The rapid adoption of autonomous multicopters in low-altitude airspace has necessitated new noise mitigation strategies. As mechanical noise reduction reaches its physical limits, trajectory optimization via intelligent agents offers a complementary solution. However, training such agents, specifically via Reinforcement Learning, is hindered by the lack of simulation environments that couple rigid-body dynamics with acoustic emission models. In this paper, we present an online acoustic simulation framework for multicopters. We introduce a parametric acoustic model that operates entirely in the frequency domain, bypassing the computational overhead of Fast Fourier Transforms (FFT). It accounts for propeller speed, flight state, and environmental factors, integrated within the Robot Operating System 2 (ROS 2) with an interface to the PX4-Autopilot Software-in-the-Loop (SITL). We validate the model by comparing simulated observer signals with real-world acoustic flight data. The results show low errors in simulated overflights for A-weighted spectra ranging between 2.5% and 9%. This validation demonstrates the viability of our modeling approach as a tool for training acoustic-aware autonomous multicopter agents.
Recent discussions on the Ethical, Legal, and Social Issues (ELSI) of social robots highlight the importance of understanding the ethical nature of human-robot relationships beyond issues of safety and regulation. Drawing on the philosophy of technology-from Ivan Illich's concept of convivial tools to contemporary relational approaches-this paper proposes a conceptual framework for human-robot coexistence based on two complementary paradigms: convivial robotics and silicopathy. Convivial robotics, inspired by Illich's notion of conviviality and exemplified by the “weak robot” approach, emphasizes robots that enhance human capabilities through imperfect yet mutually supportive interaction. Silicopathy, by contrast, refers to the emergence of artificial empathy grounded in representations of pain and affective states, through which ethical behavior is generated from internally constructed value dynamics, enabling morally meaningful relationships between humans and robots. Together these paradigms form a dual ethical framework that integrates human capability enhancement with empathic moral relations. The framework is further discussed from a cross-cultural perspective, contrasting Western autonomy-centered ethics with Eastern relational philosophies emphasizing harmony and compassion, thereby offering a conceptual basis for future human-robot coexistence, including its governance across diverse cultural and societal contexts.
Creating social robots is a multidisciplinary challenge that requires understanding human perception, behavior interpretation, and social–ethical expectations in human–robot interaction. Animal-like robots are a key category in this domain, yet quantitative evidence on user acceptance and perceived safety, especially for quadruped robots, remains limited. We therefore conducted a multilingual survey on the perception and acceptance of quadruped robots with $\mathbf{N} \boldsymbol{=} \mathbf{4 5 4}$ participants. Importantly, participants formed their judgments from indirect exposure to standardized visual stimuli (images and short videos) rather than from direct physical interaction. The results suggest that reported comfort under indirect exposure is associated with robot appearance and the coherence of design cues across animal-like and machine-like constructions, with mismatched zoomorphic cues tending to be viewed less favorably. Concerns about job loss and accidents are negatively associated with perceived safety, while concerns about unfair decisions and privacy show no significant influence. Prior experience with robotics is associated with higher acceptance and reduced anxiety. Appearance shows up to support initial sense-making, but did not correspond to uniformly high reported trust; responsibility attribution in the event of malfunction also remained contested, with responses clustering into technology-centered, human-centered, and shared-responsibility positions. Overall, the findings provide exploratory quantitative evidence on mediated first impressions of quadruped robots and motivate follow-up studies using direct human–robot interaction to test, validate, and extend these observations.
Service robots must reliably execute natural-language instructions in dynamic, partially observable environments, yet existing language-driven planners lack verifiability, adaptability, and computational efficiency. To address these challenges, we propose HSP-Plan, a hierarchical and verifiable planning framework that combines high-level LLM-based task decomposition with symbolic planning (via PDDL), efficient sub-goal execution using a smaller language model, and a scene graph-based memory capturing spatial and temporal relations for adaptive replanning. In ALFWorld, HSP-Plan achieves a task success rate of 0.78, outperforming AutoGen (0.77) and significantly surpassing ReAct (0.55). It further improves sub-goal completion to 0.88 and reduces high-cost LLM calls to approximately 6 per episode, compared to about 18 for ChatGPT-only execution. These results demonstrate that integrating hierarchical reasoning, symbolic verification, and structured memory enables more reliable and efficient language-guided robot planning in complex environments.
This paper presents UNCOM, a novel hybrid framework for interpreting natural human commands in tabletop scenarios. The system integrates multiple sources of information - speech, gestures, and scene context - to extract structured, actionable instructions for robots. Addressing the need for general-purpose human-robot interaction in domestic environments, UNCOM is designed for zero-shot operation, without reliance on predefined object models or training data specific to a given task. Using foundational and task-specific deep learning models, it allows out-of-the-box speech recognition, natural language understanding, gesture detection, and object segmentation. The modular architecture enhances transparency and explainability by explicitly parsing commands into object-action-target representations, enabling integration with symbolic robotic frameworks. We demonstrate the system in a TIAGo++ robot and provide an evaluation on a real-world data set of human-robot interaction scenarios; achieving an 82.39% success rate over our benchmark data set, highlighting the robustness of the system to diversity, noise, and communication ambiguity. The data set, evaluation scenarios, and the code are publicly available to support future research.
Intelligent wheelchairs for passenger transport are gaining prominence in the healthcare and transportation sectors due to demographic shifts. While social navigation among pedestrians is a well-studied problem in mobile robotics, existing approaches largely ignore the need to balance the priorities of passengers and nearby pedestrians, often disadvantaging the passenger. In this paper, we propose a social navigation layer based on the Social Force Model (SFM) to enable mutual navigation between pedestrians and passengers, thereby improving passenger comfort. We first show that current proxemics-based social navigation layers disadvantage passengers by prioritizing pedestrian comfort. To address this problem, our approach predicts pedestrian motion using the SFM while incorporating the passenger's social forces and adapts proxemics along the predicted trajectories. We implement the proposed method as a costmap layer, enabling deployment across a wide range of navigation setups. Simulation and real-world experiments demonstrate that our approach enables mutual navigation between pedestrians and passengers, while virtual reality user studies show a significant improvement in passenger comfort.
For older adults, sustained loneliness is a major risk factor for mental and physical health, motivating the use of socially assistive robots as long-term companions rather than short-term interfaces. Current research in the field of social robotics is attempting to attain a health-supporting role, bolstered by Large Language Models (LLMs). However, general dialogs that are not tailored to the specific interaction partner quickly undermine trust, perceived social presence, and long-term acceptance in Human–Robot Interaction (HRI). The required personalization in HRI for older adults therefore demands long-term episodic memory. However, relying on cloud-based LLMs introduces personalization issues and latency bottlenecks. Existing Retrieval-Augmented Generation (RAG) approaches often fail to balance data usage with the need for coherent, longitudinal dialog history. To address this, we present Smalltalk-KG, a local architecture that grounds conversational AI in an on-device knowledge graph powered by episodic memory. Smalltalk-KG utilizes a two-layer memory model to enforce personalization and ensure verifiable provenance. We evaluate the system through an 8-week offline simulation (N= 2,880 turns). Results demonstrate that Smalltalk-KG reduces retrieval latency to 43 ms relative to history length while maintaining high factual consistency and strictly adhering to local-only data processing constraints.
When industrial collaborative robotic assistants make mistakes, trust between the human and robot is likely to be enhanced if the robot issues an apology when it makes a mistake in its assigned task. To evaluate user perceptions toward an industrial robot, we design a virtual reality (VR) digital twin of a warehouse where a worker and the twin of a Universal Robots UR10 manipulator arm collaborate to lift boxes from a conveyor belt and sort them based on the adhesive tape color. Using VR enables us to collect data at scale due to the challenges of replicating the warehouse in a typical lab setting. We evaluate four versions of the robot responding to the event that the human misses a box—an idle robot that never lifts a box, a robot that attempts to lift human-missed boxes but does not apologize for a missed box, and versions of the robot embodied with female and male voices to issue apologies for a missed box. We conduct a human subject study with 32 participants, including 7 identifying as female, 1 Nonbinary, 1 Man, and 23 male to understand how perceptions change based on the gendered voice and expressed apology from the robot. We show significant differences in likeability, animacy, and perceived intelligence between idle robots and robots that issue apologies. We show that self-identified gender and prior experience in warehousing play roles in how participants perceive gendered voice and expressed apology in robots.
Autonomous cleaning auditing requires effective path planning to gather representative dirt samples over large areas. This paper presents a novel audit-path planning method using deep reinforcement learning (DRL) guided by priority maps. A 2D priority map is constructed from the spatial distribution of objects likely to contribute to dirt accumulation and provided as input to a DRL agent that learns to select motion primitives and sampling actions. Three DRL algorithms, namely A3C, PPO, and MARWIL, are evaluated in simulation with and without priority map guidance. Results show that incorporating the priority map improves learning efficiency, accelerates convergence, and yields better audit paths in high-priority regions. The trained policies are validated on BELUGA, an in-house developed audit robot, in a real indoor environment. The MARWIL-based policy achieved higher dirt sample collection per unit distance than PPO, confirming the practical benefit of priority map guidance for autonomous cleaning audits.
Sleep is essential for infant health and development, yet assessing infant sleep remains challenging because internal states must be inferred from observable behavior. This work presents a non-contact method for sleep-related behavioral inference based on facial analysis from video data. Unlike contact-based sensors or handcrafted feature approaches, the proposed system learns eye and mouth behavior directly from facial imagery and integrates these cues over time. Video frames are processed using facial region detection and landmark extraction, followed by MobileNet-based convolutional neural networks for binary classification of eye and mouth states. Frame-level predictions are aggregated through temporal decision logic to estimate sleep-related behavioral states, with blink dynamics and yawning serving as complementary indicators of transitional phases. Leveraging transfer learning enables efficient and robust inference suitable for real-world settings. Experimental results show reliable estimation of eye and mouth states and stable sleep-related behavioral inference. The proposed framework provides a practical, non-contact alternative to traditional infant sleep monitoring systems. Future work will address dataset expansion, robustness to occlusions and lighting variability, and the integration of multimodal inputs.
Supporting older adults in aging safely and independently at home is a key application area for service and assistive robotics. While commercial mobile manipulators provide a promising foundation, their ability to autonomously perceive and locate humans in unconstrained indoor environments, particularly for floor-level detection, remains an underexplored engineering challenge. In this work, we present a unified framework for autonomous monitoring and person-finding using the Hello Robot Stretch 3. We evaluate the robot's autonomous search and perception performance in distinguishing a person lying on the floor from a person who is seated or standing in its frontal view, enabling appropriate initiation of interaction or assistance. Experiments are conducted in a realistic, cluttered indoor environment show $\mathbf{9 0 - 1 0 0 \%}$ success rate with random robot starting positions. The results demonstrate the feasibility of using the Stretch 3 for assistive service tasks such as fall detection and in-home user monitoring, while also revealing perception challenges in cluttered scenes that must be addressed for robust real-world deployment. Our findings provide practical insights into leveraging commercial service robots for assistive aging-in-place applications.
Recent advances in radiance field representations, such as neural radiance fields (NeRF) and 3D Gaussian splatting (3DGS), have demonstrated remarkable performance in high-fidelity scene reconstruction and novel view synthesis. Along with improvements in visual quality, uncertainty estimation in reconstructed scenes has become crucial for reliability-sensitive applications. While conventional uncertainty estimation methods for neural networks have been applied to NeRF, Gaussian splatting approaches require modeling the parameters of their Gaussian primitives. Stochastic Gaussian splatting (SGS) introduced a Bayesian formulation for the positions, opacities, and colors of Gaussian primitives, enabling variational inference for uncertainty estimation. However, SGS omits scales and attitudes and replaces the negative log-likelihood with L1 and structural dissimilarity (D-SSIM) losses. In this work, we address these limitations by incorporating uncertainty over attitudes within a Lie group and extending SGS with a geometry-aware 2D Gaussian splatting (2DGS) pipeline for more accurate reconstruction with fewer parameters. Furthermore, we adopt a Laplace negative log-likelihood and area under the sparsification curve (AUSC) losses to maintain rendering quality after variational inference training. Experiments on the Mip-NeRF 360 dataset demonstrate that our approach enhances rendering fidelity while preserving uncertainty estimation, offering a more robust and geometrically informed method for scene reconstruction. In addition, an ablation study provides insights into the effects of the proposed losses offering guidance on balancing desirable performance factors, including rendering fidelity, normal consistency, and uncertainty estimation.
As robots become increasingly integral to every-day environments, they must interact with humans safely and naturally, especially during object handovers. We propose an approach to generate robot-to-human (R2H) handovers by being informed by how human givers hand objects to receivers. Our handover approach addresses the pre-handover phase of robot movement by predicting and performing optimal robotic grasp aligned with human giver grasp, and using data on how humans pose objects at transfer to predict the optimal object transfer pose and move the robot end effector to position the object at the predicted transfer pose. We use the HOH-Grasps dataset, a large-scale dataset of human-to-human handover demonstrations augmented with robot grasps at the transfer pose to learn how to conduct grasp and transfer pose in a manner that human givers do. We contribute a novel attention-based neural architecture that uses point cloud input on the object and giver hand anchor prediction to generate a frame-independent transfer pose representation consisting of a 6D rotation output and translation vector. Evaluations of our handover approach using HOH-Grasps and through real-world robot experiments show improved outcomes in terms of quantitative metrics over object pose and robot grasp prediction baselines.
Infant soothing is a major concern for new parents, while the prenatal environment exposes the fetus to continuous rhythmic stimuli such as breathing, heartbeat, and locomotor motion. Current commercial baby rockers offer generic, repetitive movements that fail to replicate the complex motion patterns associated with human walking and breathing. This paper presents the design and development of a novel biomimetic system capable of learning and reproducing personalized walking and respiratory patterns. The system comprises two core components: a 3-DOF (Degrees of Freedom) robotic platform utilizing four mecanum wheels for planar motion (X, Y, rotation) and a scissor lift mechanism for vertical motion (Z-axis), and a data acquisition and processing pipeline. An accelerometer-based sensor captures kinematic data during rest and ambulation. This data is then filtered and fed into a machine learning algorithm that extracts individual motion signatures. The algorithm then translates these signatures into control commands for the platform, effectively recreating personalized gait and breathing rhythms. Preliminary results from a study with 10 participants demonstrate the system's feasibility in capturing and replicating personalized motion patterns, paving the way for future personalized infant-soothing devices.
Ship recycling is the most sustainable and economical option for disposing of obsolete ships because it allows ship parts to be reused, recycled, and upcycled, hence reducing carbon emissions by providing scrap steel to the industry. Yet, the ship-breaking process involves multiple hazards that impact workers' health and safety. This paper investigates workers' perceptions of two ship-dismantling robots designed to reduce workers' exposure to hazardous conditions. The robots consist of a mobile manipulator for internal cutting and a magnetic crawler for cutting ship hulls, both of which are being developed as part of the SHEREC project. We interviewed 12 workers with ship-dismantling experience from different shipyards to investigate potential benefits and drawbacks of using both robots for dismantling ships. The results suggested that both robots could improve the participants' safety by reducing the exposure to toxic fumes and fires. The magnetic crawler received more favorable feedback from the participants, as it could eliminate the need for workers to work at height and enhance hull-cutting efficiency. Regarding potential drawbacks, the participants noted that both robots lack the flexibility required to cut ship frames. The participants identified additional challenges for the mobile manipulator, such as the problem with moving on uneven or obstructed surfaces and the limited reach of the manipulator arm. Finally, the participants also suggested three modifications to make the robots more suitable for ship dismantling, such as using batteries as an energy source, increasing the robots' structural durability to prevent damage, and adding fire-extinguishing capabilities to the robots.
This paper explores whether cognitive load affects intervention behavior of a driver in terms of taking control of the driving situation in autonomous driving. A video of a simulated autonomous driving scenario from a driver's perspective including critical situations is presented to the participants. In parallel, extraneous cognitive load is induced by letting participants mentally solve arithmetic problems as a secondary task. Whenever participants feel that they want to intervene and take control of the driving situation, they are required to press the button of a stopwatch. In contrast to expectations of increased impulsive behavior, results reveal an inverse association of cognitive load and counts of interventions paired with increased trust in the autonomous vehicle, but reduced perceived safety and situation awareness. This behavior may be explained with impaired switching of attentional focus and lack of possibility of internalization of the primary task under depletion of cognitive resources. The observed effects are suggested to be accounted for towards shared autonomy concepts in order to promote the design of safe autonomous vehicles and their integration in society.
This paper presents a theoretically grounded approach to multi-agent decision aggregation in robotics based on online learning of the classical Weighted Majority Rule (WMR). Unlike prior methods that require known agent reliability or use heuristic weight updates, the proposed approach learns agent competence probabilities online and applies the provably optimal WMR aggregation rule with log-odds weights. This enables agents with persistent errors to naturally acquire negative influence on group decisions while maintaining convergence guarantees. Through comprehensive Monte Carlo simulations with 100 independent trials across 20 agents and 100 decision rounds, results demonstrate a mean accuracy of 99.3% compared to 95.0% for simple majority voting. Critically, the method achieves 500 out of 2000 negative weights (25%), all correctly assigned to poor agents with 100% precision. The approach operates online without prior knowledge of agent reliability and provides a principled framework for multi-robot coordination and sensor fusion.
This study investigates how robot categorization is dynamically reconfigured in children's group interaction. Drawing on Membership Categorization Analysis and Conversation Analysis, we examine a video-recorded photo-sharing activity mediated by the social robot Haru. Rather than treating the robot's role as fixed by design, we analyze how children sequentially and multimodally negotiate what the robot is, what it knows or can do, and how it should be related to in the unfolding interaction. The findings demonstrate that children rapidly shift between treating the robot as person-like, socially accountable, knowledgeable, or mechanical, sometimes within only a few conversational turns. When oriented to as person-like, the robot becomes subject to moral evaluation and affiliative touch; when treated as knowledgeable, it is addressed as capable of understanding and procedural action; when framed as mechanical, its conduct is attributed to circuitry rather than comprehension. These categorial reconfigurations occur without corresponding changes in the robot's behavior, revealing an asymmetry between children's flexible sense-making and the robot's fixed system responses. Building on this analysis, we propose a three-domain model of dynamic robot categorization encompassing ontological, epistemic, and social-relational orientations. The study highlights that robots in group settings are not socially stable entities but are continuously re-specified in interaction. This interactionally grounded perspective contributes to research on socially impactful robotics by demonstrating how robot participation is collaboratively accomplished and renegotiated in practice.
Stroke survivors often develop hand and finger spasticity, limiting independence and quality of life. This paper presents a soft robotic hand-rehabilitation device inspired by party blowers, fabricated from Thermoplastic Polyurethane (TPU) using Fused Deposition Modeling (FDM) 3D printing, and featuring an inflatable chamber that replicates the blower's extensibility. We describe the device's design, optimization via computational simulation, fabrication, and experimental validation through force–resistance testing. When pressed against a calibrated force gauge at multiple angular positions along its curved profile, the device generates up to 19.7 N at 40 psi (275 kPa), corresponding to 94% of the maximum voluntary extension force benchmark (20.9 N) reported for stroke populations. Under graded resistance at clinically relevant angles corresponding to finger joints, it produces 2.7–4.5 N at 100-150 kPa (14.5–22 psi) at 45°, aligned with the proximal interphalangeal (PIP) joint. These engineering results establish a quantitative foundation for future human-subject evaluation, and its intuitive design, validated force–pressure behavior, and compact form factor offer potential advantages over existing glove-based rehabilitation systems.
Research in child-robot interaction indicates that robots are becoming embedded in the learning and social environments, shaping how future generations relate to and participate in increasingly robot-mediated societies. As robot manipulators become more prevalent in collaborative, assistive, and domestic contexts, understanding how children engage with such systems is relevant not only for education but also for broader questions of participation in sociotechnical life. Despite this, little is known about the factors associated with how children perform manipulator programming tasks. These tasks require operating digital interfaces, structuring action sequences, and anticipating a robot's movements in space. In an empirical study with 63 children aged 9-10, we explored whether digital habits such as frequency of computer use, mobile interface use, along with playing video games, were co-varied with their performance in structured robot manipulation tasks. Children completed pick-and-place tasks in a within-subject design with two robot arms: one programmed through a combination of haptic hand guidance and a computer interface, and one programmed via a tablet interface. Results showed that frequent use of mobile interfaces was associated with fewer programming errors and with perceived task difficulty, and task difficulty was moderated by robot interface, with the tablet-based robot perceived as more difficult. These findings provide an initial empirical indication that everyday digital experiences may shape how people approach robot manipulator interfaces, highlighting prior interface exposure as a relevant factor for inclusive and accessible human-robot interaction.
