ETH Zurich researchers taught an ANYmal robot to scan, plan and solve a giant Rubik’s cube autonomously, using its feet to hold and turn the layers.
The system combines onboard depth cameras with a conventional cube solver and movement skills trained through reinforcement learning. One foot holds the cube steady while another turns its layers; other skills flip the cube and reposition the robot for the next move.
The team, including a researcher affiliated with NVIDIA, reports three complete autonomous solves on physical hardware. Each involved around 30 steps and took eight to nine minutes. The preprint demonstrates sustained, coordinated manipulation using a walking robot’s existing limbs; three runs do not establish a general success rate.