This paper presents a learning-based approach for continuously spinning pen-like objects. The authors demonstrate the ability to spin a variety of pen-like objects with different physical properties using only proprioception as feedback. The key challenge is the significant gap between simulation and the real world, which makes it difficult to directly apply simulated policies to real-world tasks. To address this, the authors first train an oracle policy in simulation using reinforcement learning to generate high-fidelity trajectory data. This data is then used to pre-train a sensorimotor policy in simulation. The policy is then fine-tuned using real-world trajectories to adapt to real-world dynamics. The authors show that with less than 50 real-world trajectories, their policy can learn to rotate more than ten pen-like objects for multiple revolutions. The authors highlight the importance of proper initial state design and the use of privileged information during policy training. They also emphasize the role of the z-reward in keeping the pen horizontal during rotation, which is crucial for stable performance in the real world. The authors compare their approach to several baselines and show that their method achieves better performance in both simulation and the real world. They also demonstrate that their method can generalize to unseen objects, which is a significant achievement in the field of dexterous in-hand manipulation. The authors' approach involves three main steps: training an oracle policy in simulation, pre-training a sensorimotor policy in simulation, and fine-tuning the sensorimotor policy using real-world trajectories. The oracle policy is trained to generate realistic trajectories that can be used for pre-training and as an open-loop controller in the real world. The sensorimotor policy is then fine-tuned using real-world trajectories to adapt to real-world dynamics. The authors show that their method can achieve continuous spinning of pen-like objects in the real world, which is a significant advancement in the field of dexterous manipulation.This paper presents a learning-based approach for continuously spinning pen-like objects. The authors demonstrate the ability to spin a variety of pen-like objects with different physical properties using only proprioception as feedback. The key challenge is the significant gap between simulation and the real world, which makes it difficult to directly apply simulated policies to real-world tasks. To address this, the authors first train an oracle policy in simulation using reinforcement learning to generate high-fidelity trajectory data. This data is then used to pre-train a sensorimotor policy in simulation. The policy is then fine-tuned using real-world trajectories to adapt to real-world dynamics. The authors show that with less than 50 real-world trajectories, their policy can learn to rotate more than ten pen-like objects for multiple revolutions. The authors highlight the importance of proper initial state design and the use of privileged information during policy training. They also emphasize the role of the z-reward in keeping the pen horizontal during rotation, which is crucial for stable performance in the real world. The authors compare their approach to several baselines and show that their method achieves better performance in both simulation and the real world. They also demonstrate that their method can generalize to unseen objects, which is a significant achievement in the field of dexterous in-hand manipulation. The authors' approach involves three main steps: training an oracle policy in simulation, pre-training a sensorimotor policy in simulation, and fine-tuning the sensorimotor policy using real-world trajectories. The oracle policy is trained to generate realistic trajectories that can be used for pre-training and as an open-loop controller in the real world. The sensorimotor policy is then fine-tuned using real-world trajectories to adapt to real-world dynamics. The authors show that their method can achieve continuous spinning of pen-like objects in the real world, which is a significant advancement in the field of dexterous manipulation.