Robot solves Megaminx in 7.647 seconds (WORLD RECORD)

[elijahmipt]

We have created a megaminx solving robot.

It rotates all 12 faces independently, in contrast to the previous robots. Average move number for a solution is just below 80. The algorithm used is called Beam Search (which is a variation of BFS with a heuristic like in A*, but here the heuristic is a trained neural network). Solution is calculated on a computer with powerful GPU in about 10 seconds.

This record is for the mechanical solution only, but these things can be relatively trivially done in parallel.This attempt was registered in Russian Book of Records (scroll down for English version) https://www.interrecord.ru/naimenshee...During the attempts WCA scrambles were used. Robot does not use any cameras, the state of the puzzle is kept in the memory, and the solving algorithm finds a path that is significantly shorter than the scrambling sequence.

 

[microchado]

wow

 

[iguana]

timofei's got some work to do

 

[academicmaslow]

We have created a megaminx solving robot.

It rotates all 12 faces independently, in contrast to the previous robots. Average move number for a solution is just below 80. The algorithm used is called Beam Search (which is a variation of BFS with a heuristic like in A*, but here the heuristic is a trained neural network). Solution is calculated on a computer with powerful GPU in about 10 seconds.

This record is for the mechanical solution only, but these things can be relatively trivially done in parallel.This attempt was registered in Russian Book of Records (scroll down for English version) https://www.interrecord.ru/naimenshee...During the attempts WCA scrambles were used. Robot does not use any cameras, the state of the puzzle is kept in the memory, and the solving algorithm finds a path that is significantly shorter than the scrambling sequence.

congrats!

 

[iguana]

elite kilominx now?

 

[Michael Wimberly]

When are we gonna get a 1x1 solving robot

 

[iguana]

When are we gonna get a 1x1 solving robot

never

 

[elijahmipt]

timofei's got some work to do

I've texted him about that btw
Really wanted to share

 

[elijahmipt]

elite kilominx now?

We are thinking about 4 by 4. It has less states, but it also cannot be solved optimally yet. And the mechanical difficulty is that it has no fixed centers, so some kind of regripping should be used. No wonder fastert robot is 3 times slower than Tymon

 

[Sub 20 Cuber]

Wow

 

[TimoChristianCuber]

wow. congrats! how long did it take to build?

 

[elijahmipt]

wow. congrats! how long did it take to build?

Thank you!

We began in September 2025 and finished in ~June 2026. There were 7 people involved, and all of them definitely not full-time, is was a side job for each of us.

Milestones:
- robot itself (this is the 3rd version, previous were working, but not as fast/smooth/convenient)
- a course on combinatorial group theory at MIPT (all the recordings are on the same youtube channel; in Russian though, but autotranslation might work, if anyone is interested)
- an algorithm that gives solutions with under 80 moves (this alg, developed by one of the team members, was top 1 on Kaggle for couple of month); a BSc thesis defended on the subject

There were 7 people involved:
- 2 working on the robot
- 2 teaching the course
- 1 working on the simulator
- 1 working on the algorithm
- 1 me helping everyone and organizing

An additional challenge was to register the record properly, which is bureaucratic in nature.
In future we are planning to build a 444-solving robot.

 

[Smiley01Sun]

I read pyraminx at first so my first thought was

"I solo that robot"

 

[iguana]

We are thinking about 4 by 4. It has less states, but it also cannot be solved optimally yet. And the mechanical difficulty is that it has no fixed centers, so some kind of regripping should be used. No wonder fastert robot is 3 times slower than Tymon

Why does it matter if it can't be solved optimally?

 

[elijahmipt]

Why does it matter if it can't be solved optimally?

Pardon, I honestly did not catch your question, can you please elaborate?

 

[iguana]

Pardon, I honestly did not catch your question, can you please elaborate?

You said that you couldn't make a 4x4-solving robot because it isn't known how to solve it optimally - why does it make a difference to the robot's ability to solve a 4x4 if it can't find the shortest solution?

 

[elijahmipt]

You said that you couldn't make a 4x4-solving robot because it isn't known how to solve it optimally - why does it make a difference to the robot's ability to solve a 4x4 if it can't find the shortest solution?

Probably I have formulated what I wanted to say not clear enough.

We can and we will make a 4x4-solving robot. But I have also mentioned that 4x4 (as well as megaminx) cannot be solved optimally yet. 3x3 has ~4*10^19 states, it can be solved optimally easily on modern computer. 4x4 has ~10^45, and there is no way to find an optimal solution __and__ prove that it is optimal, if we are talking about some random state.

Mega is even worse, it has ~10^68 states. Notoriously difficult to find short paths. And we don't know the graph diameter (shortest distance to the furthest state) for both these puzzles.

There is a project CayleyPy (Cayley graphs in Python), they work on path finding in big Cayley graphs. Nice recreational maths there.

 

[iguana]

Probably I have formulated what I wanted to say not clear enough.

We can and we will make a 4x4-solving robot. But I have also mentioned that 4x4 (as well as megaminx) cannot be solved optimally yet. 3x3 has ~4*10^19 states, it can be solved optimally easily on modern computer. 4x4 has ~10^45, and there is no way to find an optimal solution __and__ prove that it is optimal, if we are talking about some random state.

Mega is even worse, it has ~10^68 states. Notoriously difficult to find short paths. And we don't know the graph diameter (shortest distance to the furthest state) for both these puzzles.

There is a project CayleyPy (Cayley graphs in Python), they work on path finding in big Cayley graphs. Nice recreational maths there.

I understand, I thought you meant it had to be solved optimally for a robot to be able to solve it, but I get what you mean now.

 

[abunickabhi]

Insane yoooooo ❤

 

[abunickabhi]

I've texted him about that btw
Really wanted to share

What's his reaction to this beautiful piece of engineering?