Session 3A – Human–Robot Interaction 1
Date: Wednesday, 18th November 2026
Time: 16:00 – 17:30
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.
16:00 – Safe Speech-to-Action Human-Robot Interaction with Voice-Gated Access for ASTI-Controlled ProBOT-L Mobile Robots
Authors: Arman Ahmed Khan¹, Sadegh Refaeiabdolhosseinzadehneishabouri¹, Steffen Junginger² and Kerstin Thurow¹
Affiliations: ¹ Center for Life Science Automation (celisca), University of Rostock, Rostock, Germany; ² Institute of Automation, University of Rostock, Rostock, Germany.
Abstract
We present a safety-constrained speech-to-action architecture for ASTI-controlled ProBOT-L mobile robots in life science laboratories. The contribution is a modular interaction layer that connects natural-language requests to existing logistics services through explicit, inspectable execution steps. Cascaded automatic speech recognition, constrained large-language-model interpretation, and text-to-speech synthesis allow individual components to be replaced while preserving a common service interface. Typed function proposals, argument validation, voice-gated access, and configurable confirmation connect flexible language to bounded operations. A shared interaction controller coordinates dialogue, clarification, service-grounded feedback, and a complementary graphical interface, while scheduling, navigation, and physical safety remain within the existing robot infrastructure. An eight-participant feasibility pilot across four acoustic environments achieved 315 successful speech-command trials out of 318 (99.1%). Separate voice-identification trials yielded 73.1% correct enrolled-speaker identification and 26.7% unknown-speaker false acceptance, motivating stronger identity cues. GUI tasks and participant feedback complemented the technical evaluation. The results demonstrate practical hands-free access to laboratory robot services and support the architecture as a basis for further multimodal development.
16:15 – Constraint-Based Motion Planning for Safe Human-Robot Interaction
Authors: Pol Ramon-Canyameres¹,², Isiah Zaplana¹ and Jan Rosell¹
Affiliations: ¹ Institute of Industrial and Control Engineering, Universitat Politècnica de Catalunya (UPC), Barcelona, Spain; ² Eurecat, Technology Centre of Catalonia, Robotics & Automation Unit, Barcelona, Spain.
Abstract
Motion planning for robotic manipulators in human-shared environments requires strategies that explicitly account for human safety. Incorporating artificial constraints into the planner is an effective way to limit the workspace and the types of motions the robot is allowed to perform. In this context, this paper presents a constraint-based motion planning framework designed to support safety-oriented task execution in collaborative scenarios. Building on the constrained planning foundation of the Open Motion Planning Library (OMPL), the framework provides practical tools including XML-based specification, collaboration-aware geometric constraints, dynamic tolerance adaptation, and visualization capabilities. The framework is demonstrated through experiments with a dual-arm robot in a shared workspace with a human, showing the practical effectiveness of the proposed constraints as safety-oriented mechanisms for restricting robot motion during human-robot collaboration. Videos and code can be found here: https://robotics.ioc.upc.edu/constraint motion planning/
16:30 – Human-Robot Pose Alignment Based on Orientation Estimation and F-Formation
Authors: Qiaoyue Yang, Jesus Enrique Aleman Gallegos and Sven Wachsmuth
Affiliation: Faculty of Technology, Bielefeld University, Bielefeld, Germany.
Abstract
Socially aware robots deployed in human-populated environments are expected to behave in a human-like manner and comply with human social norms. In social navigation, a robot plans its motion while accounting for the presence of people. However, the goal of its motion is yet underexplored. Socially compliant spatial coordination between robots and humans is required for smooth human-robot interaction (HRI) and for mitigating misunderstandings. Therefore, we propose a human–robot pose alignment framework that enables a robot to autonomously determine a navigation goal pose aligned with its interaction partner. The framework consists of (i) a perception module that estimates the human pose from RGB-D sensor input and (ii) a decision module that generates a human-aware, collision-free target pose for the robot. It integrates empirical data with customizable F-formation-based rules to balance flexibility and transparency in decision-making. We explored both human standing and sitting scenarios including human body occlusions and environmental obstacles in a user study. The results reveal that the F-formation is also applicable between robots and people, but its orientations require fine-tuning to establish “eye contact” and improve the legibility of the robot’s interaction intent. Our code is publicly available at GitHub https://github.com/Q-Y-Yang/Human-Robot-Alignment.git .
16:45 – Go, Wait, or Detour? Human-Aligned Aisle Decisions for Retail Robots from RGB-D
Authors: Sarin Kittisares¹, Shusuke Matsuda², Manabu Nakanoya¹, Natsuhiko Sato¹, Tatsuya Yoshimoto¹, Hiroshi Yoshida¹ and Kousuke Nogami¹
Affiliations: ¹ Cyber-Physical Intelligence Research Laboratories, NEC Corporation, Kawasaki, Japan; ² Business Innovation Department, NEC Corporation, Kawasaki, Japan.
Abstract
A shelf-scanning robot sharing narrow retail aisles with customers, who often stand still to examine goods, must decide whether to proceed, wait, or take a detour. This decision depends not only on geometric clearance but on the disturbance the robot would cause, which is a subjective human judgment. We construct a controlled RGB-D dataset in a test store at three aisle widths, in which multiple annotators label each scene with a go/wait/detour action and a graded disturbance score. Human agreement on this task is only moderate, confirming that it is inherently subjective and that human-level agreement is the appropriate target. To test whether the dataset supports learning human-aligned decisions, we train a lightweight model on a frozen pretrained RGB-D backbone. It reaches an action accuracy of 0.785, comparable to the human agreement level of 0.736, whereas three open-weight vision–language models (0.46–0.51) and a width-agnostic rule-based baseline (0.558) perform substantially worse. The graded score also proves useful beyond describing disturbance: as an auxiliary training target, it tends to improve the action decision. Finally, we discuss the implications, including the unverified relationship between human disturbance judgments and actual customer experience.
17:00 – ADAPT: A Safe Dual-Arm Teleoperation Framework for Data Collection on robot ADAM
Authors: Laura Pons, Adrian Prados, Alberto Mendez, Gonzalo Espinoza and Ramon Barber
Affiliation: RoboticsLab, System and Automation, Universidad Carlos III, Madrid, Spain.
Abstract
To successfully train robotic systems through Learning from Demonstrations, large and high-quality datasets of human demonstrations are normally required. However, dual-arm data collection is severely bottlenecked by prohibitive hardware costs, complex dual-arm coordination, and a persistent risk of self-collisions, which restricts the deployment of these platforms. In this work, we present ADAPT (ADAM Dual-Arm Platform for Teleoperation), a comprehensive hardware and software framework based on GELLO that enables safe, intuitive, and low-cost data collection for the ADAM mobile manipulator. The proposed system incorporates an active weight compensation mechanism to bridge the physical discrepancies between the leader device and the robot, alongside an active safety system based on Signed Distance Fields that prevents self-collisions without restricting the operator’s workspace. The framework has been evaluated in both simulated environments and on the real ADAM platform, demonstrating that inexperienced users can safely collect complex dual-arm demonstrations. Website project: https://mobile-robots-group-uc3m.github.io/adapt/
17:15 – Intent-Aware Mixed Reality Shared Autonomy for Safe Robotic EV Battery Dismantling
Authors: Abdussalam Alajami, Fergus Proctor and Manuel Vinagre
Affiliation: LEITAT Technological Center, Robotics and Automation, Barcelona, Spain.
Abstract
Electric vehicle (EV) battery dismantling is a safety-critical de-manufacturing process that requires precise manipulation, continuous situational awareness, and human supervision due to electrical, thermal, chemical, and mechanical hazards. This paper presents an intent-aware mixed reality shared-autonomy framework for industrial robot-assisted EV battery dismantling. The proposed system integrates an immersive digital twin, RGB-D-based workspace representation, operator motion cues, field-of-view-gated feature selection, rule-based geometric intent inference, and safety-bounded semi-autonomous task primitives. Instead of relying on direct teleoperation alone, the framework interprets the operator’s spatial focus, tool pose, approach direction, and task context to identify intended dismantling actions such as cutting, unscrewing, connector removal, and component extraction.
A confirm-then-execute strategy is introduced to preserve human authority: candidate task features are scored using interpretable geometric rules, stabilised through dwell time and hysteresis, presented to the operator for confirmation, and executed only after explicit approval. The framework was evaluated in a Gazebo Fortress simulation environment integrated with ROS 2 Humble and MoveIt 2, using a UR10 manipulator and a representative EV battery workcell. In a precision cut-edge task to seperate battery cell rows, intent-triggered autonomous handover achieved a mean deviation of 7.6 mm and RMS deviation of 11.2 mm with full edge coverage, compared with 28.8 mm and 29.7 mm for manual VR teleoperation. These results indicate that intent-aware mixed reality assistance can improve accuracy, repeatability, and task completion while keeping the operator in control, supporting safer and more adaptive robotic EV battery de-manufacturing.