pollen-robotics/grabette-slam

Grabette SLAM → LeRobot

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

1. 1.Sign in with your Hugging Face account (OAuth).
2. 2.Give a source dataset repo_id (raw OAK-D recording on the Hub), a target repo_id to create, and a task description.
3. 3.For each episode the Space runs, in-process:
4. 4.convert_episode — expand the recording into the oak/ layout
5. 5.run_oak_slam — RGBD-inertial odometry via the bundled offline_vslam binary
6. 6.build_dataset — assemble a LeRobot v3 dataset
7. 7.push_to_hub — upload under your account
8. 8.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/oak-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_oak_slam(..., binary=...) calls it directly.

Self-contained build

The grabette-postprocess package (with offline_vslam.cpp) is vendored into the build context as ./grabette-postprocess — no git clone, no private-repo secret. deploy.sh assembles that layout from the working tree and uploads it.

HF="uv run --project ../../packages/grabette-postprocess hf" \
  ./deploy.sh pollen-robotics/grabette-slam

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