The journey
The road so far, and the detours.
Everything we have built and learned, one line per track. The grey crosses are ideas we tried and dropped, and we wrote down why, so you don't have to repeat them.
- 9days so far
- 32steps done
- 6ideas dropped
- 5tracks at once
Map of the journey
Coupling with a rigid bar
A small bar per finger keeps the fingertip within 0.6 degrees of the target.
- Done
- Tried and dropped
- Latest step on a track
Day by day
What happened, chapter by chapter
- Day 126 September 2026
The start
A public repo, the first firmware and the first simulated hand.
V0 thumb and index printed
Hardware. The first prototype: 8 joints, tendon-driven, 3D printed.
Firmware v0.1 for 8 steppers
Firmware and electronics. Smooth speed-up and slow-down, motors switch off when idle, text protocol over USB.
First MuJoCo model
Simulation and software. Converted from the CAD export, checked against the real hand.
Public repo and licenses
Website and open source. Code Apache-2.0, hardware CERN-OHL-S-2.0, docs CC BY 4.0.
- Day 227 September 2026
Becoming a project
A website, a Python library and a smarter way to route the tendons.
Tendons cross each joint on its axis
Hardware. New routing for the index knuckle and middle joint, designed in CAD (not printed yet).
Python library and digital twin
Simulation and software. One Hand interface for the simulation and the real hand.
A dark, robotic lookDropped
Website and open source. The first website direction.
Design system: clean, open, friendly
Website and open source. Light by default, one blue accent.
Website v1
Website and open source. 26 pages and a 3D hand.
- Day 328 September 2026
Thinking big
The full five-finger hand is designed, the AI plan is written, and a webcam moves the hand.
21 joints, 21 servosDropped
Hardware. The first V1 plan: one servo for every joint.
V1: five fingers, servos in the forearm
Hardware. The full hand designed in CAD, with human finger proportions.
A 5-joint thumbDropped
Hardware. We added a fifth thumb joint, like a human thumb, and removed it a day later.
Steppers as the final motorDropped
Firmware and electronics. Cheap stepper motors drove the first prototype. We decided V1 would not use them.
Smart-servo driver
Firmware and electronics. Feetech SCS0009 on one shared bus, with position feedback.
Webcam control of the hand
Simulation and software. Your own hand moves the simulated one.
V1 model with 20 joints
Simulation and software. Every tendon is simulated.
AI research hub and the big goal
AI. A robot that does what humans do, with the hand as the focus.
The real hand model on the site
Website and open source. Built from the CAD file.
- Day 429 September 2026
Learning from the real hand
The first test on real hardware, a fix for the knuckle and the first grasp learning.
Back to a 4-joint thumb
Hardware. A simpler thumb, with the tendon routing reworked.
5 mm spool for the index knuckle
Hardware. The knuckle bent easily but extended weakly. A smaller spool pulls twice as hard.
First V0 test rig
Firmware and electronics. The index finger wired to an ESP32-S3.
Grasp learning system
AI. Reinforcement learning in simulation, with a human-like reward.
Reward hacking: hovering beat winningDropped
AI. The first reward ended the episode on success.
A reward that pays for holding
AI. Success no longer ends the episode, so holding the object is what counts.
One place for photos and video
Website and open source. Without GPS data in the pictures.
- Day 530 September 2026
Looking around
We studied the best robot hands in the world to see where we stand.
Study of the best robot hands
AI. The biggest gap: fingertip force.
- Day 61 October 2026
Fewer servos
Sixteen servos instead of twenty-one, by copying how a human finger works.
16 servos for 20 joints
Hardware. Each finger's last joint follows the middle joint, so it needs no servo.
Coupling the fingertip with a tendonDropped
Hardware. The first idea for making the fingertip joint follow the middle joint: a second tendon.
Firmware 0.4.0 for 16 servos
Firmware and electronics. Torque off at boot, nothing moves before a command.
- Day 72 October 2026
Arms and a cloud GPU
A better fingertip linkage, two simulated arms and the first big training run.
Coupling with a rigid bar
Hardware. A small bar per finger keeps the fingertip within 0.6 degrees of the target.
36 tendons and the finger linkage
Simulation and software. The simulation matches the 16-servo design.
A simulated arm
Simulation and software. OpenArm's shoulder and elbow stand in until we build our own.
Two arms, mirrored left hand
Simulation and software. The first body: two arms on a pole.
77% grasp success in 49 minutes
AI. A cloud GPU, simulating many hands at once. It used the older hand design, so we will repeat it.
Training the hand on the arm
AI. The arm brings the hand to the object (about 87%).
- Day 83 October 2026
The journey page
The linkage is modelled in CAD and this page exists.
The finger linkage modelled in CAD
Hardware. Bars and plates for all four fingers, with no collisions.
This journey page
Website and open source. You are here.
- Day 95 October 2026
Ready to go live
The website gets safe, fast and easy to find, ready for its first deploy.
Website ready to launch
Website and open source. Placeholders gone, security headers on, and findable by search engines and AI assistants.
Dropped ideas
What we tried, why it failed, and what we learned
Mistakes are part of an open project. Here is each one, so you can skip it.
21 joints, 21 servos
The first V1 plan: one servo for every joint.
Why it did not work. Every servo draws current and takes space in the forearm. Twenty-one of them needed a huge power supply and left no room for the tendon channels.
Lesson. Count the power and the space before you count the joints.
Instead: 16 servos for 20 joints
A 5-joint thumb
We added a fifth thumb joint, like a human thumb, and removed it a day later.
Why it did not work. The extra joint needed another hinge and more tendon channels in the most crowded part of the hand. It was not worth the complexity for a first full hand.
Lesson. Copy the human hand where it matters, not everywhere at once.
Instead: Back to a 4-joint thumb
Coupling the fingertip with a tendon
The first idea for making the fingertip joint follow the middle joint: a second tendon.
Why it did not work. A tendon has to be tensioned, stretches over time and goes slack. The fingertip joint has no servo of its own to pull the slack out.
Lesson. When nothing can take up slack, use something stiff.
Instead: Coupling with a rigid bar
Steppers as the final motor
Cheap stepper motors drove the first prototype. We decided V1 would not use them.
Why it did not work. A stepper has no idea where it is, so someone must straighten the hand by hand before every start. Eight of them also use up nearly every pin on the controller.
Lesson. Position feedback is worth paying for.
Instead: Smart-servo driver
Reward hacking: hovering beat winning
The first reward ended the episode on success.
Why it did not work. The robot learned that hovering just below the success height scored more than finishing, because finishing ended the game. A learner finds every loophole in its reward.
Lesson. Read what your reward really pays for, not what you meant it to pay for.
Instead: A reward that pays for holding
A dark, robotic look
The first website direction.
Why it did not work. It felt cold and nerdy. The project wants to be friendly to anyone who is curious, not only to engineers.
Lesson. Design for the person you want to invite in.
Instead: Design system: clean, open, friendly