Researchers have demonstrated a quadruped robot that can autonomously solve a giant Rubik's cube using only its feet. The robot completed three full hardware solves, each involving about 30 steps and taking 8–9 minutes.

The system uses an onboard RGB-D camera to estimate the cube's position and color state from partial views as the robot moves around it. A planner converts the cube solution into executable skill sequences. The robot learned two types of whole-body skills through reinforcement learning: reorienting the cube without disturbing its layer configuration, and turning a layer by pinning the cube with one foot while pushing with the other.

The skills were trained with privileged observations and then distilled into policies that rely only on depth and proprioception. The deployed policies tolerate variations in robot-cube placement and recover from external disturbances, which was critical because unintended layer motion can invalidate subsequent operations.

The work shows that combining onboard perception with coordinated pedipulation enables a quadruped to complete extended multi-step physical tasks that require repeated precise interactions. It points toward legged robots manipulating objects in settings where arms are not available or practical.