title	Grabette Slam
emoji	🌍
colorFrom	blue
colorTo	indigo
sdk	docker
app_port	7860
hf_oauth	true
hf_oauth_scopes	read-repos write-repos manage-repos

Grabette SLAM → LeRobot

A HuggingFace Space that turns a raw Grabette recording dataset into a LeRobot dataset and pushes it to the Hub.

1. Sign in with your Hugging Face account (OAuth).
2. Give a source dataset repo_id (raw Grabette recording on the Hub), a target repo_id to create, and a task description.
3. For each episode the Space runs, in-process:
   • convert_episode — expand the recording into the slam_input/ layout
   • run_slam — RGB-D odometry (visual-inertial when the camera has an IMU) via the bundled offline_vslam binary
   • build_dataset — assemble a LeRobot v3 dataset
   • push_to_hub — upload under your account
4. A link + embedded view of the LeRobot visualizer is shown for the result (the dataset must be public to be visualized).

Access the spaces

• main space
• test space

Why a Docker Space

The SLAM step is a compiled C++ binary (offline_vslam, built on RTAB-Map). Locally it runs in the pollenrobotics/grabette-vslam Docker image, but Spaces cannot run Docker-in-Docker. So this Space is that image: it builds the binary at image-build time and run_slam(..., binary=...) calls it directly.

Self-contained build

The pipeline lives in this repo: ./grabette-postprocess is the package the Space installs (with offline_vslam.cpp, compiled at image-build time) and is its source of truth. The Docker build needs nothing outside this tree.

Note: RTAB-Map is compiled with --parallel 2 to stay within the HF build runner's memory (full parallelism OOMs it).

Deployment

Both Spaces build from this repo, through GitHub Actions:

• prod: every push to main syncs it to the main Space (sync-to-hf-space.yml).
• test: run Deploy a branch to HF test Space (deploy-branch-to-hf-test-space.yml) from the Actions tab, on any branch, tag or SHA.

Local development

Native (fast, for the UI and pipeline glue). From the repo root:

uv venv --python 3.12
uv pip install -e ./grabette-postprocess -r requirements.txt
hf auth login    # once: the local Sign-in button is mocked with this account
.venv/bin/python app.py    # → http://localhost:7860

The SLAM step needs the offline_vslam binary (VSLAM_BINARY, default /usr/local/bin/offline_vslam), which is only built inside the Docker image. For anything that runs SLAM, use the image.

Docker (the same image as the Space).

docker build -t grabette-slam .
docker run --rm -p 7860:7860 -e SYSTEM= \
  -v ~/.cache/huggingface:/home/user/.cache/huggingface grabette-slam

SYSTEM= clears the image's SYSTEM=spaces, so Gradio mocks the sign-in with your local Hugging Face login instead of expecting the Space's OAuth variables.

On Apple Silicon, turn on Docker Desktop → Settings → General → "Use Rosetta for x86_64/amd64 emulation": the image is amd64-only, and under the default qemu emulation the app segfaults at startup.

Tests: see grabette-postprocess/README.md.