March 30, 2026
Researchers Dr Maria Dagioglou, Christos Spatharis and Dimitrios Koutrintzes from the Institute of Informatics & Telecommunications (IIT) of NCSR Demokritos participated in 21st ACM/IEEE International Conference on Human-Robot Interaction (HRI 2026), presenting a Late Breaking Report (LBR) on human–robot co-learning.
The work, titled “First Impressions Matter: Primacy Bias in Deep Reinforcement Learning during Human-Robot Co-learning” , investigates how early interactions shape collaboration between humans and reinforcement learning agents.
Experiments with novice users and a Soft Actor-Critic agent controlling a UR3 cobot show that initial interactions are critical: early negative experiences can hinder or even prevent learning, while positive ones facilitate effective collaboration. The study also highlights that agents may fail to adapt even as humans improve, leading to persistent miscoordination.
The findings emphasize the importance of high-quality early interaction data and point to limitations of current reinforcement learning approaches in handling human variability. Future work explores mitigation strategies such as transfer learning and improved replay buffer techniques.