robostudio-engine

robostudio-engine is the headless Python dataset engine and robostudio CLI for inspecting, indexing, reviewing, clustering, probing, and exporting robotics dataset metadata.

At a Glance

Install

Core commands have no required third-party runtime dependencies:

python -m pip install "robostudio-engine==0.1.2"

Install only the extras needed by a workflow:

python -m pip install "robostudio-engine[hdf5]==0.1.2"
python -m pip install "robostudio-engine[hf]==0.1.2"
python -m pip install "robostudio-engine[ml]==0.1.2"
python -m pip install "robostudio-engine[trust]==0.1.2"

The ml extra includes large model and clustering dependencies. The trust extra installs the published companion trust tools used by the wrapper commands.

Verified Quickstart

Run from a source checkout:

robostudio inspect examples/mock_multi_format/lerobot_v3

The bundled fixture is detected as lerobot and reports two episodes. inspect is read-only.

Core Command Surface

Adapters cover the checked-in LeRobot v2/v3, RLDS, OpenX, HDF5 profile, rosbag, and folder-of-MP4-plus-JSONL shapes. Support is fixture-backed metadata parsing, not a claim of compatibility with every producer or version.

Companion Trust Tools

The engine contains executable wrappers for:

• lerobot-quality-gates;
• robot-recovery-bench;
• vla-robustness-kit;
• embodiment-card;
• robotics-reviewkit;
• failure-gallery.

These projects are not required by the core engine. Install the trust extra before using quality-gates, recovery, probe, card, or failure-gallery preview behavior. robotics-reviewkit remains a source-distributed companion and must be installed separately for validate-review. In the AuraOne monorepo, the engine can also discover sibling src directories.

Runtime, Data, and Network Boundary

• Core adapters, indexing, QA, hash clustering, plugin validation, and local exports operate on local files.
• index and thumbs write inside the dataset's .robostudio directory. ffmpeg is invoked locally for supported thumbnail paths.
• inspect, metadata bridge exports, and some reports can contain local paths or source metadata. Review outputs before sharing.
• Training-ready manifest and intake exports scrub absolute paths and secret-like fields and exclude raw media. Metadata bridge JSONL preserves normalized episode metadata and needs separate review.
• Intake export writes a local ZIP with an intended destination; it does not transmit the packet.
• Hugging Face upload occurs only with the hf extra plus HF_TOKEN or HUGGINGFACE_TOKEN. Without them, the command prepares local files and a blocker note.
• CLIP clustering can download model files into ~/.cache/robostudio; a custom encoder or BYO VLA policy runs an explicit local executable.

Limitations

• Adapter support is fixture-backed parsing for known layouts, not a claim of compatibility with every dataset producer, codec, or schema revision.
• The wheel contains only robostudio_engine; companion trust tools are installed through the trust extra or discovered from a source checkout.
• robotics-reviewkit is not included in the trust extra because it is not a standalone PyPI distribution.

Robotics Studio Open Integration

Robotics Studio Open documents that the robostudio CLI ships with its desktop installers. That is a direct executable product integration. The standalone package remains useful for CI and headless review without installing the desktop application.

Publication Status

Verified on 2026-07-13:

• PyPI: robostudio-engine==0.1.2
• Source: auraoneai/robostudio-engine
• GitHub release: v0.1.2
• The separate Robotics Studio Open product has a v0.2.0 release; that product release is not a second robostudio-engine package release.
• Checked-in datasets and media are synthetic fixtures.

Next Action

Run inspect against a copy of one review candidate, then run qa and inspect all emitted paths and metadata before creating an export.