This paper presents a method for solving a Rubik's cube using a robotic hand trained with reinforcement learning and automatic domain randomization (ADR). The key components of the approach are ADR, which automatically generates a distribution over randomized environments of increasing difficulty, and a custom robot platform designed for machine learning. The ADR algorithm enables the model to learn to solve complex manipulation tasks in simulation and transfer this knowledge to the real world. The system uses a five-fingered humanoid hand trained with ADR to solve a Rubik's cube, which involves both control and state estimation problems. The control policy is trained using a recurrent neural network and reinforcement learning, while the vision-based state estimator uses a convolutional neural network to predict the pose and face angles of the cube. The system also uses a custom cube with embedded sensors (Giiker cube) to provide face angle information. The results show that the model can solve a Rubik's cube with high accuracy, demonstrating the effectiveness of ADR in enabling sim-to-real transfer. The paper also discusses the challenges of solving a Rubik's cube, including the need for high precision and accurate state estimation, and presents a detailed analysis of the policy's performance and the factors that contribute to its success.This paper presents a method for solving a Rubik's cube using a robotic hand trained with reinforcement learning and automatic domain randomization (ADR). The key components of the approach are ADR, which automatically generates a distribution over randomized environments of increasing difficulty, and a custom robot platform designed for machine learning. The ADR algorithm enables the model to learn to solve complex manipulation tasks in simulation and transfer this knowledge to the real world. The system uses a five-fingered humanoid hand trained with ADR to solve a Rubik's cube, which involves both control and state estimation problems. The control policy is trained using a recurrent neural network and reinforcement learning, while the vision-based state estimator uses a convolutional neural network to predict the pose and face angles of the cube. The system also uses a custom cube with embedded sensors (Giiker cube) to provide face angle information. The results show that the model can solve a Rubik's cube with high accuracy, demonstrating the effectiveness of ADR in enabling sim-to-real transfer. The paper also discusses the challenges of solving a Rubik's cube, including the need for high precision and accurate state estimation, and presents a detailed analysis of the policy's performance and the factors that contribute to its success.