2026-10-31

v13×3

Atomic Moves, Agentic Hands

Xunyi Zhao · ~2 min read · project page

Two Shadow hands on UR10e arms holding a 3×3 cube: the whole robot, the hands up close, and what the robot sees

Two robot hands pick up a scrambled Rubik's cube and solve it — the way a person does. No end-to-end policy, no 20-move optimal solver: learned atomic moves that always come home to the same grip, and an agent that plans with CFOP. Nine random scrambles, nine solves.

Nine random scrambles, each solved by Opus from the pattern and an algorithm library, and done by the hands, whole-cube rotations included. Two Shadow hands on UR10e arms, in simulation; turns at about 2×, the way back to the grip at about 4–5×.

v1 · nowThe 3×3 cube, solved end to end.

One grip to come home to

Every move starts here and ends here. That is what lets moves chain.

Atomic moves, learned with RL

Turn the layer, then return to the grip. Every face, both directions.

R
R'
U
U'
L
L'
D
D'
F
F'
B
B'

Chain them

Because every move comes home, any sequence just plays one move after another.

R U R' U' and back. Left: first person; right: wide view.

L U L' R: the left hand takes over for L.

Solve it

Cross, F2L, OLL, PLL — the agent plans, the hands execute.

One random scramble, start to finish.

Or let a language model solve it

Swap the planner for a language model. No optimal solver: like a speedcuber, it recognises the case at each CFOP stage — which F2L pair and where its pieces are, which OLL, which PLL — looks the algorithm up in a library, rotates the cube or turns U when that helps, and plays it. One answer per stage, and under five minutes of thinking for a whole solve on average (4.5 min; it took 14.6 when it worked every algorithm out in its head). Recognition is the model's own; the library only remembers algorithms. The hands don't notice.

Game 1: the hands executing the model's solution — case recognised, algorithm looked up — its thought for each stage on top; the corner says the playback speed.

Play with it

Three interactive demos on the project page.

How it works, in one picture

cube state agent CFOP planner · or an LLM moves executor one move at a time two Shadow hands on UR10e arms · sim atomic move library R R' U U' L L' D D' F F' B B' · x y z reference trajectory + RL residual every move ends in the standard grip after each CFOP stage, the real state goes back to the agent

(Swipe the figure sideways.) The agent never sees a finger; the skills never see the plan. The standard grip is the interface between them.

Next up

Whole-cube rotations are in: x, x', y, y', z and z' each end in the standard grip, and the agent uses them in the solves above. Bigger cubes are next.

x': the whole cube turns over in hand, and both hands land back in the grip.

The same hands holding a 7×7 cube, and first-person views of 4×4, 5×5, 6×6 and 7×7 cubes in the standard grip

Next: 4×4 to 7×7. The same hands already hold the bigger cubes in the standard grip.

References

[1] OpenAI, I. Akkaya et al., "Solving Rubik's Cube with a Robot Hand," 2019. arXiv:1910.07113.

[2] Amiet et al., quadruped solving a giant Rubik's cube with whole-body skills, 2026. arXiv:2610.24610.

[3] J. Fridrich, the CFOP method.

[4] The Yau method for big cubes (reduction to a 3×3).

← Back to blog  ·  Project page