Keaghan Paterson · Mechanical design

rough formatting, to update post completion

Started April 2026. Work in progress.
Done: scanner, solver, all parts printed. In sim the hand reorients the cube and turns faces on command.
Next: servos arrive → assemble → sim-to-real → juggle.

Robot

A full solve in simulation, half speed

Physical build

Open-Source DexHand V2.0

Photograph the cube → read the state → solve it. Works on different cubes and in different lighting.

Scan → import → solve (screen recording)

Teaching a simulated hand to turn the cube, against physics measured off the real one.

Solving one face, half speed

Turning one face is the hard half of a robot solve. The grip has to hold the cube still and leave one face free at the same time, and breaking a face out of its magnets takes more than twice the torque of keeping it turning.
Measured that torque on the real cube with golf balls, built it into the simulation, then trained the hand against it: rewards for turning toward the target angle, penalties for drops, for drifting out of the grip, and for stalling.
Reached 3.5 solved moves per 10-second attempt at 76% success, with no assisting torque. Time per move fell from 3.8 seconds to 2.0.

Rolling the cube to a commanded pose, half speed

A solver names a face to turn, so the hand first has to present that face to the fingers. Every reorientation risks dropping the cube off an open palm.
Adapted NVIDIA's cube-reorientation rewards to this 18-servo hand, then added a penalty for resting the cube on the palm. That forced the fingers to do the rolling instead of letting the hand lean on gravity.
Reached about 11 commanded poses per 10-second attempt at 90% success. These weights seeded the face-turning work above, so it did not start from scratch.

Every gain here came from reward design. The cube's measured friction and magnet strength stayed fixed throughout, so the policy had to improve rather than the task getting easier.
Two failures came out of watching playback: the hand freezing in a crossed-finger grip to farm holding points, and fast flicks that swept past the target between control steps and scored nothing. Added a time cost, a fingertip clearance penalty, and crossing detection.
The curve resumed climbing after each fix and has not flattened yet: 0.4 to 3.5 solved moves per attempt over 21 hours.

Six-face cube: carry the one-face skills to all six faces, then chain real solver moves.
Fingertip force sensing: custom PCBA, schematic and layout in progress.
Sim → real: move the trained policies onto the real hand.
Two hands, three cubes: second hand, then juggling.