Session 1A – Assistive, Rehabilitation and Social Robotics
Date: Wednesday, 18th November 2026
Time: 11:00 – 12:15
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.
11:00 - Active Sensor Health Monitoring and Lifelong Spatial Mapping for Resilient Robot Localization
Authors: Francisco Rodrigues¹, John Lewis Devassy¹, Alberto Vale², Meysam Basiri¹, João Santos³ and Pedro U. Lima¹
Affiliations: ¹ Institute for Systems and Robotics, Instituto Superior Técnico, Lisbon, Portugal; ² Instituto de Plasmas e Fusão Nuclear, Instituto Superior Técnico, Lisbon, Portugal; ³ The Toro Company, Lisbon, Portugal.
Abstract
Autonomous robots operating in complex structured environments frequently experience transient sensor degradation, leading to localization divergence. While multi-sensor fusion mitigates some errors, traditional filters often fail to isolate and reject severely corrupted data before it compromises the state estimate. This paper presents a real-time Sensor Health Monitor that systematically evaluates heterogeneous sensory data streams and assigns dynamic quality scores based on metrics such as registration fitness, inlier RMSE, and noise energy. Operating prior to the filter update step, the monitor actively suppresses degraded measurements. Furthermore, we extend this framework for lifelong localization by introducing a spatial Health Map, which aggregates historical sensor performance to anticipate systematic failures in previously traversed regions. Evaluated on the public Nebula dataset (subterranean mines) and a novel golf course dataset, the proposed framework reduces the maximum Absolute Pose Error (APE) by over 97% during severe synthetic failure events and by 71% in challenging nominal conditions. Additionally, leveraging the lifelong spatial memory dropped peak errors by a further 76% during repeated traversals. The system imposes minimal computational overhead, offering a highly resilient enhancement for autonomous navigation stacks.
11:15 - Intelligent Robotic Massage with Anatomical Adaptation via 3D Point Cloud Processing
Authors: Nancy Barbosa¹,², Lucía Sánchez-Urán¹, Ignacio Secades², Francisco Javier Badesa¹ and Manuel Ferre¹
Affiliations: ¹ Centre for Automation and Robotics (UPM-CSIC), Universidad Politécnica de Madrid, Spain; ² Canonical Robots, Madrid, Spain.
Abstract
This paper addresses the problem of anatomical personalization for robotic massage through 3D point cloud processing. The proposed framework integrates an RGB-D camera into a robotic massage setup to reconstruct the surface of the subject’s back. From the processed point cloud, the system automatically extracts geometric reference landmarks and uses them to register a generic massage template to the subject-specific geometry through a triangulated representation and affine-invariant barycentric mapping. This pipeline enables predefined massage trajectories to adapt to different body shapes, sizes, and postural configurations while preserving their relative spatial correspondence. The framework was experimentally evaluated using 28 3D scans obtained from 20 healthy subjects with different body morphologies and postural configurations. The results show that the proposed geometric landmark extraction and trajectory-transfer pipeline can generate subject-specific massage trajectories, with a mean processing time of 4.6 s. By combining 3D surface reconstruction, geometric landmark extraction, and structured trajectory transfer, the proposed methodology provides a computationally efficient link between offline surface perception and subsequent robotic execution.
11:30 - From Intervention Tool to Supervised Support Resource: A Social Robot System Architecture for Co-constructed Play with Neurodivergent Children
Authors: Huiguo Ma¹, Raquel Cedazo¹, Yuqi Bao² and Jiaxing Han¹
Affiliations: ¹ Centro de Automatica y Robotica (CSIC-UPM), Escuela Técnica Superior de Ingeniería y Diseño Industrial (ETSIDI), Universidad Politécnica de Madrid, Spain; ² Jiangxi University of Software Professional Technology, Nanchang, China.
Abstract
Play is a fundamental right through which neurodivergent children develop abilities such as inhibitory control and joint attention, yet the competitive rules and sensory stimulation of traditional group games often restrict their participation. Existing socially assistive robots are mostly used for one-to-one training and are often assigned the role of instructor or referee, which can suppress children’s agency and cause frustration. This paper presents JARI 2.0, which repositions the robot as a teacher-supervised support resource for co-constructed inclusive play. The core of the system is a normalized, scale-invariant visual tolerance model that preserves a pixel-level tolerance band for small involuntary movements in the Musical Statues Game, thereby reducing false violations and punitive feedback. The system also fuses facial, vocal, and motion cues to estimate coarse-grained affective states and uses these estimates to lower interaction arousal. Implemented on Raspberry Pi 5 and ROS2, the system integrates a soft-touch safe body, an adaptive state machine, structured logging, and a highest-priority teacher override interface based on rosbridge. Preliminary classroom validation across 12 sessions with seven children showed 11.7 ± 1.3 fps visual processing, 91.2 ± 4.0% player-ID retention, and a 95.8% complete-game operation rate, while supporting concurrent play by one to four children under teacher supervision.
11:45 - Emotion-Aware Ethical Supervision for Teleconference-Mediated Interaction in Social Service Robots
Authors: António Alves¹, Fadi Aldeeb² and Paulo J.S. Gonçalves³
Affiliations: ¹ Instituto Politécnico de Castelo Branco, Portugal; ² IDMEC, Instituto Superior Técnico, Universidade de Lisboa, Portugal; ³ IDMEC, Instituto Politécnico de Castelo Branco, Portugal.
Abstract
Social Service Robots (SSRs) for older-adult care must operate in sensitive interaction contexts where apparently simple actions, such as starting or maintaining a conversation, may have ethical implications. Previous work addressed the ethical assessment of action initiation using ontology-based reasoning and SWRL rules. This paper extends that approach towards continuous ethical supervision during human–robot interaction, with particular focus on teleconference-mediated conversations. The proposed architecture integrates a ROS-based robot software framework, an onboard hardware platform, a cloud-based Ethical Engine, a MongoDB decision log, and an emotion-recognition module. The robot evaluates whether an interaction may start, monitors the older adult’s emotional state during the call, and determines whether the teleconference may continue or requires intervention. An ensemble of convolutional neural networks was implemented for facial emotion recognition and deployed as a ROS 2 node. Experimental results with 80 participants show that the module correctly identified most tested emotional expressions, supporting its integration into the ethical supervision loop. The proposed system demonstrates how ontological reasoning and emotion-aware monitoring can be combined to support transparent, auditable, and ethically controlled SSR interactions with older adults.
12:00 - Low-Burden Pain Recognition for Robot-Assisted Rehabilitation Using Facial Action Units
Authors: Ane San Martin Igarza¹, Gabriele Kubiliute¹,² and Johan Kildal¹
Affiliations: ¹ Autonomous and Intelligent Systems, Tekniker, Eibar, Spain; ² Department of Computer Science and AI, EHU, Spain.
Abstract
Robot-assisted rehabilitation systems commonly adapt therapy based on biomechanical variables, but they rarely consider patients’ internal states such as pain or discomfort. This work investigates low-burden pain recognition from facial behavior as a potential feedback signal for future adaptive rehabilitation robots. Using the Pain Expression Multimodal Faces dataset, facial Action Units were automatically extracted from RGB videos with OpenFace 2.0 and used to distinguish neutral and spontaneous pain expressions. Conventional machine learning models and recurrent deep learning architectures were evaluated under different sliding-window lengths to analyze the effect of temporal context. Results show that temporal information improves pain recognition, with recurrent models outperforming conventional approaches. Random Forest achieved the best conventional performance, reaching an F1-score close to 0.89, while the BiLSTM with attention obtained the best overall result, reaching 0.976 with a 15-frame window. SHAP analysis identified eye-, brow-, cheek-, and mouth-related Action Units as the most relevant cues. These findings suggest that AU-based vision sensing can provide an interpretable and non-contact approach for pain-aware rehabilitation systems.