ETH Zurich quadruped solves a Rubik's cube with its feet
Three full autonomous solves on real hardware, using depth images and joint sensing alone.
A four-legged robot has autonomously solved a giant Rubik's cube using the same two feet it stands on. Lukas Ziegler flagged the work, which comes from ETH Zurich and NVIDIA Robotics. The team calls it pedipulation, manipulation with feet.
Pin and push at once, on legs that are also holding the robot up.
A party trick on the surface. Underneath, it’s a quadruped completing a long, precise, multi-step physical task with nothing but its feet and its own eyes.
The pipeline is the interesting part. The robot walks around the cube and scans it with an onboard depth camera. SAM 3 segments the cube, ICP locks in its 3D pose, and the facelet colors are classified to rebuild the full cube state. A standard cube solver plans the move sequence. Learned skills then execute those moves.
Pin with one foot, push with the other
The hardest skill is turning a layer. One foot pins the cube in place while the other pushes the top layers 90 degrees. Both of those legs are also holding the robot up, so the balance problem and the manipulation problem are the same problem.
The policies were trained in simulation with reinforcement learning, then distilled into controllers that run on depth images plus the robot's own joint sensing. No motion capture, no external tracking, no extra cameras in the room.
The numbers
Pose estimation lands at roughly 2 cm position error and under 3 degrees of rotation error out to 2 m. Facelet color reading hit 100% accuracy indoors and 98.7% outdoors, which matters because a single misread sticker sends the solver down a wrong plan.
On real hardware the team reports three full autonomous solves, around 30 steps each, taking 8 to 9 minutes. That is a long horizon for a legged robot. Most quadruped manipulation demos are one grab or one push, and the failure mode that usually bites is drift over many contacts. Here the robot has to re-find the cube, re-grip it with feet that are not grippers, and keep the cube state consistent across dozens of moves.
Shown versus claimed
Three solves is three solves. The source does not give a success rate across attempts, so how often the robot fails is not stated. The outdoor figure applies to color reading, not to a full outdoor solve.
The cube itself is the party trick. The thing worth watching is a quadruped finishing a long, precise, multi step physical task with nothing but its feet and its own onboard sensing, which is a different capability from walking over rough ground. Ziegler calls it a real step toward legged robots that do more than walk.
Sound like you were at the demo
One email a day. Under three minutes. The demos, the deployments and the part the press release left out.