Open-source toolkit for collecting robotic manipulation demonstrations and turning them into training-ready datasets.
A GRABETTE rig records synchronized camera + IMU streams from hand-held or gripper-mounted devices, recovers camera trajectories with SLAM, and exports LeRobot datasets for policy learning. The data-collection pipeline is robot-agnostic.
π© Build the hardware: Bill of Materials Β· CAD β Onshape Β· Assembly guides: Grabette Β· Gripette
grabette_readme_demo.mp4
| Package | Role | Target | Interface |
|---|---|---|---|
grabette |
Hand held data-collection device | Raspberry Pi | HTTP/WebSocket, :8000 |
gripette |
Robot mounted Gripper motor | Raspberry Pi Zero 2W | gRPC, :50051 |
grabette-postprocess |
Data postprocess + SLAM β LeRobot dataset generation | Workstation | CLI |
casquette (WIP) |
POV head-mounted device | Raspberry Pi Zero 2W | HTTP/WebSocket, :8001 |
| Package | Role |
|---|---|
openarm_gripette |
Code to control the OpenArm + Gripette robot |
openarm_gripette_simu |
MuJoCo simulation of OpenArm + Gripette and synthetic data collection |
openarm_gripette_model |
Robot description (URDF / MuJoCo XML) and mesh assets, generated from Onshape |
DiffusionPolicy |
Diffusion Policy training code |
Pi05 |
Ο0.5 VLA fine-tuning, gates, and remote-GPU deployment (shares the DiffusionPolicy dataset pipeline) |
Using GRABETTE with a different robot arm:
the core in packages/ carries no
OpenArm dependency. To target another platform, add an
integrations/<your-arm>/ alongside openarm/ β the OpenArm integration is the
reference example.
The repo uses Git LFS for mesh assets (*.stl, see .gitattributes). Install LFS once per workstation, then clone normally:
sudo apt install git-lfs # Debian / Ubuntu / Pi OS β install the binary first
# (macOS: brew install git-lfs; see git-lfs.com for others)
git lfs install # one-time per user β configures git filters
git clone git@github.com:pollen-robotics/grabette.gitIf you cloned before git lfs install, the .stl files are 130-byte pointer text files. Fetch the real binaries:
cd grabette
git lfs pullVerify:
file packages/grabette/urdf/grabette_right/assets/*.stl | head -3
# expected: "Binary" | bad: "ASCII text" (pointer file β run `git lfs pull`)For on-device installs where you don't need the meshes (Pi services don't load them), skip LFS to save disk + bandwidth:
GIT_LFS_SKIP_SMUDGE=1 git clone git@github.com:pollen-robotics/grabette.gitRequires uv. Python β₯ 3.11.
uv sync --all-packages # full workspace dev environment
uv run --package grabette python packages/grabette/main.py # run a service (mock backend by default)The one rule to know: this repo is a single uv workspace β one shared
.venvand oneuv.lockat the root. A bareuv sync, run from anywhere in the repo, builds the whole workspace and installs every package's dependencies β gigabytes of torch/mujoco on a Raspberry Pi if you're not careful. Therefore:
- Single package / deployment β
uv sync --package <name>(extras attach to it:uv sync --package grabette --extra rpi).- Full dev environment β
uv sync --all-packages.integrations/DiffusionPolicyis deliberately standalone (ownuv.lock, heavy training pins) β inside it a plainuv syncis correct.
On-device install (Raspberry Pi): each device package ships a make install-rpi target that builds the --system-site-packages venv picamera2
needs (a bare uv sync skips it and the service falls back to the mock
backend). Follow the package's own README for the exact steps:
grabette, gripette,
casquette (WIP).
Note β lerobot / Python 3.12: grabette-postprocess's LeRobot commands,
the real OpenArm driver, and the sim's dataset/eval extras require Python
β₯ 3.12 (gated by env markers so the device packages remain installable on
3.11). The OpenArm sim also needs the system liburdfdom package for placo.
Apache-2.0. See LICENSE.