The pipeline is calibrated for GoPro Hero9 or Hero10 cameras that are now hard to find, and the CAD has no stated license.
A hand-held, 3D-printed parallel-jaw gripper from Stanford, Columbia and Toyota Research Institute used to record human demonstrations for robot learning. A GoPro with a fisheye lens and side mirrors captures video, and a SLAM pipeline recovers the gripper's motion. Policies trained on the data run on robot arms such as the UR5e or Franka fitted with a matching gripper.
At a glance
| Parts cost | $370 (2024) project · Stated by the project source ↗ |
|---|---|
| Difficulty | 2/5 Easy · The gripper is a simple print-and-bolt build with linear rails, but the data pipeline and robot deployment are much harder. |
| Build time | Not stated |
| Tools needed | 3D printer (PLA+ body and TPU fingers, 0.6 mm nozzle), Linux PC with Docker for the SLAM pipeline and a GPU for policy training (unconfirmed) |
| Motors | None on the hand-held gripper (human-powered); deployment uses a WSG-50 gripper on the robot |
| Controller | None on the gripper; GoPro records to SD card |
| Sensors | GoPro Hero9 or Hero10 with Max Lens Mod, IMU in the GoPro, ArUco tags for gripper width |
| Published | ✓ Parts list✓ CAD/STL✓ Assembly guide✗ Wiring✓ Software✓ Video Google Doc hardware guide + YouTube printing and assembly tutorials + Onshape + GitHub code |
| Activity | Last update 2026-05-26 · 1,597 GitHub stars · 290 forks; very active issue tracker and many derivative projects |
| Kits | None found |
Licenses
No license is stated for the design files. They are published for building, but without a license the right to copy, modify or redistribute them is not granted.
| Design files | None found (public Onshape documents with no license) source ↗ |
|---|---|
| Code | MIT source ↗ |
| Documentation | None found source ↗ |
Builder notes
Practical points from the project's issue tracker, forum and docs, each linked to its source.
- The guide specifies a GoPro Hero9, which is hard to buy now. The authors say a Hero10 works as-is, while a Hero11 or 12 needs new camera and IMU calibration. source ↗
- Builders using a GoPro Hero13 could not get mapping to work, since no firmware supports the needed 4:3 2.7K 60 fps mode. source ↗
- SLAM mapping failures on custom recordings were fixed by recording in brighter, more textured scenes and holding the gripper still for a few seconds at the start. source ↗
- The example pipeline expects videos in a raw_videos subfolder and needs Docker, which the README does not spell out. source ↗
- Using the collected data on a robot needs a separate deployment setup, with a UR5e or Franka arm, a Weiss WSG-50 gripper (about 3,875 EUR) and a capture card, which costs far more than the hand-held gripper. source ↗
Other hands plans
- LEAP Hand · $1,800 (2023) · difficulty 2/5
- AmazingHand · $140 · difficulty 2/5
- RUKA hand · $1,300 · difficulty 3/5
- Aero Hand Open · $320 · difficulty 3/5
- ORCA Hand · $2,000 · difficulty 4/5
- YUBI glove and gripper (Toyota) · $550–$750 · difficulty 4/5
- Yale OpenHand (Model T42 / Model O) · $1,400 · difficulty 3/5
- DexHand (The Robot Studio) · $300 (2023) · difficulty 4/5