing actions by modeling object relations based on contacts or articulations. Our framework is validated both in real and simulated environments containing different shaped multi-part objects connected via different types of joints and objects with different masses, and it outperforms image-based representations on physics prediction. Our approach enables the robot to predict and adapt the effect of a pushing action as it observes the scene. It can also be used for tool manipulation with never-seen tools. Further, we demonstrate 6D effect prediction in the lever-up action in the context of robot-based hard-disk disassembly.

Cheng-Yu Kuo, Andreas Schaarschmidt, Yunduan Cui et al.

This letter presents contact-safe Model-based Reinforcement Learning (MBRL) for robot applications that achieves contact-safe behaviors in the learning process. In typical MBRL, we cannot expect the data-driven model to generate accurate and reliable policies to the intended robotic tasks during the learning process due to sample scarcity. Operating these unreliable policies in a contact-rich environment could cause damage to the robot and its surroundings. To alleviate the risk of causing damage through unexpected intensive physical contacts, we present the contact-safe MBRL that associates the probabilistic Model Predictive Control's (pMPC) control limits with the model uncertainty so that the allowed acceleration of controlled behavior is adjusted according to learning progress. Control planning with such uncertainty-aware control limits is formulated as a deterministic MPC problem using a computation-efficient approximated GP dynamics and an approximated inference technique. Our approach's effectiveness is evaluated through bowl mixing tasks with simulated and real robots, scooping tasks with a real robot as examples of contact-rich manipulation skills. (video: https://youtu.be/sdhHP3NhYi0)

Felix Hundhausen, Julia Starke, Tamim Asfour

We present a novel underactued humanoid five finger soft hand, the KIT \softhand, which is equipped with cameras in the fingertips and integrates a high performance embedded system for visual processing and control. We describe the actuation mechanism of the hand and the tendon-driven soft finger design with internally routed high-bandwidth flat-flex cables. For efficient on-board parallel processing of visual data from the cameras in each fingertip, we present a hybrid embedded architecture consisting of a field programmable logic array (FPGA) and a microcontroller that allows the realization of visual object segmentation based on convolutional neural networks. We evaluate the hand design by conducting durability experiments with one finger and quantify the grasp performance in terms of grasping force, speed and grasp success. The results show that the hand exhibits a grasp force of 31.8 N and a mechanical durability of the finger of more than 15.000 closing cycles. Finally, we evaluate the accuracy of visual object segmentation during the different phases of the grasping process using five different objects. Hereby, an accuracy above 90 % can be achieved.

Jianfeng Gao, You Zhou, Tamim Asfour

Compliant robot behavior is crucial for the realization of contact-rich manipulation tasks. In such tasks, it is important to ensure a high stiffness and force tracking accuracy during normal task execution as well as rapid adaptation and complaint behavior to react to abnormal situations and changes. In this paper, we propose a novel approach for learning predictive models of force profiles required for contact-rich tasks. Such models allow detecting unexpected situations and facilitates better adaptive control. The approach combines an anomaly detection based on Bidirectional Gated Recurrent Units (Bi-GRU) and an adaptive force/impedance controller. We evaluated the approach in simulated and real world experiments on a humanoid robot.The results show that the approach allow simultaneous high tracking accuracy of desired motions and force profile as well as the adaptation to force perturbations due to physical human interaction.

Fabio Ferreira, Lin Shao, Tamim Asfour et al.

Endowing robots with human-like physical reasoning abilities remains challenging. We argue that existing methods often disregard spatio-temporal relations and by using Graph Neural Networks (GNNs) that incorporate a relational inductive bias, we can shift the learning process towards exploiting relations. In this work, we learn action-conditional forward dynamics models of a simulated manipulation task from visual observations involving cluttered and irregularly shaped objects. We investigate two GNN approaches and empirically assess their capability to generalize to scenarios with novel and an increasing number of objects. The first, Graph Networks (GN) based approach, considers explicitly defined edge attributes and not only does it consistently underperform an auto-encoder baseline that we modified to predict future states, our results indicate how different edge attributes can significantly influence the predictions. Consequently, we develop the Auto-Predictor that does not rely on explicitly defined edge attributes. It outperforms the baseline and the GN-based models. Overall, our results show the sensitivity of GNN-based approaches to the task representation, the efficacy of relational inductive biases and advocate choosing lightweight approaches that implicitly reason about relations over ones that leave these decisions to human designers.

Christian R. G. Dreher, Mirko Wächter, Tamim Asfour

Recognizing human actions is a vital task for a humanoid robot, especially in domains like programming by demonstration. Previous approaches on action recognition primarily focused on the overall prevalent action being executed, but we argue that bimanual human motion cannot always be described sufficiently with a single action label. We present a system for frame-wise action classification and segmentation in bimanual human demonstrations. The system extracts symbolic spatial object relations from raw RGB-D video data captured from the robot's point of view in order to build graph-based scene representations. To learn object-action relations, a graph network classifier is trained using these representations together with ground truth action labels to predict the action executed by each hand. We evaluated the proposed classifier on a new RGB-D video dataset showing daily action sequences focusing on bimanual manipulation actions. It consists of 6 subjects performing 9 tasks with 10 repetitions each, which leads to 540 video recordings with 2 hours and 18 minutes total playtime and per-hand ground truth action labels for each frame. We show that the classifier is able to reliably identify (action classification macro F1-score of 0.86) the true executed action of each hand within its top 3 predictions on a frame-by-frame basis without prior temporal action segmentation.

Jonas Rothfuss, Fabio Ferreira, Simon Boehm et al.

Modelling statistical relationships beyond the conditional mean is crucial in many settings. Conditional density estimation (CDE) aims to learn the full conditional probability density from data. Though highly expressive, neural network based CDE models can suffer from severe over-fitting when trained with the maximum likelihood objective. Due to the inherent structure of such models, classical regularization approaches in the parameter space are rendered ineffective. To address this issue, we develop a model-agnostic noise regularization method for CDE that adds random perturbations to the data during training. We demonstrate that the proposed approach corresponds to a smoothness regularization and prove its asymptotic consistency. In our experiments, noise regularization significantly and consistently outperforms other regularization methods across seven data sets and three CDE models. The effectiveness of noise regularization makes neural network based CDE the preferable method over previous non- and semi-parametric approaches, even when training data is scarce.

Jonas Rothfuss, Dennis Lee, Ignasi Clavera et al.

Credit assignment in Meta-reinforcement learning (Meta-RL) is still poorly understood. Existing methods either neglect credit assignment to pre-adaptation behavior or implement it naively. This leads to poor sample-efficiency during meta-training as well as ineffective task identification strategies. This paper provides a theoretical analysis of credit assignment in gradient-based Meta-RL. Building on the gained insights we develop a novel meta-learning algorithm that overcomes both the issue of poor credit assignment and previous difficulties in estimating meta-policy gradients. By controlling the statistical distance of both pre-adaptation and adapted policies during meta-policy search, the proposed algorithm endows efficient and stable meta-learning. Our approach leads to superior pre-adaptation policy behavior and consistently outperforms previous Meta-RL algorithms in sample-efficiency, wall-clock time, and asymptotic performance.

Christian R. G. Dreher, Nicklas Kulp, Christian Mandery et al.

There have been many proposals for algorithms segmenting human whole-body motion in the literature. However, the wide range of use cases, datasets, and quality measures that were used for the evaluation render the comparison of algorithms challenging. In this paper, we introduce a framework that puts motion segmentation algorithms on a unified testing ground and provides a possibility to allow comparing them. The testing ground features both a set of quality measures known from the literature and a novel approach tailored to the evaluation of motion segmentation algorithms, termed Integrated Kernel approach. Datasets of motion recordings, provided with a ground truth, are included as well. They are labelled in a new way, which hierarchically organises the ground truth, to cover different use cases that segmentation algorithms can possess. The framework and datasets are publicly available and are intended to represent a service for the community regarding the comparison and evaluation of existing and new motion segmentation algorithms.

Ignasi Clavera, Jonas Rothfuss, John Schulman et al.

Model-based reinforcement learning approaches carry the promise of being data efficient. However, due to challenges in learning dynamics models that sufficiently match the real-world dynamics, they struggle to achieve the same asymptotic performance as model-free methods. We propose Model-Based Meta-Policy-Optimization (MB-MPO), an approach that foregoes the strong reliance on accurate learned dynamics models. Using an ensemble of learned dynamic models, MB-MPO meta-learns a policy that can quickly adapt to any model in the ensemble with one policy gradient step. This steers the meta-policy towards internalizing consistent dynamics predictions among the ensemble while shifting the burden of behaving optimally w.r.t. the model discrepancies towards the adaptation step. Our experiments show that MB-MPO is more robust to model imperfections than previous model-based approaches. Finally, we demonstrate that our approach is able to match the asymptotic performance of model-free methods while requiring significantly less experience.

Fabio Ferreira, Jonas Rothfuss, Eren Erdal Aksoy et al.

We release two artificial datasets, Simulated Flying Shapes and Simulated Planar Manipulator that allow to test the learning ability of video processing systems. In particular, the dataset is meant as a tool which allows to easily assess the sanity of deep neural network models that aim to encode, reconstruct or predict video frame sequences. The datasets each consist of 90000 videos. The Simulated Flying Shapes dataset comprises scenes showing two objects of equal shape (rectangle, triangle and circle) and size in which one object approaches its counterpart. The Simulated Planar Manipulator shows a 3-DOF planar manipulator that executes a pick-and-place task in which it has to place a size-varying circle on a squared platform. Different from other widely used datasets such as moving MNIST [1], [2], the two presented datasets involve goal-oriented tasks (e.g. the manipulator grasping an object and placing it on a platform), rather than showing random movements. This makes our datasets more suitable for testing prediction capabilities and the learning of sophisticated motions by a machine learning model. This technical document aims at providing an introduction into the usage of both datasets.

Jonas Rothfuss, Fabio Ferreira, Eren Erdal Aksoy et al.

We present a novel deep neural network architecture for representing robot experiences in an episodic-like memory which facilitates encoding, recalling, and predicting action experiences. Our proposed unsupervised deep episodic memory model 1) encodes observed actions in a latent vector space and, based on this latent encoding, 2) infers most similar episodes previously experienced, 3) reconstructs original episodes, and 4) predicts future frames in an end-to-end fashion. Results show that conceptually similar actions are mapped into the same region of the latent vector space. Based on these results, we introduce an action matching and retrieval mechanism, benchmark its performance on two large-scale action datasets, 20BN-something-something and ActivityNet and evaluate its generalization capability in a real-world scenario on a humanoid robot.

Nikolaus Vahrenkamp, Eduard Koch, Mirko Waechter et al.

In this work, we present a grasp planner which integrates two sources of information to generate robust grasps for a robotic hand. First, the topological information of the object model is incorporated by building the mean curvature skeleton and segmenting the object accordingly in order to identify object regions which are suitable for applying a grasp. Second, the local surface structure is investigated to construct feasible and robust grasping poses by aligning the hand according to the local object shape. We show how this information can be used to derive different grasping strategies, which also allows to distinguish between precision and power grasps. We applied the approach to a wide variety of object models of the KIT and the YCB real-world object model databases and evaluated the approach with several robotic hands. The results show that the skeleton-based grasp planner is capable to autonomously generate high-quality grasps in an efficient manner. In addition, we evaluate how robust the planned grasps are against hand positioning errors as they occur in real-world applications due to perception and actuation inaccuracies. The evaluation shows that the majority of the generated grasps are of high quality since they can be successfully applied even when the hand is not exactly positioned.

Matthias Plappert, Rein Houthooft, Prafulla Dhariwal et al.

Deep reinforcement learning (RL) methods generally engage in exploratory behavior through noise injection in the action space. An alternative is to add noise directly to the agent's parameters, which can lead to more consistent exploration and a richer set of behaviors. Methods such as evolutionary strategies use parameter perturbations, but discard all temporal structure in the process and require significantly more samples. Combining parameter noise with traditional RL methods allows to combine the best of both worlds. We demonstrate that both off- and on-policy methods benefit from this approach through experimental comparison of DQN, DDPG, and TRPO on high-dimensional discrete action environments as well as continuous control tasks. Our results show that RL with parameter noise learns more efficiently than traditional RL with action space noise and evolutionary strategies individually.

Matthias Plappert, Christian Mandery, Tamim Asfour

Linking human whole-body motion and natural language is of great interest for the generation of semantic representations of observed human behaviors as well as for the generation of robot behaviors based on natural language input. While there has been a large body of research in this area, most approaches that exist today require a symbolic representation of motions (e.g. in the form of motion primitives), which have to be defined a-priori or require complex segmentation algorithms. In contrast, recent advances in the field of neural networks and especially deep learning have demonstrated that sub-symbolic representations that can be learned end-to-end usually outperform more traditional approaches, for applications such as machine translation. In this paper we propose a generative model that learns a bidirectional mapping between human whole-body motion and natural language using deep recurrent neural networks (RNNs) and sequence-to-sequence learning. Our approach does not require any segmentation or manual feature engineering and learns a distributed representation, which is shared for all motions and descriptions. We evaluate our approach on 2,846 human whole-body motions and 6,187 natural language descriptions thereof from the KIT Motion-Language Dataset. Our results clearly demonstrate the effectiveness of the proposed model: We show that our model generates a wide variety of realistic motions only from descriptions thereof in form of a single sentence. Conversely, our model is also capable of generating correct and detailed natural language descriptions from human motions.

Timothee Habra, Markus Grotz, David Sippel et al.

Gaze stabilization is fundamental for humanoid robots. By stabilizing vision, it enhances perception of the environment and keeps points of interest in the field of view. In this contribution, a multimodal gaze stabilization combining classic inverse kinematic control with vestibulo-ocular and optokinetic reflexes is introduced. Inspired by neuroscience, it implements a forward model that can modulate the reflexes based on the reafference principle. This principle filters self-generated movements out of the reflexive feedback loop. The versatility and effectiveness of this method are experimentally validated on the Armar-III humanoid robot. It is first demonstrated that each stabilization mechanism (inverse kinematics and reflexes) performs better than the others as a function of the type of perturbation to be stabilized. Furthermore, combining these three modalities by reafference provides a universal gaze stabilizer which can handle any kind of perturbation.

Matthias Plappert, Christian Mandery, Tamim Asfour

Linking human motion and natural language is of great interest for the generation of semantic representations of human activities as well as for the generation of robot activities based on natural language input. However, while there have been years of research in this area, no standardized and openly available dataset exists to support the development and evaluation of such systems. We therefore propose the KIT Motion-Language Dataset, which is large, open, and extensible. We aggregate data from multiple motion capture databases and include them in our dataset using a unified representation that is independent of the capture system or marker set, making it easy to work with the data regardless of its origin. To obtain motion annotations in natural language, we apply a crowd-sourcing approach and a web-based tool that was specifically build for this purpose, the Motion Annotation Tool. We thoroughly document the annotation process itself and discuss gamification methods that we used to keep annotators motivated. We further propose a novel method, perplexity-based selection, which systematically selects motions for further annotation that are either under-represented in our dataset or that have erroneous annotations. We show that our method mitigates the two aforementioned problems and ensures a systematic annotation process. We provide an in-depth analysis of the structure and contents of our resulting dataset, which, as of October 10, 2016, contains 3911 motions with a total duration of 11.23 hours and 6278 annotations in natural language that contain 52,903 words. We believe this makes our dataset an excellent choice that enables more transparent and comparable research in this important area.

Christian Mandery, Júlia Borràs, Mirjam Jöchner et al.

When executing whole-body motions, humans are able to use a large variety of support poses which not only utilize the feet, but also hands, knees and elbows to enhance stability. While there are many works analyzing the transitions involved in walking, very few works analyze human motion where more complex supports occur. In this work, we analyze complex support pose transitions in human motion involving locomotion and manipulation tasks (loco-manipulation). We have applied a method for the detection of human support contacts from motion capture data to a large-scale dataset of loco-manipulation motions involving multi-contact supports, providing a semantic representation of them. Our results provide a statistical analysis of the used support poses, their transitions and the time spent in each of them. In addition, our data partially validates our taxonomy of whole-body support poses presented in our previous work. We believe that this work extends our understanding of human motion for humanoids, with a long-term objective of developing methods for autonomous multi-contact motion planning.

Júlia Borràs, Tamim Asfour

Exploiting interaction with the environment is a promising and powerful way to enhance stability of humanoid robots and robustness while executing locomotion and manipulation tasks. Recently some works have started to show advances in this direction considering humanoid locomotion with multi-contacts, but to be able to fully develop such abilities in a more autonomous way, we need to first understand and classify the variety of possible poses a humanoid robot can achieve to balance. To this end, we propose the adaptation of a successful idea widely used in the field of robot grasping to the field of humanoid balance with multi-contacts: a whole-body pose taxonomy classifying the set of whole-body robot configurations that use the environment to enhance stability. We have revised criteria of classification used to develop grasping taxonomies, focusing on structuring and simplifying the large number of possible poses the human body can adopt. We propose a taxonomy with 46 poses, containing three main categories, considering number and type of supports as well as possible transitions between poses. The taxonomy induces a classification of motion primitives based on the pose used for support, and a set of rules to store and generate new motions. We present preliminary results that apply known segmentation techniques to motion data from the KIT whole-body motion database. Using motion capture data with multi-contacts, we can identify support poses providing a segmentation that can distinguish between locomotion and manipulation parts of an action.

Manfred Kröhnert, Nikolaus Vahrenkamp, Johny Paul et al.

Humanoid robots are designed to operate in human centered environments where they execute a multitude of challenging tasks, each differing in complexity, resource requirements, and execution time. In such highly dynamic surroundings it is desirable to anticipate upcoming situations in order to predict future resource requirements such as CPU or memory usage. Resource prediction information is essential for detecting upcoming resource bottlenecks or conflicts and can be used enhance resource negotiation processes or to perform speculative resource allocation. In this paper we present a prediction model based on Markov chains for predicting the behavior of the humanoid robot ARMAR-III in human robot interaction scenarios. Robot state information required by the prediction algorithm is gathered through self-monitoring and combined with environmental context information. Adding resource profiles allows generating probability distributions of possible future resource demands. Online learning of model parameters is made possible through disclosure mechanisms provided by the robot framework ArmarX.

Johny Paul, Walter Stechele, Manfred Kröhnert et al.

Humanoid robots are designed to operate in human centered environments. They face changing, dynamic environments in which they need to fulfill a multitude of challenging tasks. Such tasks differ in complexity, resource requirements, and execution time. Latest computer architectures of humanoid robots consist of several industrial PCs containing single- or dual-core processors. According to the SIA roadmap for semiconductors, many-core chips with hundreds to thousands of cores are expected to be available in the next decade. Utilizing the full power of a chip with huge amounts of resources requires new computing paradigms and methodologies. In this paper, we analyze a resource-aware computing methodology named Invasive Computing, to address these challenges. The benefits and limitations of the new programming model is analyzed using two widely used computer vision algorithms, the Harris Corner detector and SIFT (Scale Invariant Feature Transform) feature matching. The result indicate that the new programming model together with the extensions within the application layer, makes them highly adaptable; leading to better quality in the results obtained.

Jeannette Bohg, Antonio Morales, Tamim Asfour et al.

We review the work on data-driven grasp synthesis and the methodologies for sampling and ranking candidate grasps. We divide the approaches into three groups based on whether they synthesize grasps for known, familiar or unknown objects. This structure allows us to identify common object representations and perceptual processes that facilitate the employed data-driven grasp synthesis technique. In the case of known objects, we concentrate on the approaches that are based on object recognition and pose estimation. In the case of familiar objects, the techniques use some form of a similarity matching to a set of previously encountered objects. Finally for the approaches dealing with unknown objects, the core part is the extraction of specific features that are indicative of good grasps. Our survey provides an overview of the different methodologies and discusses open problems in the area of robot grasping. We also draw a parallel to the classical approaches that rely on analytic formulations.