This paper presents a novel learning-based approach to spin pen-like objects, a challenging task in in-hand manipulation. The authors address the limitations of current methods, which struggle with high-quality demonstrations and the gap between simulation and real-world performance. They train an oracle policy using reinforcement learning in simulation, generating a high-fidelity trajectory dataset. This dataset is then used to pre-train a sensorimotor policy in simulation and fine-tune it using real-world trajectories. The method successfully learns to rotate over ten pen-like objects with different physical properties for multiple revolutions, demonstrating the effectiveness of combining simulation and real-world data. The paper includes a comprehensive analysis of design choices and shares lessons learned during development, highlighting the importance of proper initial state design, privileged information, and the role of simulation in bridging the sim-to-real gap.This paper presents a novel learning-based approach to spin pen-like objects, a challenging task in in-hand manipulation. The authors address the limitations of current methods, which struggle with high-quality demonstrations and the gap between simulation and real-world performance. They train an oracle policy using reinforcement learning in simulation, generating a high-fidelity trajectory dataset. This dataset is then used to pre-train a sensorimotor policy in simulation and fine-tune it using real-world trajectories. The method successfully learns to rotate over ten pen-like objects with different physical properties for multiple revolutions, demonstrating the effectiveness of combining simulation and real-world data. The paper includes a comprehensive analysis of design choices and shares lessons learned during development, highlighting the importance of proper initial state design, privileged information, and the role of simulation in bridging the sim-to-real gap.