Toward User-Mediated Self-Repair in Ubiquitous Robots Through Goal-Oriented Agentic AI

I. INTRODUCTION

• Movement, perspiration, and varied pressure expose body-worn robots to mechanical strain,, while sustained physical contact can cause failures.The paper identifies user-facilitated self-repair as an underresearched opportunity for sustainable devices.Source: §I · p. 1
• Complex repair logic can overwhelm working memory and obscure dependencies, while branching tasks can exceed users’ immediate cognitive capacity,,,,.The paper contrasts these challenges with static manuals and linear tutorials, which often fail for self-repair tasks.Source: §I · p. 1

A. Agentic AI in a Ubiquitous Context

• The proposed multi-layered agentic AI mediates robot self-repair and lets the robot introduce its repair needs to the user –.The paper aims for an anthropomorphic interaction that frames repair as collaborative.Source: A. Agentic AI in a Ubiquitous Context · p. 1
• The physical testbed uses ten numbered wires, clips, a color-coded sticker, and a confirmation lamp in a repair procedure whose wire connections branch by sticker color.Participants interacted with an animated robot face during the repair.Source: Figure 1 · p. 1
• Concealed or eyes-free devices are poorly served by visual-only instructions; dialogue supports hands-on repair while users focus on tactile interaction,,,,,,.The paper describes this dialogue-based social-partner framing as a way to enrich learning and support sustainability,, –.Source: A. Agentic AI in a Ubiquitous Context · pp. 1–2

B. Open-Ended and Self-Repairing

• A strategic agent monitors repair objectives and context, allowing explanations to adapt to confusion or partial progress and letting users pause and resume without penalty.The framework replaces a fixed diagnostic tree with non-linear interaction.Source: B. Open-Ended and Self-Repairing · p. 2
• Task decomposition converts unconstrained explanations into goal and subgoal hierarchies that model dependencies and failure-recovery paths.The paper presents delegated task management as a way to turn technical repair into collaborative interaction.Source: B. Open-Ended and Self-Repairing · p. 2
• 95% task completion was achieved in the physical repair experiment, despite lower perceived social presence and trust than the online baseline (p < .05).Participants also reported high perceived helpfulness and post-interaction self-efficacy.Source: B. Open-Ended and Self-Repairing · p. 2
• The paper’s contributions include separating goal decomposition, persistent state tracking, strategic management, and conversational execution, and evaluating screen-free physical repair against a prior online baseline.The physical scenario required completion of a conditionally branched hardware task without a traditional screen-based instructional interface.Source: B. Open-Ended and Self-Repairing · p. 2

A. Speech-Driven Interfaces in Ubiquitous Robotics

• Speech interfaces progressed from rigid finite-state grammars to situated multi-turn dialogue for resolving ambiguity, while eyes-free wearables cannot rely on visual joint attention.Earlier audio-only augmented-reality work guided maintenance when visual attention was needed on machinery,.Source: A. Speech-Driven Interfaces in… · p. 2

B. Task-Oriented Agentic AI

• Prior work treats chain-of-thought decomposition as a basis for sequential repair and interleaves reasoning with actions, while SayCan grounds instructions in physical feasibility.The paper positions these as developments in task-oriented agentic planning.Source: B. Task-Oriented Agentic AI · p. 2
• The paper adapts a bifurcated slow-thinking and fast-thinking architecture to physical hardware, assigning structural logic to one agent and social fluency and latency reduction to another.The authors distinguish their setting from digital exploration and note that generic planners can lack specialized hardware constraints.Source: B. Task-Oriented Agentic AI · p. 3

A. System Architecture

• The system comprises goal decomposition, strategic, conversational, and persistence agents; it was evaluated through open-ended physical repair interactions with 20 university participants.The study included physical and online procedures, collected measures, and comparative analysis.Source: A. System Architecture · p. 3
• Before live interaction, the Goal Decomposition Agent uses Llama 3.1 70B to translate instructions into goals and subgoals, identifying dependencies and conditional branches.In the test task, sticker color determined the wire connections, followed by confirmation-bulb verification.Source: 1) Goal Decomposition Agent · p. 3

1) Goal Decomposition Agent:

• The decomposition agent’s goals are stored as unformatted plaintext, providing lightweight state tracking; high-level logic is derived dynamically from system state rather than static scripts.The persistent agent maintains those goals throughout the interaction.Source: 1) Goal Decomposition Agent · p. 3, 2) The Persistence handler Agent · p. 3

2) The Persistence handler Agent:

• The persistence handler stores conversation history, user-provided goal variables, and intermediate states to support continuity, error recovery, and conditional evaluation.The described implementation uses ChromaDB and separates memory management from strategy and immediate conversation.Source: 2) The Persistence handler Agent · p. 3

B. Conversation handling agent

• The conversational agent explains procedures, clarifies user input, and validates one active goal, then reports its completion without advancing the workflow itself.This reactive layer uses Qwen 30B for synchronous dialogue; overall structural logic remains with the strategic layer.Source: B. Conversation handling agent · p. 3
• The strategic agent tracks repair steps as pending, active, completed, or skipped, and uses user input and system state to select goals and enforce conditional progression.A centralized goal board holds the steps and their governing conditional logic.Source: 1) The Strategic Agent · p. 3

1) The Strategic Agent:

• The strategic agent prioritizes goals asynchronously and can revisit objectives needing clarification while the conversational agent responds around the active goal.Goal decomposition creates the initial hierarchy before interaction, and persistence stores conversation history, task variables, goal status, and intermediate states.Source: 1) Social Robotic Interface · p. 4, Figure 2 · p. 4

1) Social Robotic Interface:

• Blinking, smiling, and gaze shifts project social agency; eye contact triggers engagement, sound thresholds trigger Whisper transcription, and generated responses are sanitized for TTS.The interface uses a high-definition camera for face detection and removes special characters and emojis before vocalizing responses.Source: C. Experimental Procedure and Testbed · p. 4

• Twenty participants aged 18–44 completed a physical repair task using a wooden testbed with a status lamp, ten numbered wires, and color-coded task variations.Successful repair required identifying the sticker color, following configuration-specific wiring instructions, and confirming that the lamp illuminated after four correct connections.Source: 2) Physical study · p. 4
• The agentic interaction began with the robot introducing the repair problem and requesting help, and ended after users confirmed the status lamp was active.Participants first completed a questionnaire and a two-minute casual conversation to become familiar with turn-taking.Source: 2) Physical study · pp. 4–5
• Self-efficacy was assessed only after the interaction, so the measure captures perceived capability following system use but not change from a pre-interaction baseline.One participant's objective performance metrics were missing and excluded from objective comparisons.Source: 2) Physical study · p. 5
• The physical study was benchmarked against a prior online study using the same architecture and branching repair logic, but a digital rather than physical task interface.The comparison was intended as an initial indication of how situated physical interaction may affect perception.Source: 2) Physical study · p. 5

A. Task Completion and Behavioral Outcomes

• 95% of participants completed the physical repair task, objectively verified by termination of the repair sequence and confirmation of the status lamp.Nineteen of twenty participants succeeded despite variation in interaction length.Source: A. Task Completion and Behavioral… · p. 5

B. Perceived Helpfulness, Trust, and Social Presence

• The Social Presence and Trust and Competence composite scales had Cronbach’s α values of .80 and .93, respectively.Source: B. Perceived Helpfulness, Trust, and… · p. 5
• Trust and Competence averaged 4.87 (SD = 1.57), while Social Presence averaged 4.09 (SD = 1.45) on the 7-point scale.Trust and Competence had a medium effect size (d = 0.56), while Social Presence had a negligible effect size (d = 0.06), measured against the neutral midpoint of 4.0.Source: B. Perceived Helpfulness, Trust, and… · p. 5

C. Relationship Between Prior AI Familiarity and Post-Interaction Trust

• Prior AI familiarity was not significantly correlated with Trust and Competence (r = −0.07, p = .775).The authors state that trust was independent of participants’ prior AI familiarity.Source: C. Relationship Between Prior AI… · p. 5

D. Quantitative Analysis of Interaction Efficiency and Robustness

• 19.7 turns on average (SD = 20.6) exceeded the theoretical optimal path of 9 turns (t(19) = 2.27, p = .035).The high variability included interactions extended by chit-chat or repeated clarification, while others stayed close to the optimal path.Source: D. Quantitative Analysis of… · p. 5
• 10.5% of sessions involved linguistic code-switching and 21.1% involved divergent meta-queries; the system briefly engaged before redirecting or politely refused to refocus on repair.Table II presents system robustness and recovery against non-normative inputs.Source: D. Quantitative Analysis of… · pp. 5–6, Table II · p. 6

1) Baseline Methodology and Comparison:

• The physical study involved 20 participants and was compared with a 70-participant online study using a virtual wiring task.The baseline was included to assess agentic AI logic and examine how physical embodiment and hidden sensors affect user perceptions.Source: E. Comparative Baseline Analysis · p. 6, 1) Baseline Methodology and Comparison · p. 6

E. Comparative Baseline Analysis

• System Helpfulness was M = 4.60 in the physical implementation versus M = 4.21 online, while Perceived Understanding remained stable (M = 4.21 vs. M = 4.25).The comparison contextualizes the physical hardware testbed against the online baseline.Source: E. Comparative Baseline Analysis · p. 6, Table III · p. 6
• The physical task averaged 19.7 conversational turns versus 11.4 online, while participants reported Perceived Self-Efficacy of M = 4.95 after the physical task.The authors suggest the added turns reflect the physical and interactional demands of wire manipulation; the post-only measure does not demonstrate an increase in self-efficacy.Source: E. Comparative Baseline Analysis · p. 6
• Social Presence and Trust and Competence ratings were significantly lower in the physical task than online, while the physical system retained high System Helpfulness.The authors suggest real-time situated feedback was valued amid the tactile complexity of hardware repair.Source: E. Comparative Baseline Analysis · p. 6

V. DISCUSSION

• 95% task completion was achieved, despite a mean of 19.7 turns (SD = 20.6) versus the 9-turn theoretical optimum and 11.4 turns online.The authors attribute the longer interaction trend likely to speech-recognition latency and the physical manipulation of crocodile connectors.Source: A. Efficacy of Situated Ubiquitous… · p. 6
• The framework effectively mediated technical information in environments with reduced traditional visual feedback.The authors describe its near-total completion rate as evidence of robustness despite the interaction's temporal cost.Source: A. Efficacy of Situated Ubiquitous… · pp. 6–7

B. Perceptual Disparity: Virtual vs. Physical Environments

• Social Presence (M = 4.09) and Trust and Competence (M = 4.87) were significantly lower in the physical task than online (M = 5.55, p = .0005; M = 5.74, p = .037, respectively).The paper suggests the tangible task's physical demands and perceived consequences may have contributed to the lower ratings.Source: B. Perceptual Disparity: Virtual vs.… · p. 7

C. Goal-Oriented Safety and Limitations

• The physical evaluation used a conditionally branched wiring procedure and did not represent maintenance with interacting faults or substantial diagnostic uncertainty.The authors call for evaluation with larger samples, longer procedures, multiple fault types, richer diagnostic uncertainty, and repeated sessions.Source: C. Goal-Oriented Safety and Limitations · p. 7

VI. CONCLUSION

• The architecture transforms unconstrained natural language into a structured hierarchy of goals to guide non-experts through hardware repair via situated dialogue.Source: §VI · p. 7
• The framework maintained high task success and robustness to linguistic variations, but findings are initial feasibility evidence rather than proof of causal self-efficacy improvement or broad generalizability.Self-efficacy was assessed only after interaction, and the physical evaluation involved 20 participants.Source: §VI · p. 7, C. Goal-Oriented Safety and Limitations · p. 7