ETH Zürich’s Quadruped Robot Masters the Rubik’s Cube

The research team achieved this result by training the robot in virtual environments before transferring those skills to the physical world. Unlike robotic arms designed for fine motor control, this quadrupedal bot must balance its own weight while using its stubby limbs to reorient the cube and rotate individual sides. The process relies on a distillation technique where the robot learns from a simplified version of its own neural network, effectively teaching itself to handle the physical constraints of the cube.

This demonstration highlights the potential for quadruped robots to perform tasks far beyond simple navigation or inspection. By mastering the precise coordination required to solve a puzzle, the system proves that reinforcement learning can overcome the inherent limitations of a robot not originally built for manual dexterity.