GRABETTE

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 arm-agnostic — OpenArm is provided as one worked integration, not a requirement.

This is a uv workspace monorepo. It supersedes the former per-component repositories (grabette, casquette, gripette, grabette-data, openarm_gripette_simu, openarm_gripette_model), which are archived read-only. Electronics and screen firmware live in a separate hardware repo.

Components

packages/ — arm-agnostic core

Package	Role	Target	Interface
grabette	On-device data-collection service (camera + IMU + angle/OAK-D)	Raspberry Pi	HTTP/WebSocket, :8000
casquette	POV head-mounted camera + IMU collection service	Raspberry Pi Zero 2W	HTTP/WebSocket, :8001
gripette	Gripper motor + camera service	Raspberry Pi Zero 2W	gRPC, :50051
grabette-postprocess	SLAM/VIO (OAK-D + RTAB-Map, Dockerized) → LeRobot dataset generation	Workstation	CLI

integrations/openarm/ — reference integration (OpenArm 7-DOF arm)

Package	Role
openarm_gripette_simu	MuJoCo simulation of OpenArm + Gripette, with gRPC gripper (:50051) and arm (:50052) control and synthetic data collection
openarm_gripette_model	Robot description (URDF / MuJoCo XML) and mesh assets, generated from Onshape

Using GRABETTE with a different arm: the core in packages/ carries no OpenArm dependency. To target another platform, add an integrations/<your-arm>/ alongside openarm/ — the OpenArm integration is the reference example.

Layout

packages/                       arm-agnostic core (uv workspace members)
integrations/openarm/           reference integration for the OpenArm arm
pyproject.toml                  uv workspace root
uv.lock                         single lock for the whole workspace

Development

Requires uv. Python ≥ 3.11.

uv sync --all-packages          # build the full workspace environment
uv run --package grabette python packages/grabette/main.py   # run a service (mock backend by default)

Work on a single component without pulling the rest:

uv sync --package grabette-postprocess
uv run --package grabette-postprocess python scripts/arducam_slam/generate_dataset.py --help

Notes

• Python / lerobot: lerobot (used by grabette-postprocess and the sim's dataset extra) requires Python ≥ 3.12, so it is gated by an environment marker. The on-device services still install and run on Python 3.11 (Raspberry Pi OS Bookworm) — they don't depend on lerobot.
• OpenArm sim system deps: placo (sim kinematics) dynamically links liburdfdom; install it from your distro's packages before running the sim.
• Raspberry Pi install: use the make install-rpi target (see below) — a bare uv sync skips the apt deps and --system-site-packages venv that picamera2 needs, and the service silently falls back to the mock backend.

Running on a Raspberry Pi device

The on-device services (grabette, casquette, gripette) install through a make target that builds the --system-site-packages venv at the workspace root, with the device's apt-provided picamera2/libcamera:

# clone skipping the OpenArm meshes (not needed on-device)
GIT_LFS_SKIP_SMUDGE=1 git clone git@github.com:pollen-robotics/grabette.git
cd grabette/packages/grabette

make install-rpi                               # apt deps + venv + sync + verify imports
uv run --package grabette python -m grabette   # run the service (auto-detects hardware)
make install-systemd                           # optional: enable boot-time autostart

On success the log shows RPi hardware detected, using RpiBackend (not MockBackend). casquette follows the same pattern from packages/casquette; run make help in either package for the available targets.

License

Apache-2.0. See LICENSE.