For those interested, the full source-code is available here: [ Ссылка ] (robot), [ Ссылка ] (solving algorithm).
This succeeds my own mirrcub3r ([ Ссылка ]; also featured in the 2019 Royal Institution Christmas Lectures [ Ссылка ]) as the new fastest Lego-based Rubik's Cube solving robot in the world. Most, obviously it can now turn all 6 sides making solutions ~12% shorter on average. This is however by far not the only improvement that has been made. First, the now exclusively Technic-based construction is a lot more stable and centers the cube much better allowing even more aggressive turning while at the same time maintaining noticeably better consistency. Further, the solving algorithm has been carefully revised, making it even more efficient and implementing some extra features. One of those is the ability to return more than a single solution thus allowing post-selection based on historic timing data (collected during previous solves) to also incorporate information not directly considered by the solving algorithm into the process of finding the solution that will take the shortest time to execute. Perhaps the biggest improvement of SquidCuber is the color recognition algorithm (although it is not very relevant in terms of speed). Being a combination of machine learning and full constraint propagation it is able to reliably handle the extremely difficult scanning conditions of this robot (most prominently, the strong reflections caused by the steep viewing angles and unfavorable surface texture of the cube), which completely broke all my previous approaches.
Overall, this machine pushes the Lego hardware to its absolute limits and includes basically every optimization, no matter how small, I could come up with. For now, I believe this is about as good as it is going to get. However, I think it is already pretty crazy that consistent 1 second times are possible with Lego at all.
