In this work, we propose a novel Personalised Large Language Model (PLLM) agent, designed to advance the integration and adaptation of large language models within the field of human–robot interaction and human–computer interaction. While research in this field has primarily focused on the technical deployment of LLMs, critical academic challenges persist regarding their ability to adapt dynamically to user-specific contexts and evolving environments. To address this fundamental gap, we present a methodology for personalising LLMs using domain-specific data and tests using the NeuroSense EEG dataset. By enabling the personalised data interpretation, our approach promotes conventional implementation strategies, contributing to ongoing research on AI adaptability and user-centric application. Furthermore, this study engages with the broader ethical dimensions of PLLM, critically discussing issues of generalisability and data privacy concerns in AI research. Our findings demonstrate the usability of using the PLLM in a human–robot interaction scenario in real-world settings, highlighting its applicability across diverse domains, including healthcare, education, and assistive technologies. We believe the proposed system represents a significant step towards AI adaptability and personalisation, offering substantial benefits across a range of fields.
A safety margin derived from a vision–language verdict is only as valid as that verdict is recent, yet run-time safety filters model state noise, not the staleness of an intermittent label. We present AEGIS, a safe-action projection whose hold radius is an explicit function of a semantic belief and of the age of the verdict that produced it, with a conditional grasp-site separation result and a fresh-evidence feasibility property under stated assumptions. AEGIS runs on one fixed Intel RealSense D435i: localisation through a calibrated homography with a sub-millimetre mean fitting residual, a workspace-gated intruder detector fused with metric depth, and an on-device verdict returned in a fraction of a second. Beneath it a geometric layer checks and, when necessary, modifies commands against a set of geometric constraints, so it wraps any policy without retraining, reducing accumulated constraint cost in simulation without a significant loss of task success. In a buffer-matched comparison, indexing the margin by verdict age is significantly safer than trusting the latest verdict, raising the fraction of episodes free of any commanded approach to a person, and no constant belief reproduces the behaviour. On the physical Niryo NED3 Pro the logged radius expands as a verdict ages and collapses on a fresh one: AEGIS held positive measured clearance, one trust-latest run ended up inside the detected footprint, and the worst-case policy delivered least. Claims are reported against explicitly separated evidence tiers.
Adaptive attention allocation in dynamic social environments remains a fundamental challenge for autonomous robots, requiring the integration of perceptual saliency, social context, and real-time decision-making.We present a bio-inspired reinforcement learning framework for robotic gaze control that incorporates a habituation mechanism to regulate the exploration–exploitation trade-off, mirroring how biological attention systems filter redundant stimuli whilst remaining responsive to novel events. Through a comprehensive ablation study comparing Deep Q-Learning (DQL), Vanilla Q-Learning (VQL), and Multi-Objective Q-Learning (MOL), we uncover a critical insight: habituation significantly enhances DQL performance, improving response efficiency and policy stability, yet causes systematic degradation in MOL due to fundamental incompatibilities between fixed-threshold resets and the extended episodes required for multi-objective optimisation. This differential effect reveals that bio-inspired mechanisms cannot be applied universally across learning architectures but must be carefully matched to algorithmic characteristics. Real-world deployment on the ARI humanoid robot validates the framework’s practical applicability, achieving robust gaze prediction accuracy across diverse interaction scenarios with well-calibrated confidence metrics that reliably distinguish correct from incorrect predictions. Our findings provide evidence-based guidelines for integrating biological principles into cognitive robotics, demonstrating both the promise and the pitfalls of bio-inspired mechanism design.