Personalization in Social Robots

Updated 2 February 2026

Personalization in social robots is the autonomous adaptation of behavioral, physical, and interaction traits to meet unique user needs in varied contexts.
Layered architectures integrate perception, user modeling, decision-making, and execution to adjust social cues and improve human–robot interaction.
Advanced methods such as reinforcement learning, federated learning, and theory-of-mind models ensure efficient adaptation while maintaining privacy and transparency.

Personalization in social robots refers to the autonomous adaptation of a robot’s behavior, appearance, and interaction strategies to meet the individualized needs, preferences, and expectations of each user within specific social contexts. This capability underpins effective, natural, and sustainable human–robot interaction (HRI) by enabling robots to modulate their social cues, engagement pacing, and action selection according to a user’s dynamic profile, history, and situational requirements. Personalization spans domains as varied as therapy, education, healthcare, companionship, and service provision, and is realized through an array of algorithmic, architectural, and user-modeling techniques that integrate multimodal perception, memory, personality, and responsible computing.

1. Foundational Models and Architectures

The architectural basis for personalization in social robots is typically a multi-layered, modular pipeline spanning perception, user modeling, decision-making, and execution. In ROS-based deployments such as BRILLO (Rossi et al., 2022), personalization arises from the integration of four layers:

Perception Layer: Aggregates biometric (face, body pose), speech, and sentiment data streams for real-time user identification and state estimation.
Beliefs/User Modeling Layer: Synthesizes short-term, working, long-term, and semantic memories to maintain persistent user profiles, historical preferences, and semantic associations (e.g., drink–ingredient graphs).
Decision-Making Layer: Employs context-aware finite state machines, influence diagrams, and graph similarity measures to select personalized actions, dialogue strategies, and physical gestures considering the current belief state.
Execution Layer: Realizes action plans as coordinated speech, facial expression, and physical behaviors, explicitly tailored to the user’s profile and live engagement state.

This tightly coupled loop enables the robot to update its knowledge of the user on each encounter, modulate interaction pacing, and deploy personalized content through synchronized multi-modal channels (Rossi et al., 2022, Asprino et al., 2020, Balossino et al., 8 Aug 2025).

2. User Modeling: Profiles, Preferences, and Memory

Personalization hinges on robust user modeling encompassing:

Profiles: Composite vectors of explicit (e.g., stated preferences, demographic metadata) and implicit (e.g., sensor-inferred engagement, feedback history) features. Feature vectors are updated either by weighted averaging (e.g., taste vectors in BRILLO) or probabilistic memory retrieval conditioned on robot personality (Matcovich et al., 2024).
Memory Systems: Episodic memory tracks recent interactions; semantic memory encodes persistent user traits and context patterns (Tang et al., 2 Feb 2025). In personality-driven models, retrieval biases and decay rates are parametrized by robot personality (e.g., Big Five OCEAN traits) such that, for example, conscientious robots retain more factual information, while neurotic robots selectively recall emotionally salient events (Matcovich et al., 2024).
Preference Encoders: In recommender-system approaches (Huang et al., 27 Jan 2026), profiles span long-term embeddings (via matrix factorization), session-based sequential embeddings (RNN/Transformer), and fine-grained attribute vectors (knowledge graphs, concept bottlenecks).
Self-Disclosure and Sparse Feedback: Explicit self-disclosure (open-ended prompts for emotional associations or task adaptations) is crucial for art therapy (Cooney, 2020) and general household assistance (Patel et al., 2024), enabling robots to capture high-variance, user-specific mappings that escape generic inference.

3. Algorithmic Methods for Personalization

Multiple computational strategies drive personalization:

Reinforcement Learning:
- In personalized stroke rehabilitation (Lee et al., 2020), RL agents dynamically select salient kinematic features for assessment and feedback, achieving patient-specific adaptation without full retraining for each individual. RL reward functions combine accuracy, action cost, and patient engagement.
- In multimodal emotion-aware HRI (Xie et al., 2021), RL agents optimize for both positive affect (estimated via fused multimodal signals) and regularize toward user’s affective baseline and intrinsic preferences.
Federated and Continual Learning:
- To scale personalization and preserve privacy, decentralized federated continual learning aggregates parameter updates (not raw data) across robots, leveraging Elastic Weight Transfer to reconcile inter-client knowledge sharing with intra-client retention (Guerdan et al., 2022). This method demonstrably reduces forgetting and enables rapid adaptation per user in socially-aware navigation.
Concept Bottleneck and Explanation:
- TAACo (Patel et al., 2024) achieves data-efficient generalization to open-set tasks by mapping robot actions and state features through human-meaningful abstract concepts (fragility, risk, mundaneness), training a Transformer on sparse user feedback, and producing faith-aligned explanations via attention mechanisms.
Theory-of-Mind Model-Based Control:
- Closed-loop controllers identify dynamical models of human perception, cognition, and decision-making, estimating invisible mental states in real time and steering robot actions to maintain engagement and regulate frustration, yielding higher subjective awareness and engagement compared to model-free policies (Patrício et al., 30 Apr 2025).

Psychologically parameterized personality and emotion models are central to higher-order personalization:

Big Five/CEA Personality Models:
- Robots encode personality as trait vectors [OCEAN], or reduced axes [CEA], which in turn modulate memory formation, attention allocation, language style, and emotional expression (Tang et al., 2 Feb 2025, Nardelli et al., 17 Apr 2025, Matcovich et al., 2024).
Emotion Generators and Appraisal Theory:
- LLM-driven appraisal functions update internal emotional state in response to perceived human cue and robot memory/context, then condition action selection on momentary emotional shifts and trait parameters (Tang et al., 2 Feb 2025, Nardelli et al., 17 Apr 2025).
Comfort/Engagement Variables:
- Adaptive controllers maintain internal “comfort” functions (rising with user stimuli, decaying otherwise), with growth/decay rates adapted online for individual users to regulate engagement/withdrawal cycles (Tanevska et al., 2020).
Affordance Personalization and Social Role:
- Systematic tuning of appearance, voice, and behavior (affordances) demonstrates that perceptions of warmth and competence, and preferred human-likeness, are highly context- and individual-dependent (Huang et al., 2023). One-size-fits-all strategies systematically underperform compared to multidimensional, closed-loop personalization mappings.

5. Applications and Evaluation Metrics

Personalization in social robots is validated across domains:

Socially Assistive Therapy and Rehabilitation: Personalized feedback systems outperform baseline wrappers and match therapist agreement, modulating exercise challenge and feedback to maximize patient engagement and performance (Lee et al., 2020, Clabaugh et al., 2019).
Education and Child Development: Hierarchical RL controllers adapt challenge and feedback levels to each child’s learning pattern in long-term home deployments, leading to measurable skill gains and sustained engagement (Clabaugh et al., 2019).
Service Robots: In bartender, barista, and museum-guide contexts (Rossi et al., 2022, Lim et al., 2023, Balossino et al., 8 Aug 2025), role-appropriate personality and history-aware interaction policies measurably increase preference, trust, loyalty, and perceived empathy.
Artistic Collaboration and Creative Expression: Personalized, metaphor-driven art making rooted in elicited emotion profiles enables more accurate and emotionally congruent visual communication in art therapy scenarios (Cooney, 2020).

Evaluation employs quantitative metrics (F1, engagement time, trust scales, art ranking accuracy) and qualitative assessments (user satisfaction, perceived lifelikeness, empathy, awareness), with ablation studies confirming the necessity of personality, memory, and explanation modules for coherent, engaging interaction (Tang et al., 2 Feb 2025, Nardelli et al., 17 Apr 2025, Patel et al., 2024).

6. Responsible Computing: Privacy, Fairness, Transparency

With increasing autonomy and memory comes responsibility:

Privacy: Federated learning, active unlearning, and parameter-only updates minimize risk of data exposure in multi-user deployments (Guerdan et al., 2022, Huang et al., 27 Jan 2026).
Fairness: Ranking modules in RS-based frameworks can enforce demographic parity, minimum exposure, and bias-corrected loss functions to ensure inclusive adaptation (Huang et al., 27 Jan 2026).
Transparency and Explainability: White-box, theory-of-mind dynamic models and concept bottleneck architectures yield explicit, queryable explanations for robot decisions and action selection, essential for trust and regulatory compliance (Patrício et al., 30 Apr 2025, Patel et al., 2024).
Ethical Design Guidelines: Personalization should incorporate explicit adaptation bounds, reflection on comfort/engagement levels, and scenario-tagging to avoid misaligned expectations or unintended affective consequences (Huang et al., 2023, Balossino et al., 8 Aug 2025).

7. Future Research Directions

Key research challenges persist:

Robust real-time affect and engagement estimation, including vision/audio fusion in noisy and short-window environments (Balossino et al., 8 Aug 2025).
Group personalization for multi-user social settings, combining individual profiles via social compatibility metrics.
Lifelong and non-stationary learning, to accommodate evolving user habits and preferences over months/years (Patel et al., 2024, Guerdan et al., 2022).
Integration with IoT and context sensors, enabling pervasive adaptation of speech, locomotion, and proxemics.
Richer user models—layered explicit, stereotype-based, and implicitly inferred traits—supporting cold-start scenarios and dynamically modulating trust, empathy, competence, and perceived friendliness.

Personalization in social robots remains a rapidly advancing intersection of engineering, psychology, and HRI, with the empirical consensus converging on modular, scenario- and profile-aware adaptation, robust interpretability, and responsible data stewardship as prerequisites for widespread acceptance and efficacy.