"CQ(\u03bb) Human-in-the-Loop Reinforcement Learning Dataset for Robotic Bag-Shaking Control"

"This dataset supports comparative research on autonomous and human-guided reinforcement learning for robotic manipulation. It extends prior work on human\u2013robot collaborative learning (Kartoun, Stern, & Edan, IEEE SMC, 2006; Journal of Intelligent & Robotic Systems, 2010) by providing structured, reproducible experimental data for modern policy evaluation and adaptive control analysis.The dataset accompanies a simulation study of CQ(\u03bb) \u2014 Cooperative Q-learning with eligibility traces \u2014 applied to a bag-shaking task in which a robotic agent must extract knotted objects under partially observable conditions (hidden knot tightness and bag entanglement). It comprises 10,000 episodes across 10 independent runs and two learning conditions: standard Q(\u03bb) and CQ(\u03bb), in which a human expert provides linguistic corrective guidance (\"significantly increase\", \"slightly decrease\", etc.) whenever rolling task performance drops below a defined threshold. Approximately 500,000 step-level state\u2013action transitions and 1,200 intervention records are included, alongside episode-level reward trajectories, success flags, exploration schedules, and timing data.CQ(\u03bb) achieves roughly 30% higher accumulated reward, a 21 percentage-point improvement in success rate, and 14% faster episode completion relative to autonomous Q(\u03bb), demonstrating the practical value of performance-triggered linguistic policy shaping. The data are intended to support research in human-in-the-loop reinforcement learning, sample efficiency analysis, intervention strategy design, and benchmarking of policy shaping methods within collaborative robotic frameworks."