A robot Rubik’s Cube demonstration has delivered a wonderfully strange sight: a four-legged machine solving a large puzzle without using hands. The robot moves around the cube, inspects its coloured faces, holds it in position with a foot and uses another foot to turn the layers. In completed physical trials, it did this repeatedly until the puzzle was solved.
The research comes from ETH Zurich and collaborators including NVIDIA. The team reports three complete autonomous solves, each requiring roughly 30 steps and taking eight to nine minutes. That is enough to establish a serious laboratory achievement, but not enough to say the robot would reliably succeed under every conceivable condition.
Why a robot Rubik’s Cube challenge is difficult
A human solving a cube generally keeps it in their hands. They can feel how much pressure to apply, adjust a grip almost without thinking and rotate a face while keeping the rest of the puzzle steady. A quadruped designed for walking has a very different body. Its feet are built to support weight and move across a floor, not to manipulate a free-standing puzzle.
The researchers call this approach pedipulation, meaning the use of legs or feet to handle objects. In the team's published demonstration, the robot must balance its own body while controlling the cube’s position and turning individual layers without accidentally dislodging others. A mistaken movement is not simply untidy. It can invalidate the series of moves calculated to solve the puzzle.
That makes the experiment interesting beyond the novelty of a robot dog playing with a toy. The machine has to combine seeing, planning, balance and delicate physical contact over a lengthy sequence. These skills may be relevant to future robots that have to interact with tools or objects in an environment not arranged specifically for their convenience.
First the robot has to work out the colours
Before making a move, the machine needs to understand the cube’s current state. A conventional puzzle solver can calculate a solution once it knows the colour arrangement, but the robot cannot see every face at once. It has to inspect the object from different viewpoints and combine partial observations into a complete representation.
The researchers describe an onboard vision system using colour images and depth information. The robot moves and repositions the cube to inspect the hidden faces. The resulting observations are used to estimate both the puzzle’s arrangement and its position in physical space. The solver then provides a target sequence, which must be translated into movements the robot can actually execute.
This is a useful distinction. The core puzzle-solving calculation is not the dramatic physical achievement. Computers have long been able to solve the logical problem. The challenge is carrying out those instructions with a moving machine that must balance while applying forces to a real object.
What the feet actually do
The research preprint dated October 2026 describes learned skills for repositioning the cube, flipping it and rotating its layers. During a turn, one foot can pin the object while another pushes a layer. At other times the robot changes its own stance or turns the cube so the next operation becomes accessible.
There is a constant negotiation between stability and dexterity. The robot needs enough force to turn a mechanical layer, but too much or badly directed force could move the whole cube or throw it out of alignment. A human can sense these problems through fingertips. The robot instead relies on its sensors, learned movement policies and feedback from its body.
The team trained movement skills using reinforcement learning, in which a system practises actions and receives feedback about their outcomes. Much of that preparation can happen in simulation before policies are transferred to the physical robot. The technique does not imply the machine has become a human-like puzzle enthusiast. It means particular motor skills were learned and coordinated to serve a defined task.
The three successful runs matter, and so do their limits
The paper reports three complete autonomous hardware solves with roughly 30 steps each, lasting eight to nine minutes. Those are concrete results. They are stronger evidence than a mere claim that the system could potentially solve the puzzle, because the task was executed on a real robot.
There is also a limit to what can be inferred. Three successes do not provide a meaningful failure rate across thousands of cubes, lighting conditions, damaged puzzles or unexpected human interruptions. The researchers describe robustness to certain disturbances and variations in placement, but that should not be turned into a promise of flawless real-world performance.
A demonstration can be both impressive and narrow. The experiment tests a coordinated sequence of manipulation skills with a particular object. It does not show that the robot can walk into an unfamiliar home, find a toy in a cupboard and solve it without any preparation or supervision.
Why scientists want walking robots to handle objects
Walking machines may eventually need to move things when they have no convenient arm or gripper available. A robot working in a cramped industrial location could use a leg to steady a panel. A field machine might need to move an obstacle or operate a control while keeping its balance. Those are possible applications of leg-based manipulation, not capabilities established for this particular robot by the cube experiment.
Robotics development often divides into locomotion, the ability to get somewhere, and manipulation, the ability to change the environment once there. The ETH work is interesting because it makes the same limbs responsible for both. The robot moves to acquire a better view, changes its stance and uses its feet for interactions that might otherwise require a separate tool.
That broadening of the role of a leg is different from research into robots that use specialised walking hands. The two approaches illustrate the variety of bodies engineers are exploring. A foot is not automatically the best gripper, and adding a hand creates its own engineering compromises. Successful machines will depend on the requirements of the work rather than the entertainment value of their shape.
Every successful movement depends on a good picture
The vision problem is inseparable from the physical one. If the robot misreads a coloured face, the calculated move sequence may be wrong. If it estimates the cube’s position poorly, the correct move may be executed in the wrong place. If a layer shifts unexpectedly, the earlier plan may need reconsideration.
This relationship between perception and action helps explain why apparently easy everyday tasks remain hard for robots. A person moving a chair can continually revise their grip and direction as the furniture shifts. A robot needs an adequate way to observe and respond to those changes. A flawless plan based on stale information is not enough.
It also connects to work on keeping humanoid robots balanced during interruptions. Balance is not an optional extra while a machine handles an object. It is part of the same physical problem, particularly when a foot that ordinarily supports weight is busy doing something else.
A Rubik’s Cube is a test, not the destination
Why use a giant puzzle rather than a practical factory part? A Rubik’s Cube has a clear finish line. It requires a sequence of precisely controlled actions, and success is immediately visible. Those qualities make it a useful research challenge. Observers can see whether the robot completed a demanding task, rather than relying only on an abstract benchmark score.
The size of the puzzle also makes the interaction visible and physically meaningful. The robot must use its existing limbs to hold and rotate a large object. That helps distinguish the result from an ordinary computer displaying a solved cube on a screen.
There are reasons not to overinterpret the spectacle. The skills are tailored to the challenge, the environment remains constrained and the paper reports a small set of demonstrations. No evidence here establishes consumer availability, price or the ability to perform domestic chores at scale.
What the robot actually teaches us
A quadruped completing a giant Rubik’s Cube is memorable because the behaviour appears so unlikely. The technical lesson is less theatrical: useful physical intelligence comes from combining perception, planning and motor control in a way that survives contact with a real object.
Future research will need to test how those abilities transfer to unfamiliar tasks and environments. Systems will also need to recognise when they are uncertain and know when to stop, especially around people. The related work on robot spatial memory addresses another part of that challenge: keeping track of the world while it changes.
For now, the achievement is specific and worth celebrating on its own terms. The robot did solve the puzzle, on real hardware, using its feet. It has not solved the far larger puzzle of making machines reliably useful everywhere else.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.