A Scoping Review and Experimental Study on Reinforcement Learning from Human Feedback for Human-Robot Collaboration

Human feedback quality matters for safety: The research found that when humans actively initiate feedback (rather than the system prompting them), the feedback better captures what people actually perceive as safe, suggesting that passive feedback collection may miss important safety concerns in human-robot collaboration.

Multiple ways to teach robots are emerging: The review identified diverse feedback methods—including different formats and timing—that can be integrated at various stages of robot training, indicating that there's no one-size-fits-all approach to teaching robots to work safely alongside humans.

Safety requires two-way learning: The study emphasizes "bidirectional, closed-loop design," meaning robots must not only learn from human feedback but also adapt their behavior in ways that help humans provide better feedback—a more complex but more realistic model than one-directional instruction.

Testing with real humans is essential: Pilot experiments using both human metrics (like psychological safety measured through questionnaires) and robot metrics (like collision avoidance measurements) were standard practice across reviewed studies, highlighting that responsible AI development requires evaluating how humans actually experience working with robots, not just technical performance.

The field is still developing best practices: With only 20 peer-reviewed publications on this specific topic identified between 2020-2025, the research reveals that practical challenges around safety during development and human feedback quality remain largely unsolved, indicating this is an emerging area needing further standardization.