Session 4A - Human–Robot Interaction 2
Date: Thursday, 19th November 2026
Time: 09:45 – 10:45
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.
09:45 – Agentic Social Robotic System for Work Environments
Authors: David Morilla-Cabello, Maria Peribañez, Fernando Peña, Ruben Martinez-Cantin, Ana C. Murillo and Eduardo Montijano
Affiliations: Instituto de Investigación en Ingeniería de Aragón (I3A), Universidad de Zaragoza, Spain.
Abstract
Large language models (LLMs) have made natural-language interaction with social robots increasingly capable, but the memory that personalizes this interaction is typically built for a single user, as a personal assistant, rather than for the collective knowledge of a shared workplace. Moreover, such agents are predominantly virtual, without the physical presence that embodiment affords. We present an embodied lab companion that maintains an explicit memory of the people, interactions, and shared knowledge of a research group. Its memory is stored as human-readable records that an LLM-based agent can read and modify through format-preserving tools, which assist the agent by reducing the required generation length, avoiding hallucinations from corrupting the memory, and allowing the records to be programmatically managed and human-audited. The proposed system integrates spoken interaction, local LLM inference with tool use, and autonomous navigation running on a mobile robot platform. We validate it through real interactions with lab members, evaluating its agentic tool use and the faithfulness with which it builds and recalls group memory.
10:00 – A Webots Automated Warehouse Logistics Scenario for Robotics Education
Authors: João Braun¹, Arezki A. Chellal¹, Paulo Costa² and José Lima¹
Affiliations: ¹ CeDRI, SusTEC, Bragança Polytechnic University, Bragança, Portugal; ² INESC TEC, Faculty of Engineering of University of Porto, Porto, Portugal.
Abstract
This paper presents an open-source Webots scenario adapting the RobotAtFactory 4.0 warehouse logistics workflow for robotics education. The environment includes four movable parts, two processing machines, incoming and outgoing warehouses, and a differential-drive robot. A logistics supervisor manages part assignments, machine processing, and scoring independently of the student controller through a text-based protocol. The C++ scaffold provides named robot-centre poses, waypoint and arc route planning, feedback-controlled motion primitives, and non-blocking manipulation actions. Compact C and Python examples access the same protocol without the C++ navigation wrapper. The reference C++ policy implements the required processing routes for one part, leaving multi-part coordination and scheduling as extensions. Fifteen simulation trials covering the three initial part types achieved delivery with the expected scores and no supervisor-rejected releases. These results provide a nominal technical baseline. Classroom learning outcomes remain to be evaluated.
10:15 – Emotion-Aware Navigation for Social Robots: A Fuzzy Logic Approach Integrating Facial Expression Recognition
Authors: Carlotta Pifferi¹, Jiaxi Huo², Zhenhao Zhang³, Daniel Galan³ and Paloma de la Puente³
Affiliations: ¹ Electronics and Telecommunications, Politecnico di Torino, Torino, Italy; ² Mechanical and Aerospace Engineering, Politecnico di Torino, Torino, Italy; ³ Centre for Automation and Robotics (CAR), Universidad Politécnica de Madrid, Madrid, Spain.
Abstract
This paper presents an approach that integrates real-time facial expression recognition in a fuzzy-logic-based emotional model to dynamically modulate navigation behavior. The system enables the mobile robot to adapt its movement parameters based on detected user emotions, for more natural and empathetic human-robot interactions. Two internal emotional dimensions, mood and alertness, are processed through Mamdani fuzzy controllers to influence navigation speed and distance to obstacles tuning. Two distinct robot personalities, shy and intense, are implemented through differentiated fuzzy rule configurations. Experimental validation with 28 participants using standardized Human-Robot Interaction (HRI) questionnaires (Godspeed Animacy and Likeability scales, Escalas Visuales Analógicas de Estado de Ánimo (EVEA)) shows significant improvements in perceived animacy and likeability when the emotional system is active. Statistical analysis using mixed Analysis of Variance (ANOVA) suggests that emotional modulation improves perceived robot naturalness and social competence in University environments. On a public dataset, the recognition module reaches a macro-average F1-score of 0.47, with large differences across expression classes, and we discuss how recognition errors propagate through the fuzzy controller. Since the neutral condition always came first and the sessions did not record kinematic data or controller states, the questionnaire results describe perceived rather than measured robot behavior.
10:30 – Characterizing Human Hand Movement: Towards a Minimal Representation for Robotic Hand Systems
Authors: Anna Cristóbal Mirats¹, Marina Pujol Closa² and Enric Xavier Martín Rull³
Affiliations: ¹ Technical University of Catalonia (UPC), Barcelona, Spain; ² Institute of Industrial and Control Engineering (IOC), Technical University of Catalonia (UPC), Barcelona, Spain; ³ B2slab, Institut de Recerca i Innovació en Salut (IRIS), Technical University of Catalonia (UPC), Barcelona, Spain.
Abstract
This paper presents a data-driven framework for mapping human hand motion to the joint space of an anthropomorphic robotic hand while investigating whether the required control information can be represented compactly. Human motion was recorded with a 22-channel CyberGlove, and the 16 joints of the Allegro Hand were used as robotic reference outputs. The dataset comprised 6 static hand configurations and 10 dynamic transitions selected to activate different combinations of finger joints and capture coordinated movement patterns beyond object-specific grasps. Following user-specific calibration, the glove signals were normalized to a common 0–100 range, visually validated, and paired with the corresponding robotic postures and trajectories. Separate feedforward neural networks were trained for static and dynamic mapping. On this normalized 0–100 scale, the direct static model achieved a mean absolute error of 0.210 and a root mean square error of 0.247, whereas the dynamic model obtained 0.679 and 1.042, respectively. Bottleneck variants were then evaluated using 6 latent variables for static postures and 8 for dynamic trajectories. Despite the dimensionality reduction, the increase in prediction error remained limited, indicating that the main information required for the evaluated human-to-robot mapping was preserved. These results support the potential of compact latent representations for simplifying robotic hand control within the considered dataset and provide a foundation for future validation across users, movements, and grasping tasks.