This paper presents a novel approach to solving complex manipulation tasks, specifically the Rubik's Cube, using a humanoid robot hand. The key contributions include: 1. **Automatic Domain Randomization (ADR)**: A novel algorithm that automatically generates a distribution of randomized environments to improve the transfer of learned policies from simulation to the real world. This method enhances the robustness and generalization of the robot's control policies and vision state estimators. 2. **Robotic Platform**: A custom-built robot platform designed for machine learning, featuring the Shadow Dexterous Hand and a motion capture system for precise tracking of the robot's movements. 3. **Task Overview**: The paper addresses two manipulation tasks: block reorientation and solving a Rubik's Cube. The Rubik's Cube task is significantly more challenging due to its complexity and requires precise control and state estimation. 4. **Physical Setup**: Detailed descriptions of the robot platform, including improvements to the Shadow Dexterous Hand and the Gikker cube, a modified Rubik's Cube with built-in sensors for state estimation. 5. **Simulation Setup**: The simulation setup uses the MuJoCo physics engine and ORRB for rendering synthetic images, ensuring accurate modeling of the physical system. 6. **Policy Training**: The control policy is trained using Proximal Policy Optimization and reinforcement learning. The policy architecture includes a recurrent neural network with LSTM layers to enable meta-learning. 7. **State Estimation**: A vision-based state estimator is developed to estimate the pose and face angles of the Rubik's Cube. This system uses three RGB cameras and a neural network to predict the cube's state. 8. **Results**: The paper demonstrates that models trained in simulation can effectively solve the Rubik's Cube on a real robot, showcasing the effectiveness of ADR and the custom robot platform. 9. **Discussion**: The authors discuss the benefits of ADR, including simplified training and improved performance, and provide insights into the emergence of meta-learning during test time. The paper highlights the potential of combining machine learning with robotics to tackle complex manipulation tasks, opening new avenues for future research and applications.This paper presents a novel approach to solving complex manipulation tasks, specifically the Rubik's Cube, using a humanoid robot hand. The key contributions include: 1. **Automatic Domain Randomization (ADR)**: A novel algorithm that automatically generates a distribution of randomized environments to improve the transfer of learned policies from simulation to the real world. This method enhances the robustness and generalization of the robot's control policies and vision state estimators. 2. **Robotic Platform**: A custom-built robot platform designed for machine learning, featuring the Shadow Dexterous Hand and a motion capture system for precise tracking of the robot's movements. 3. **Task Overview**: The paper addresses two manipulation tasks: block reorientation and solving a Rubik's Cube. The Rubik's Cube task is significantly more challenging due to its complexity and requires precise control and state estimation. 4. **Physical Setup**: Detailed descriptions of the robot platform, including improvements to the Shadow Dexterous Hand and the Gikker cube, a modified Rubik's Cube with built-in sensors for state estimation. 5. **Simulation Setup**: The simulation setup uses the MuJoCo physics engine and ORRB for rendering synthetic images, ensuring accurate modeling of the physical system. 6. **Policy Training**: The control policy is trained using Proximal Policy Optimization and reinforcement learning. The policy architecture includes a recurrent neural network with LSTM layers to enable meta-learning. 7. **State Estimation**: A vision-based state estimator is developed to estimate the pose and face angles of the Rubik's Cube. This system uses three RGB cameras and a neural network to predict the cube's state. 8. **Results**: The paper demonstrates that models trained in simulation can effectively solve the Rubik's Cube on a real robot, showcasing the effectiveness of ADR and the custom robot platform. 9. **Discussion**: The authors discuss the benefits of ADR, including simplified training and improved performance, and provide insights into the emergence of meta-learning during test time. The paper highlights the potential of combining machine learning with robotics to tackle complex manipulation tasks, opening new avenues for future research and applications.