Safe and Trustworthy Reinforcement Learning for Human–Robot Collaboration

Reinforcement Learning (RL) has rapidly emerged as a powerful approach for enabling robots to acquire adaptive, data-driven behaviors in real-world environments. In Human–Robot Interaction (HRI), RL offers a pathway toward robots that can collaborate effectively with people, adapt to user preferences, and operate under changing task and environmental conditions. However, HRI introduces distinct challenges: human behavior is dynamic and diverse, interaction is often partially observable, and many settings are safety-critical and socially constrained. As a result, conventional autonomy pipelines—and RL methods optimized primarily for performance—can struggle to deliver interactions that are safe, predictable, and intuitive, especially under uncertainty and distribution shift.In this Research Topic, HRI is understood primarily as embodied, physical interaction: robots acting in shared spaces with humans (e.g., assistive and healthcare robotics, manufacturing cobots, service robots, and mobile manipulators). We include communication and intent inference (language, gesture, gaze, and multimodal feedback) when these signals are used to shape robot decision-making and control. We explicitly prioritize work in which RL policies influence real robot actions and in which safety, trust, and accountability are treated as first-class objectives.Recent studies demonstrate RL’s potential to learn from human demonstrations and ongoing feedback, incorporate real-time cues, and integrate constraints during training and deployment. Yet, a major gap remains in producing robust and trustworthy robot behaviors that generalize across users and contexts while respecting safety requirements and ethical considerations (e.g., fairness across populations, transparency of learned strategies, and alignment with human norms). This Research Topic aims to catalyze interdisciplinary research across robotics, machine learning, and human factors, fostering the development of RL methods and evaluation practices that move beyond task reward optimization toward safe, human-compatible collaboration.We welcome Original Research, Review, and Methods articles addressing, but not limited to, the following topics:Safe and risk-aware RL for embodied HRI (constraint satisfaction, uncertainty-aware control, out-of-distribution robustness)Human-in-the-loop RL: preference learning, corrective feedback, interactive policy shapingCooperative and multi-agent RL for human–robot collaboration (shared autonomy, mixed-initiative control)Sim-to-real transfer and model-based RL for interactive tasks with humans in the loopSocially compliant RL for navigation and manipulation in shared spacesPersonalization and adaptation across users while preserving safety and equityExplainable and interpretable RL to support transparency, trust calibration, and debuggingFairness, bias mitigation, and ethical governance frameworks for embodied RL in HRIHierarchical, multi-task, and long-horizon RL for complex collaborative tasksImitation and demonstration learning for coordinated teamwork and safe bootstrappingApplication case studies in assistive, manufacturing, and healthcare robotics, including deployment lessonsReward and value design for HRI objectives (comfort, legibility, trust, workload, consent, and safety margins) Topic editors: Steven Fernandes, João Manuel R. S. Tavares, Asha C S Source: https://www.frontiersin.org/research-topics/75773/safe-and-trustworthy-reinforcement-learning-for-human-robot-collaboration