Solving Rubik's Cube with Bimanual Robot Hands

This work presents a simulation-to-learning pipeline for 2×2 Rubik's Cube manipulation using a humanoid robot. A high-fidelity simulation environment was constructed by integrating Isaac Lab with LeRobot, enabling VR-based teleoperation data collection using a Unitree G1 humanoid equipped with a multi-DoF Inspire hand. To address simulation inefficiencies, the cube asset was simplified to reduce collision overhead, and force transmission issues in the robotic hand were resolved through targeted parameter tuning. Approximately 120 teleoperated episodes were collected and used for imitation learning. Training was performed with GR00T N1.5, resulting in stable grasping and rotation behaviors. Quantitative evaluation demonstrated a peak success rate of 60% after stepwise training optimization, while qualitative analysis revealed emergent behaviors such as multi-rotation execution despite single-rotation training data. However, the learned policy showed limited generalization to background variations. This study highlights the importance of simulation fidelity, data consistency, and task scoping in complex dexterous manipulation, and outlines future directions including sim-to-real transfer and high-level planning integration.