PaperlayerView exact revision on arXiv

Source-linked AI summary

Navigating human-robot interaction challenges in private and residential aged care homes

David Ahmedt-Aristizabal, Mohammad Ali Armin, Akila Pemasiri, Harshala Gammulle, Tharindu Fernando, David Silvera-Tawil, Simon Denman

arXiv:2610.31149v1cs.HCcs.CVcs.RO

TL;DR

RAC and private homes pose demanding HRI challenges because of variable environments, diverse needs, limited technical support, and strong expectations for safety and privacy. The chapter examines technical, human, and organisational constraints, concluding with a practical roadmap toward scalable, regulation-ready evidence and deployments.

  • Problem

    RAC and private homes make HRI challenging because daily variability, diverse resident needs, and limited technical support coincide with strong safety, privacy, and usability expectations.

    Source: §1 · p. 1

  • Method

    The chapter examines the technical, human, and organisational constraints shaping real-world deployment of robots and ambient intelligence in both settings.

    Source: §9 · p. 25

  • Results

    The chapter concludes that real-world HRI requires systems that remain robust, transparent, and respectful, alongside shared foundations for interaction, technology, privacy, and benchmarking.

    Source: §9 · p. 25

  • Takeaways & Limitations

    A staged roadmap proposes shared event schemas and minimal data flows, multi-month privacy-aware personalisation, and multi-site studies linking technical changes to practical outcomes.

    Source: §8.5 · p. 25

    Datasets remain scarce for privacy-labelled HRI, long-horizon multimodal recordings with ground truth, older and atypical speech, and linked robot, sensor, and outcome data.

    Source: §7.3 · p. 23

Abstract

from arXiv · show

This chapter explores how robotic systems and ambient intelligence can enhance care for older adults in residential aged care (RAC) homes and private homes. Although these settings differ, we frame both as extreme and challenging environments for human-robot interaction (HRI), characterised by cluttered and dynamic layouts, diverse resident needs, and limited on-site technical support. We review embodied and non-embodied systems that assist with daily living, safety monitoring, and social engagement, highlighting hybrid deployments where robots complement fixed and wearable sensors. To ensure these technologies are usable and effective, design considerations consider the cognitive, physical, and social needs of older adults. Interaction design prioritises accessibility for older adults with low digital literacy, using robust but simple voice interfaces, small sets of non-verbal cues, accessible screen-based companions, and safe object handovers. Privacy, trust, and acceptability are addressed via practical strategies such as on-device inference, data minimisation, visible consent controls, and transparent data handling aligned with emerging artificial intelligence (AI) and privacy regulation. Evaluation emphasises real-world metrics including false alarms per day, time to recover from faults, staff workload, and long-term system reliability. We close by outlining open research challenges and future directions for scalable, regulation-ready deployments that support aging with dignity, safety, and independence.

1 Introduction

Human–robot interaction in residential aged care and private homes is promising but challenging: variable, cluttered settings and diverse resident needs coincide with limited technical support and strong expectations for safety, privacy, and usability. These conditions shape what systems can sense, store, and share, and make robust operation and recovery central design concerns.

1 Introduction

  • Private homes vary in layout and lighting, while residential care adds shared spaces, rotating staff, and formal workflows, making both extreme HRI environments.Robots and AI may support everyday tasks, social engagement, and safety monitoring while preserving autonomy.Source: §1 · p. 1
  • Privacy and safety constrain what systems can sense, store, and share in daily life.Clear consent controls, data minimisation, transparent behaviour, fault tolerance, and recovery are presented as ways to support trust and reduce caregiver burden.Source: §1 · p. 1
  • The chapter takes a primarily technical perspective, focusing on sensors, perception algorithms, and robotic systems for safe and effective operation.The supplied passage ends before completing its qualification of this focus.Source: §1 · p. 1

2 Navigating human–robot interaction challenges

The chapter reviews robots and ambient sensing for private and residential aged-care homes, with attention to practical deployment challenges and the conditions needed for sustained acceptance. It organises design considerations, privacy strategies, and real-world evaluation into a roadmap for usable, trustworthy, regulation-ready systems.

2 Navigating human–robot interaction challenges

  • The chapter organises the design space around perception in lived-in spaces, interaction modalities for diverse users, and deployment patterns that can scale.Its review considers where embodied systems are needed and how they interact with people and care infrastructure.Source: §2 · p. 2
  • Privacy and trust requirements are translated into strategies including on-device inference, minimal data flows, visible consent and mute controls, and transparent data handling.These choices are aligned with emerging AI and privacy regulation.Source: §2 · p. 2
  • The chapter proposes real-world evaluation considerations spanning technical performance, safety, usability, and long-term reliability.The planned evaluation moves beyond accuracy to implementation, safety events, usability, and reliability.Source: §2 · p. 2
  • Its roadmap aims to support robots and ambient intelligence systems that are usable, trustworthy, and regulation-ready, enabling older adults to age with dignity, safety, and independence.The chapter also identifies open challenges and future directions, including multimodal fusion, privacy-preserving learning, and safe operation in confined, cluttered spaces.Source: §2 · p. 2

2 Background, definitions, and constraints

This section establishes the system categories, settings, stakeholders, and physical and human constraints that shape HRI in residential care and private homes. It emphasizes that deployment requires solutions adapted to distinct care contexts, varied users, and ethical and regulatory expectations.

2 Background, definitions, and constraints

  • The section frames its background around system categories, settings, stakeholder roles, and the physical and human constraints shaping practical HRI.These definitions establish a shared basis for comparing technologies and deployment.Source: §2 · p. 3

2.1 Settings and deployment context

  • Residential care offers shared spaces, rotating staff, and formal procedures, whereas private homes vary widely and typically have minimal technical support.Clinical tasks require stricter documentation and oversight; both clinical and non-clinical tasks should benefit users and carers while remaining low-effort to operate and maintain.Source: §2.1 · p. 4

2.2 Robotic and ambient systems considered

  • The chapter groups assistive technologies into socially assistive, assistive, teleoperated, service, and ambient-intelligence systems, while noting that categories can overlap.A mobile robot may be combined with ambient sensors and a smart display for reminders, safety monitoring, and family communication.Source: §2.2 · p. 4

2.3 Stakeholders and roles

  • Residents, staff, families, managers, vendors, and regulators shape deployment, and their priorities can conflict over data, automation, human contact, and workload.The chapter uses a task, risk, and burden lens to make trade-offs explicit, including cognitive load and data exposure.Source: §2.3 · p. 5

2.4 Physical and environmental constraints

  • Private homes pose variable layouts, lighting, noise, clutter, and obstacles; residential care adds shared spaces, equipment, infection control, and changing furniture and workflows.Across both, robust perception, adaptive mapping, conservative motion planning, and often sensor fusion are identified as important for reliability.Source: §2.4 · p. 5

2.5 Human factors

  • Older adults’ digital-literacy, sensory, and cognitive barriers call for intuitive interaction, minimal setup, simple cues, transparent sensing, and graceful fault handling.The design guidance includes short prompts, large fonts, actionable alerts, local operation during network loss, and reversible personalisation; configuration should fit the roles and support available in each setting.Source: §2.5 · p. 5

2.6 Ethical, legal, and cultural considerations

  • Ethical deployment requires data minimisation, transparency, and informed consent, with architecture choices such as on-device inference and short retention.Governance frameworks use risk-based duties, while cultural adaptation can improve comfort and engagement only when transparent; perceived manipulation or uncanniness can reduce trust.Source: §2.6 · p. 6

3 Interaction modalities and integration

Interaction in aged care must accommodate noisy, cluttered and variable environments, diverse resident abilities, and limited technical support. The chapter combines accessible interaction patterns with hybrid system integration and deployment evidence, emphasizing that reliability and uptake depend on fit with care routines and workflows.

3.1.1 Verbal interaction

  • Older-adult speech interfaces face noise, overlapping conversation, and speech variation; safety-critical flows should confirm actions, offer tablet fallback, and use short, unambiguous utterances.On-device voice activity detection and beamforming can improve reliability; press-to-talk, diarisation, and barge-in suppression offer further protections.Source: §3.1.1 · p. 7

3.1.2 Non-verbal interaction

  • Non-verbal interaction should use a small, consistent set of slow, exaggerated cues, backed by lighting or large screen icons when glare, clutter, occlusion, or low light make gestures hard to read.Physical handovers can announce intent and retreat when unexpected contact is detected; docking and charging should avoid spill zones and walking-aid paths.Source: §3.1.2 · p. 7

3.2 Integration with the home and care ecosystem

  • Robots can combine ambient context signals with their own mobile sensing and actuation, for example pairing ambient fall detection with a robot's verbal prompt or physical assistance.For clinical data exchange, the chapter recommends FHIR-based interfaces, minimal event payloads, audit logs, and edge processing with upstream summaries.Source: §3.2 · pp. 7–8, Figure 4 · p. 8

3.3 Design patterns and deployment insights

  • Field deployments favour scheduled interactions, quick caregiver customisation, fallback modalities, and operational instrumentation, while early temi reports describe interest in scheduled, staff-curated activities and check-ins.The chapter identifies measures including docking success, false speech activations, alert rates, false alarms per day, and time to recover after faults.Source: §3.3 · p. 8, §3.3.5 · p. 9

3.3.1 Mobile social robots in RAC homes.

  • Pepper was used about 1 h/day in a three-month RAC study, while staff uptake depended on easy customisation and was constrained by reliability and time pressure.A lightweight mobile app for configuring content and behaviours across facilities increased appropriation and reduced reliance on technical staff.Source: §3.3.1 · p. 8

3.3.2 Telepresence for connection and clinical touchpoints.

  • Telepresence robots support social contact and remote clinical interactions when navigation is smooth and controls are simple, with routine use also requiring workflow fit, training, and clear escalation.Repeated exposure can improve comfort and perceived usefulness, while in-person contact remains important; scheduled sessions and a short task menu reduce cognitive load.Source: §3.3.2 · p. 8

3.3.3 Companion/therapeutic robots.

  • PARO sessions can reduce agitation and improve mood when structured into activity schedules, with sustained benefits reported when programs and staff training continue.The chapter argues companion devices should be delivered as programs with defined dose and frequency, rather than left for ad hoc use.Source: §3.3.3 · p. 8

3.3.4 Home-focused social companions.

  • ElliQ deployments report reduced loneliness and improved daily engagement with proactive, forgiving dialogue and handoffs to phone or video when needed.Design notes favour multimodal prompts and clear boundaries; anthropomorphic cues may build trust or trigger discomfort depending on context and expectations.Source: §3.3.4 · p. 9

4.1 Sensors and devices

Sensor choice shapes system capability as well as privacy, safety, cost, and usability in care environments. The chapter surveys visual, auditory, tactile, physiological, environmental, and infrastructure options, noting their distinct strengths and deployment constraints.

4.1 Sensors and devices

  • The sensor taxonomy spans assistive/social robots, telepresence robots, domestic service robots, and ambient intelligence, while design choices reflect the constraints of lived-in care settings.Those constraints include clutter, variable lighting, mobility aids, diverse sensory abilities, privacy, safety, cost, and usability.Source: §4.1 · p. 9, Figure 6 · p. 10

4.1.1 Vision-based systems

  • Vision sensors support scene understanding, recognition, navigation, and social interaction, but privacy safeguards and modality-specific limits shape deployment.RGB cameras offer versatile visual information but face privacy, lighting, and compute constraints; RGB-D adds geometry yet remains sensitive to lighting and reflective surfaces, while LiDAR navigates without lighting dependence but costs more and lacks semantic detail.Source: §4.1.1 · p. 10

4.1.2 Auditory and speech-based sensing

  • Auditory systems support speech, environmental-event detection, alerts, and conversation, but noise, multi-speaker interference, and intrusive continuous monitoring constrain their use.Short buffered windows, on-device processing, and event summaries can reduce exposure; speech output should be controllable and coordinated with visual or contextual cues.Source: §4.1.2 · pp. 10–11

4.1.3 Tactile and physical interaction sensing

  • Tactile, force, and proximity sensing support safer physical interaction, but calibration drift, wear, limited spatial resolution, and environmental sensitivity require design care and complementary sensing.Artificial skins can detect contact location, magnitude, and texture, but durability under wear, temperature, and humidity remains a challenge in domestic settings.Source: §4.1.3 · p. 11

4.1.4 Physiological and environmental sensing

  • Wearable and ambient physiological or environmental sensors provide health and context signals, but accuracy, calibration, comfort, privacy, and false alarms remain concerns.Physiological readings can vary with motion, sweat, and skin contact; environmental streams may drift and require sensor fusion for actionable interpretation.Source: §4.1.4 · p. 11

4.1.5 Technology enabling infrastructure

  • Enabling infrastructure should support coordination while prioritising local inference, minimal data flows, and resilience when networks are lost.The underlying infrastructure includes IoT frameworks, wireless networks, edge and cloud computing, and data management architectures.Source: §4.1.5 · p. 11

4.2 Scene understanding and mapping

Scene understanding must support safe, context-aware operation in dynamic homes and care settings, where clutter, changing layouts, lighting, and occlusion challenge perception. The chapter describes multimodal mapping and privacy-preserving representations, while identifying generalisation and embedded-compute limits.

4.2 Scene understanding and mapping

  • Multimodal sensing combines visual, depth, and non-visual information, while semantic SLAM links geometric maps with object labels and spatial relationships.Complementary inertial, thermal, radar, and ambient IoT signals can add resilience when visual sensing is impaired.Source: §4.2 · p. 12
  • Clutter, dynamic lighting, and frequent occlusions degrade scene understanding, while changing layouts and long-term environmental or behavioural shifts complicate deployment.Models trained on public datasets may fail to generalise without extensive adaptation, and robots cannot assume static maps over time.Source: §4.2 · p. 12
  • Privacy-preserving maps can replace raw RGB storage with geometric, semantic, blurred, downsampled, or topological representations that retain task-relevant navigation information.Topological maps can encode key places such as a bed area or bathroom entrance without detailed imagery.Source: §4.2 · p. 12
  • Accurate real-time perception on embedded hardware is difficult under power and latency constraints, particularly during continuous multimodal processing in shared spaces.The chapter identifies efficient architectures, model compression, and on-device inference as practical deployment strategies.Source: §4.2 · p. 12

4.3 Human perception

Human-perception systems support monitoring and interaction tasks, but care settings create persistent challenges for robustness, generalisation, privacy, and real-time operation. The chapter reviews methods for detection, tracking, pose and gesture recognition, fall detection, and affect recognition, alongside their remaining limits.

Human detection and tracking

  • Human detection and tracking support corridor monitoring and return-to-bed checks, with multi-camera tracking handling occlusions and illumination changes but resource and environmental robustness challenges remaining.LiDAR-assisted and radar-based detection are emerging privacy-preserving alternatives with resilience under poor lighting and clutter.Source: §4.3 · p. 12

Pose and gesture recognition

  • Pose and gesture recognition enables fine-grained HRI, but occlusion and real-world generalisation remain open challenges where mobility aids, furniture, and limited motion complicate interpretation.Approaches range from 2D keypoints to 3D reconstruction and hybrid sensor fusion.Source: §4.3 · p. 12

Fall detection

  • Fall detection combines pose estimation and anomaly detection, with multimodal radar or depth sensing addressing some RGB limitations, while real-world fall-data scarcity and unobtrusive monitoring remain barriers.The chapter also notes lightweight edge deployment strategies for local processing.Source: §4.3 · p. 12

Affect recognition

  • Affect recognition increasingly combines pose, context, and visual, vocal, or physiological signals, yet data availability, cultural variation, privacy, and consent remain challenges.The chapter flags particular ethical concerns when systems infer affect implicitly rather than through explicit check-ins.Source: §4.3 · p. 12

4.4 Learning paradigms in robotics for RAC and home care

The chapter reviews learning approaches for adapting robots to diverse, data-scarce homes, from labelled-data training to domain adaptation. It describes their care applications alongside limits in generalisation, privacy, reliability, and practical deployment.

4.4.2 Supervised learning approaches

  • Supervised models learn from labelled multimodal inputs, supporting activity, speech, and gesture recognition in care settings.Creating labels is costly, and models trained in labs may fail in homes or struggle with rare activities and atypical communication patterns.Source: §4.4.2 · p. 13
  • 97.3% accuracy: DiG-Net uses depth-conditioned alignment for long-range gesture recognition.GestLLM combines MediaPipe features with language models to interpret culturally nuanced gestures without retraining.Source: §4.4.2 · p. 13

4.4.3 Weakly-supervised and self-supervised learning approaches

  • Weakly-supervised learning uses coarse or incomplete labels, while self-supervised learning creates predictive tasks from unlabelled data to reduce annotation costs.Examples include timestamp-level action segmentation and speech representations learned by predicting masked audio or video segments.Source: §4.4.3 · p. 14
  • WSL can inherit noise from heuristic labels, while SSL faces difficult transfer to downstream tasks, substantial pretraining costs, and privacy and interpretability concerns.Both can still struggle with environmental and cultural variation.Source: §4.4.3 · p. 14

4.4.4 Few-shot and meta-learning approaches

  • Few-shot learning adapts robots from a handful of examples, and meta-learning trains them to adapt across tasks, supporting low-effort personalisation during onboarding or routine changes.Applications described include learning manipulation tasks from demonstrations and adapting preferences from minimal user feedback.Source: §4.4.4 · p. 14
  • Few-shot and meta-learning may generalise poorly to culturally nuanced activities, while opaque or brittle adaptation and unclear feedback complicate oversight and can risk dignity.Training across diverse tasks may also require costly datasets or simulations.Source: §4.4.4 · p. 15

4.4.5 Domain adaptation for home environments

  • Domain adaptation targets distribution mismatch through invariant representations, alignment, self-training, and selective requests for caregiver or clinician labels.It is described as supporting perception, manipulation, and interaction across homes where supervised models may degrade under target shift.Source: §4.4.5 · p. 15
  • Domain adaptation can cause negative transfer or weaken fine-grained semantics, while noisy pseudo-labels may propagate errors under changing home conditions.Long-term adaptation also faces sparse or delayed rewards, drifting preferences, constrained safe exploration, and limited target data under privacy and consent requirements.Source: §4.4.5 · p. 15

5 Privacy, trust and acceptability

Privacy, trust, and acceptability shape what robots and ambient systems can sense and how people respond to them in private and RAC homes. The chapter connects these requirements to safeguards, interaction choices, and stakeholder collaboration.

5 Privacy, trust and acceptability

  • Privacy, trust, and acceptability are linked: unclear privacy can erode trust, and low trust can lead to rejection even when systems are technically capable.The chapter defines trust through reliability and perceived intentions, and acceptability through contextual suitability, usefulness, and emotional comfort.Source: §5 · p. 16

5.1.2 Challenges

  • Always-on sensing and silent uploads complicate ongoing consent, while residents differ in whether they accept monitoring for safety; settings should therefore be visible and reversible.The passage recommends conservative defaults and simple ways to pause sensing or narrow its scope.Source: §5.1.2 · p. 16

5.1.3 User perspectives

  • Older adults value safety features but resist constant recording; visible sensing states, short retention, explanations, and options to review or delete captured data can address this trade-off.Shared care settings also raise questions about who sets rules and resolves disagreements.Source: §5.1.3 · p. 16

5.1.4 Policy and legal frameworks

  • The European Union AI Act, ISO/IEC guidance, Australian privacy duties, and WCAG 2.2 provide frameworks for risk management, data governance, transparency, oversight, and accessible controls.The chapter presents these as shared baselines for technical and clinical teams.Source: §5.1.4 · p. 16

5.1.5 Privacy-by-design

  • Privacy-by-design limits data collection, favors on-device processing and summaries over raw streams, and uses consented sharing, visible indicators, pause controls, and deletion schedules.Federated learning is identified as a way to improve models without pooling raw personal data.Source: §5.1.5 · pp. 16–17

5.2.2 Design considerations

  • Performance-based trust concerns reliability, while relational trust concerns respect and honesty; trustworthy interactions should explain decisions, offer reversible controls, and defer to a person when confidence is low.Transparent states, graceful recovery, and emotionally appropriate design are also identified as daily trust factors.Source: §5.2.1 · p. 17, §5.2.2 · p. 17, §5.2.3 · p. 17

5.2.4 Stakeholder collaboration

  • Trust requires distinct supports for residents and families, care workers, operators, and designers, including consent controls, actionable alerts, logs, and privacy-preserving field learning.Short co-design cycles and transparent updates support trust as it evolves over time.Source: §5.2.4 · p. 17

5.3 Acceptability

The chapter frames acceptability as a combination of usability, ethical fit, perceived benefit, and emotional response. It links user studies and co-design to sustained engagement, explicit trade-offs, and practical risk governance.

5.3.1 Frameworks

  • Acceptability models converge on perceived effectiveness, burden and usability, ethical and relational fit, and affective engagement.These domains recur across TFA, Almere, and TAM/UTAUT model families.Source: §5.3.1 · p. 17, Figure 8 · p. 18

5.3.2 User studies

  • Acceptance improves when benefits are immediate and visible, while proactivity bounded by user opt-out can support acceptability.Examples include easier family contact, reminders, and fall alerts that work without constant recording; negative experiences can stem from frustration, interruption, or feeling monitored.Source: §5.3.2 · pp. 17–18

5.3.3 Influencing factors

  • Culture, language, and prior experience shape acceptance; friendly forms, multilingual prompts, customisable names, and caregiver configuration tools can fit diverse homes and workflows.Staff and family training is described as critical.Source: §5.3.3 · p. 18

5.3.4 Long-term engagement

  • Novelty effects can fade after the first weeks, so sustained use calls for measuring participation over months and refreshing content regularly.The passage also cautions against close human imitation and frames robots as part of a people-led care program.Source: §5.3.4 · p. 18

5.3.5 Co-design methods

  • Co-design with older adults and caregivers can reveal preferences, cultural norms, usability barriers, and privacy expectations that standard testing may miss.Early involvement is associated with more trusted and usable systems and may avoid costly redesign after field deployment.Source: §5.3.5 · p. 18

5.4 Intersections and trade-offs

  • Privacy, trust, and acceptability trade-offs vary by setting and stakeholder, so the chapter recommends making compromises explicit and negotiated.Residents may prioritize agency and dignity, while staff emphasize safety and efficiency and families may prefer more monitoring; Fig. 9 depicts evaluation shifting toward safety, workflow, privacy, and long-term reliability as deployment progresses.Source: §5.4 · p. 18, Figure 9 · p. 19

5.5 Risk management and governance

  • Risk governance links intended use and identified risks to auditable controls, fault recovery, human oversight, and regular review of field metrics with caregivers.Suggested metrics include near misses, false alarms per day, docking success, and time to recover after faults.Source: §5.5 · p. 19

6 Benchmarking and evaluation

Evaluation should progress from offline tests to long-term deployments, carrying operational, safety, privacy, and usability measures across stages. Accuracy alone cannot capture the burdens and reliability that shape daily use in aged care.

6 Benchmarking and evaluation

  • A staged evaluation ladder links interaction and privacy choices to comparable outcomes, tracking safety events, false alarms, caregiver workload, fault recovery, privacy friction, and sustained use across sites.Social robots also need measures of acceptability, usability, discomfort, and engagement.Source: §6 · p. 19

6.1 Simulations and offline evaluations

  • Offline simulations can stress-test perception and planning under clutter, occlusion, and lighting variation, but their results are prerequisites for lab trials, not proof of real-world reliability.Fig. 10 illustrates a care-home simulation, mapping and trajectory views, and a replay dashboard for safety and performance analysis.Source: §6.1 · p. 19, Figure 10 · p. 20

6.2 Lab-based trials

  • Lab trials validate safety-critical interaction loops, sensor calibration, consent, emergency stops, approach distances, and fault recovery before exposure to full household or ward complexity.Consent should be plain-language, proportionate, and capacity-aware, with ongoing checks and data minimisation for audio or video.Source: §6.2 · p. 20

6.3 Pilot studies

  • Pilots run for weeks or months to assess workflow fit, false-alarm burden, early acceptability, and reversible privacy controls; care-influencing systems need transparent reporting and human oversight.Behaviour-change interventions should report content, dose, and facilitation to support reproducibility.Source: §6.3 · pp. 20–21

6.4 Long-term deployments

  • Long-term deployments expose maintenance and reliability gaps while measuring stable engagement, alert burden, and caregiver workload beyond novelty; they can also support randomised-controlled trials (Lu et al., 2025).Teams should log failures such as docking misses, network drops, and time drifts, including recovery time (Hawes et al., 2017).Source: §6.4 · p. 21

6.5 Measuring performance and metrics

  • Evaluation should add total cost of ownership, latency, energy, robustness, adaptability, fairness, and conversational safety to accuracy, while tracking installation, training, maintenance, and support burdens.Long-term studies report that simple behaviours with reliable docking and easy staff customisation can sustain use better than feature-rich systems requiring frequent intervention (Hawes et al., 2017).Source: §6.5.2 · p. 21, §6.5.1 · p. 21, §6.5 · p. 21

6.5.3 Privacy and security benchmarks

  • Privacy and security benchmarks should measure local versus cloud processing, summary versus raw-data logging, retention, anonymisation, and resilience to attacks, aligned with relevant privacy frameworks.Layered defences include encryption and identity management; runtime safeguards can include input validation and anomaly detection.Source: §6.5.3 · p. 21

6.5.4 Ethical and regulatory compliance checks

  • Reporting checks should cover AI-specific bias, drift, oversight, consent and privacy audits, explanation fidelity and comprehensibility; existing toolkits still leave HRI-specific gaps in social interaction, privacy, and long-term autonomy.SPIRIT-AI, CONSORT-AI, and TIDieR support transparent reporting and reproducibility, while future standards should combine technical, ethical, and human-centric indicators.Source: §6.5.4 · p. 21, §6.5.5 · pp. 21–22, §6.5.6 · p. 22

7 Open research challenges

The chapter identifies persistent deployment challenges across perception, personalisation, governance, sustainability, and equity in private and residential aged care homes. These challenges include limited long-term evidence, privacy and data constraints, and unequal access to infrastructure.

7.1 Robust multimodal perception in lived-in spaces

  • Perception systems must handle occlusion, noisy or conflicting multimodal inputs, changing conditions, and variation across homes while communicating uncertainty without increasing cognitive load or alarm fatigue.Navigation and manipulation also remain difficult in narrow, cluttered spaces, where safe approach distances and reversible handovers matter.Source: §7.1 · pp. 22–23

7.2 Personalisation without overfitting or privacy loss

  • Personalisation must adapt to diverse needs with limited data while preserving privacy, predictable behaviour, safe rollback, and auditability.Federated learning, continual adaptation, and uncertainty-aware updates show promise but remain difficult to deploy safely in small, heterogeneous datasets (Casado, 2024).Source: §7.2 · p. 23

7.3 Governance, transparency, and evidence generation

  • Governance lacks widespread long-term, multi-site evidence and shared benchmarks; few privacy-labelled or linked multimodal datasets capture older voices, outcomes, staff handovers, configuration changes, and consent over time.Normalising measures across sites must retain local privacy, workflow, and staffing context; existing reporting-guideline adoption in HRI remains uneven (Liu et al., 2020; Rivera et al., 2020).Source: §7.3 · p. 23

7.4 Operational and economic sustainability

  • Operational sustainability is difficult to compare because studies rarely report maintenance time, remote-support frequency, or failure categories; systems that are easier to maintain may outperform more capable alternatives in practice.Total ownership costs include installation, training, cleaning, consumables, software updates, and support; long-running deployments associate sustained use with simple behaviours, reliable docking, and easy staff configuration (Hawes et al., 2017).Source: §7.4 · p. 23

7.5 Equity, accessibility and cultural competence

  • Cultural competence requires interaction styles and staff-tailorable content that are transparent, accessible, and evaluated for fairness over months rather than days.Advanced robots and sensor networks also require costly hardware, reliable infrastructure, and specialist support, potentially limiting access for rural, remote, and low-income settings.Source: §7.5 · pp. 23–24

8 Future research directions

The chapter proposes cross-cutting research levers to move assistive HRI from prototypes toward dependable, regulation-ready practice. These include hybrid architectures, privacy-aware personalisation, staged evaluation, and ongoing co-design.

8.1 Reference architectures for hybrid robot and ambient homes

  • Shared modular architectures should combine robots and ambient sensing while making boundaries for perception, decisions, data sharing, retention, and clinical escalation explicit.Common event schemas and documented interfaces can support audit and safe handover, including use of health-data exchange standards such as FHIR (Pavão et al., 2023; Li et al., 2024a).Source: §8.1 · p. 24

8.2 Personalisation that stays private

  • Privacy-aware personalisation should be demonstrated over months with limited labelled data, clear consent, safety limits, rollback, uncertainty signals, and role-based controls.Promising approaches include federated or split learning and privacy-preserving training that does not pool raw personal data (Casado, 2024; Brauneck et al., 2023).Source: §8.2 · p. 24

8.3 Evaluation ladders that bridge the reality gap

  • Shared evaluation ladders should carry core measures from simulation through lab checks and micro-pilots to in-the-wild studies, with staged tests for clutter, occlusion, audio noise, and privacy controls.Health-style reporting can help reviewers assess bias, drift, human oversight, and missing-data handling (Liu et al., 2020; Rivera et al., 2020).Source: §8.3 · p. 24

8.4 Co-design with caregivers and residents that lasts beyond the pilot

  • Long-term co-design requires simple configuration tools for staff and families, recurring caregiver reviews, and reporting of setting changes and their effects on safety, false alarms, and engagement.Staff perceptions of workload, trust, and accountability can evolve over time, supporting co-design as an ongoing operational capability (Ruan et al., 2025; Otaka et al., 2024).Source: §8.4 · pp. 24–25

8.5 Summary

  • The proposed roadmap moves from shared schemas and minimal data flows toward multi-month personalisation studies and multi-site trials linking technical changes to care-relevant outcomes.Longer-term studies should use SPIRIT AI and CONSORT AI reporting and measure false alarms, caregiver workload, and recovery after faults.Source: §8 · p. 24, §8.5 · p. 25

9 Conclusion

Private and residential aged care homes are challenging HRI settings because they are cluttered, dynamic, and have limited support alongside strong expectations of safety, privacy, and dignity. The chapter argues for robust, transparent systems and hybrid deployments, evaluated through real-world measures, to support useful and respectful care.

9 Conclusion

  • Dependable HRI requires robust perception and interaction that account for occlusion, mobility aids, shifting layouts, and diverse sensory and cognitive needs.Privacy, trust, and acceptability should be built into system architecture and interaction design from the outset.Source: §9 · p. 25
  • Evaluation should extend beyond accuracy to include false alarms per resident per day, caregiver workload, recovery after faults, maintainability, and long-term reliability.Practical study ladders link simulation and lab tests to in-the-wild pilots and long-term deployments, helping compare systems across sites.Source: §9 · p. 25
  • Hybrid architectures, privacy-aware personalisation, evaluation ladders, and long-term co-design can help bridge prototypes and sustained use while supporting regulation-ready evidence and interoperability.The chapter's stated goal is assistive technology that supports ageing with dignity, safety, and independence.Source: §9 · p. 25

Related papers

Loading 2610.31149v1…