Twisting a Rubik's Cube normally calls for fingers, and a robot doing it without any human at all is even rarer. That is the problem this project takes on: a quadruped manipulates a giant 3x3 cube entirely with its feet, and the whole solve runs with zero human intervention. Credit goes to Robotic Systems Lab: Legged Robotics at ETH Zürich, working with NVIDIA on the demonstration.
The platform is an ANYmal with a red chassis, four legs ending in spherical rubber feet, and a compute and sensor enclosure on top. It carries onboard RGB and depth cameras. The motion skills were trained in Isaac Lab, using reinforcement learning followed by teacher-student distillation, and the deployed skills rely only on onboard depth sensing and proprioception.
The autonomous pipeline runs in three phases: scanning every face, choosing skills, then physically solving. RGB-D perception estimates the cube's pose and color state even from partial views, and all six faces get scanned to capture the full state. Five learned skills do the work: flips forward and backward, clockwise and counter-clockwise rotations, plus a locomotion-based repositioning skill. To turn a single layer, the robot pins the cube with one foot while pushing with another, keeping the layer configuration intact as it reorients the cube. The full fifteen-step sequence ends with a solved cube, and the perception side includes segmentation, 3D pose estimation, face color identification and an unfolded cube net.
Robustness looks like the standout detail here. The skills tolerate placement variations and recover from ropes pulling legs, blocks in the way and sticks prodding the cube, and tests covered misaligned layers and a low-friction surface, plus outdoor work on cobblestones. The same skills also moved a trash can and a metal equipment case, which raises the open question of how broadly this foot-based manipulation can generalize.
Project by Robotic Systems Lab: Legged Robotics at ETH Zürich. Watch it on their YouTube channel: