What is Greg (high-five-robot)?
Greg is a 3D-printed robot arm with one job: hold up an open palm and it reaches out and high-fives you. A camera in the base streams video over Wi-Fi to a small Python server on your computer, which finds the palm with Google's MediaPipe Hand Landmarker and works out the joint angles; an ESP32-S3 board then drives five servos, striking about 3 cm past the touch point before returning home.
The repository holds the parametric CAD (Blender Python scripts), ready-made STL files in five print stages, Arduino firmware for both boards, the vision and calibration software, a wiring guide and a bill of materials. Andris Gauracs built the robot by hand; according to the README, Claude wrote the CAD, both firmwares and the vision code in one long Claude Code session.
It is MIT licensed, written mostly in Python, and was published on 5 November 2026. On the day we checked it had 1 star, no forks, 3 commits and no releases.
Is Greg (high-five-robot) worth it?
Greg is worth studying because it covers the whole path from a camera frame to a moving joint in one small, readable project: printed parts, two microcontrollers, a vision model, a calibration step and a learned correction for the wrist. The catch is that it has been built once, on one Mac, and the README says so.
The parts are cheap and ordinary. The bill of materials lists three MG995 servos for pan, shoulder and elbow, two MG90S micro servos for the wrist, an ESP32-S3 N16R8 board for the arm, a Seeed XIAO ESP32S3 Sense with a wide-angle OV2640 lens as the camera, a 5 V 6 A power supply, a perfboard, M3 screws, a wooden base and about 1 kg of PLA or PETG. The author's estimate is $140 to $230 before tax and shipping, using November 2026 retail listings, and that excludes the printer and tools.
The vision side is off the shelf. The server runs Google's MediaPipe Hand Landmarker, a float16 bundle of a palm detector and a model that places 21 keypoints on each hand; Google reports 17.12 ms per frame on a Pixel 6 CPU. Greg uses it to check for an open, raised palm, then a calibrated camera model turns the palm position into joint angles and a small learned model sets the wrist and twist from about 15 examples you teach it by voice in Chrome.
Difficulty sits in the calibration, not the code. Servo clones vary by about 1 mm, so the build starts with printed fit tests and may mean editing the CAD and regenerating parts in Blender 4.3. Camera calibration sweeps about 268 poses while tracking a red label on the hand, aiming for an error of about 5 px, and the README stresses that its fitted model only matches the original robot. A write-up on Stork.ai describes the real-world fixes along the way: servo mounts with no room for wires, wrong guesses about the camera's thickness, servos that shook until they got their own power path and shared ground. Total build time is not published; the base alone takes 8 to 10 hours to print.
As a learning project it teaches more than an assembled kit would, precisely because the rough edges are documented. Anyone wanting a robot arm for general tasks will find Greg does one motion, runs only with a laptop server alongside, and "hits hard" in the author's words, so faces stay out of reach.
Where it earns its place
A software developer with a 3D printer and no robotics background
Learn how a camera image becomes a joint movement
Print the five stages, wire the two ESP32 boards and follow the calibration steps; the code is short enough to read end to end. Expect $140 to $230 in parts and several evenings of fit tests and calibration.
A secondary-school or university robotics club
Run a term project that covers CAD, firmware and computer vision
Split the build by layer: one group prints and adapts the Blender CAD, one flashes the arduino-cli firmware, one runs the MediaPipe server and calibration. The README's warning that Greg hits hard means the club needs low-speed testing and a rule about faces.
A maker who already uses Claude Code for software
See how far an AI coding agent gets on a physical build
Read the repository and the Stork.ai write-up together: the agent produced the CAD, firmware and vision code, and the documented fixes for wire clearance, camera fit and servo power show where a person still had to measure and test.
Before you commit
- Setup
- Print fit tests, then five stages on a 225 x 225 mm bed (PLA or PETG, 0.2 mm layers, no supports); flash both boards with arduino-cli and ESP32 core 3.3.12 over 2.4 GHz Wi-Fi; install the Python requirements, download hand_landmarker.task, trim the joints, calibrate the camera and teach the wrist. Total build time is not published.
- Integrations
- MediaPipe Hand Landmarker for palm detection; Blender for regenerating the parametric CAD; arduino-cli for firmware; local web pages on ports 8090 and 8091 for jogging, voice teaching and the ARM ON switch; an optional install script to start the server on a Mac.
- Data and privacy
- Video stays on your network between the camera board and your computer; the vision model runs locally. MIT licence, with Espressif's Apache 2.0 headers kept on the camera firmware. No security advisories published.
- Cost at scale
- About $140 to $230 in parts per robot by the author's November 2026 estimate, plus a 3D printer and about 1 kg of filament; no paid edition.
- Limits
- Built and tested once, on macOS; no releases; the calibrated model fits only the original robot; one motion only; the README warns it strikes hard and should be tested at low speed first.
What’s good
- Every layer included: CAD, STLs, firmware, vision server and wiring guide
- Parts estimated at $140 to $230 from common hobby suppliers
- Uses MediaPipe's ready-made hand model, so no training from scratch
- Printed fit tests and editable CAD absorb servo size differences
What’s not
- Published on 5 November 2026 with 1 star and no other known builds
- Tested only on macOS with Python 3.12
- Calibration of about 268 camera poses has to be redone for each robot
- Needs a computer running the server; the arm cannot work alone
Best for
Makers with a 3D printer and some soldering experience who want to learn how camera vision, calibration and servo control fit together in one small robot.
Greg (high-five-robot): common questions
- Is Greg (high-five-robot) worth it?
- A charming, well-documented robotics lesson that has been built exactly once, so treat it as a careful recipe to adapt rather than a kit that works first time. We rate it 3.5 out of 5.
- How much does Greg (high-five-robot) cost?
- Free under the MIT licence; parts cost about $140 to $230, plus a 3D printer.
- What are the downsides of Greg (high-five-robot)?
- Published on 5 November 2026 with 1 star and no other known builds; tested only on macOS with Python 3.12; calibration of about 268 camera poses has to be redone for each robot; needs a computer running the server; the arm cannot work alone.
- Who is Greg (high-five-robot) best for?
- Makers with a 3D printer and some soldering experience who want to learn how camera vision, calibration and servo control fit together in one small robot.
- How much does Greg (high-five-robot) cost for a team or in production?
- About $140 to $230 in parts per robot by the author's November 2026 estimate, plus a 3D printer and about 1 kg of filament; no paid edition.
- Is Greg (high-five-robot) safe for company data?
- Video stays on your network between the camera board and your computer; the vision model runs locally. MIT licence, with Espressif's Apache 2.0 headers kept on the camera firmware. No security advisories published.
Answers checked Nov 5, 2026.