Session 2A – Deformable Object Manipulation and Tactile Robotics
Date: Wednesday, 18th November 2026
Time: 14:00 – 15:30
Room: Sala d’Actes
Session Chair: TBA
Presentation format: Each paper is allocated 15 minutes: approximately 10–12 minutes for the presentation and 3–5 minutes for questions.
14:00 - Antenna Tracking Using a Curve-Fitting Approach
Authors: Maria Giovanna Di Gaetano¹, Sedat Dogru² and Lino Marques²
Affiliations: ¹ Department of Control and Computer Engineering (DAUIN), Polytechnic University of Turin, Turin, Italy; ² Institute of Systems and Robotics, Department of Electrical and Computer Engineering, University of Coimbra, Coimbra, Portugal.
Abstract
Directional antennas can improve link quality in robotic communication systems, but their benefits depend on precise alignment with the signal source. This paper uses a pan-tilt unit (PTU) to estimate source direction from the full received signal strength indicator (RSSI) profile collected during controlled sweeps. The measured sweep is fitted to an empirical beam shape through least-squares (LSQ) minimization, providing a direction estimate whose reliability is assessed before updating the pointing direction. The proposed method is evaluated both in simulation, using moving target scenarios, and in the field using a directional antenna system. The localization stage shows precise and consistent convergence, while predictive sweep repositioning supports tracking under sensing and mechanical imperfections. The results obtained by this method support the use of full-profile RSSI fitting for directional antenna tracking in realistic robotic scenarios.
14:15 - Experimental Benchmark of Deformation Jacobian Estimation Methods for Shape Servoing of Deformable Linear Objects
Authors: Mathijs Severac¹, Alessio Caporali² and Gianluca Palli²
Affiliations: ¹ Ecole Centrale de Nantes, Nantes, France; ² Department of Electrical, Electronic and Information Engineering (DEI), University of Bologna, Bologna, Italy.
Abstract
Robotic manipulation of Deformable Linear Objects (DLOs) (e.g. cables, ropes, and wires) remains challenging due to their high-dimensional state space and nonlinear dynamics. A key enabler for closed-loop control is the deformation Jacobian, which maps robot control inputs to local changes in object shape. This paper presents an experimental benchmark of several model-based and model-free deformation Jacobian estimation methods for DLO shape servoing. The model-based set includes physics-based simulation, neural-network surrogate models, and the Cosserat rod formulation, while the model-free set includes Broyden’s update rule and polynomial approximation. We experimentally evaluate these methods on a common set of real-world DLO shape servoing tasks, comparing manipulation accuracy, modeling and data requirements, computational efficiency, implementation complexity, and suitability for online control. The results reveal distinct trade-offs between the approaches and provide practical guidelines for selecting estimation strategies tailored to task constraints.
14:30 - Cloth State Manipulation with Gradient Descent in Topological Coordinates
Authors: Franco Coltraro, Adrià Colomé, Júlia Borràs and Carme Torras
Affiliation: Institut de Robòtica i Informàtica Industrial (IRI), CSIC-UPC, Spain.
Abstract
Robotic cloth manipulation is a challenging task because cloth is highly deformable, difficult to model accurately, and only partially observable in real time. We propose a gradient-based method for controlling cloth from one qualitative state to another using topological dGLI (derivative of the Gauss Linking Integral) coordinates, a low-dimensional representation computed from eight boundary segments near the cloth corners. Given a target dGLI state, the method uses a cloth simulator to estimate how small motions of the grasped point affect the Spearman distance between the current and target dGLI coordinates. The resulting gradient defines the control direction, with an additional trajectory-shaping term to reduce self-friction and slippage. We compare this approach with gradient descent on the full-mesh RMSE and show that dGLI-based control achieves comparable final states while requiring substantially less feedback information. These results suggest that topological coordinates can provide a compact and effective feedback signal for cloth state manipulation.
14:45 - Dual-arm manipulation of a hanging T-shirt for robotic hem sewing through vision-based grasping
Authors: Filipe Almeida¹, Carlos M. Costa¹,², Cláudia D. Rocha¹, Gonçalo Leão¹,², Armando Sousa¹,², Lara Silva³, Luís F. Rocha¹ and Germano Veiga¹
Affiliations: ¹ Centre for Robotics in Industry and Intelligent Systems, INESC TEC, Porto, Portugal; ² Faculty of Engineering, University of Porto, Porto, Portugal; ³ ESI Robotics, Esmeriz, Portugal.
Abstract
The automation of garment feeding for sewing operations remains a challenging problem due to the highly deformable nature of textile materials and the need for accurate garment positioning before sewing. This paper presents an improved robotic system for automatically feeding T-shirts into an industrial bottom hem sewing machine. Building upon previous work based on predefined grasping positions, the proposed approach combines dual-arm robotic manipulation with vision-based grasp point detection to increase robustness and adaptability to variations in garment configuration. The system also incorporates improvements to the robotic workstation and the interaction with the sewing machine to enhance the reliability of the feeding process. Experimental validation was conducted through 40 trials, including both insertion-only and complete feeding-and-sewing experiments. The proposed system achieved a 95% success rate in the complete automated process while maintaining an average cycle time of approximately 31 s, representing a substantial improvement over the previous approach without compromising productivity. The obtained results demonstrate that the proposed solution provides reliable and repeatable robotic garment feeding, representing a step towards the automation of industrial textile manufacturing.
15:00 - Tactile Information in in-hand manipulation with Model-Based Lookahead and Model Predictive Actor Critic Reinforcement Learning
Authors: André Rafael Maissa Antunes¹, Catarina Barata² and Plinio Moreno²
Affiliations: ¹ Instituto Superior Técnico, Lisbon, Portugal; ² Instituto de Sistemas e Robótica, Instituto Superior Técnico, Lisbon, Portugal.
Abstract
Reinforcement learning has been one of the main approaches to effectively solve complex tasks such as in-hand manipulations and other dexterous tasks. These tasks have complex dynamic systems and the ability to control various objects using the actuators of the hand, increasing the difficulty of the task. This work presents the application of a previously developed hybrid reinforcement learning framework to an in-hand manipulation task using tactile information, evaluating its ability to improve task performance. The framework combines ideas from both model-based and model-free reinforcement learning by guiding the policy through a forward dynamics model using Model Predictive Control (MPC). Two versions of the framework are evaluated using different policy learning algorithms on multiple objects manipulated by an underactuated robotic hand. The first version employs Proximal Policy Optimization (PPO) and is compared against a baseline policy without the dynamics model. Additional experiments evaluate the effect of removing tactile information and perform generalization tests. The second version uses the Soft Actor-Critic (SAC) algorithm. The results show that, when combined with a sufficiently accurate dynamics model, PPO can improve manipulation performance and maintain robustness under small variations in object properties. However, the effectiveness of MPC strongly depends on the availability of tactile information and on the characteristics of specific objects. Additionally, the use of MPC introduces computational limitations that may affect real-time applicability. Nonetheless, these results represent a first step toward combining MPC-based methods with reinforcement learning for complex in-hand manipulation tasks involving underactuated hands and tactile sensing.
15:15 - Handling folded clothes: deformation estimation before, during and after grasping
Authors: Irene Garcia-Camacho, Júlia Borràs and Guillem Alenyà
Affiliation: Institut de Robòtica i Informàtica Industrial (IRI), CSIC-UPC, Spain.
Abstract
Manipulating textiles is challenging due to the difficulty of understaing their deformation state. State estimation is an open problem that includes different challenges, such as the need to find simplified representations and deal with self-occlusions. We claim that the requirements for state estimation are very different depending on three main categories of deformation: crumpled, flat and folded, and when approached separately, self-occlusions can become less relevant and simplified representations are easier. In this paper, we propose a method for handling folded clothes by estimating the deformation state both during grasping and after placement with a unified simple representation. In addition, we demonstrate that having a comprehensive characterization of cloth and explicitly integrating cloth properties, we are able to predict cloth deformation induced by a grasp action before its execution, enabling grasp planning decisions.