LeRobot Viewer
Browse, scrub and inspect Hugging Face LeRobot datasets directly inside VS Code — no Python, no Rerun install required.
Features
- Multi-source dataset browser — local folders, Hugging Face repos, and remote SSH hosts in one tree
- Episode preview with synchronized video, signal graphs, 2D trajectory plot, and a smooth 60 fps timeline cursor
- Native parquet decoding (via
hyparquet) for observation.state, action, velocity, effort, environment_state, next.reward, next.done, next.success, next.truncated, task_index
- Dataset metadata viewer — KPIs, splits, tasks, schema groups, dataset-wide signal range cards, raw
info.json
- Format coverage — LeRobot v2.0 / v2.1 (per-episode parquet/mp4) and v3.0 (sharded), with metadata-driven shard lookup
- Remote SSH datasets — pick a host from
~/.ssh/config, browse remote folders interactively, mirror metadata to local cache, lazy-fetch parquet/video on demand
- Keyboard shortcuts —
Space play/pause, ← → ±1 frame, J/L ±1 second, R loop, Home/End jump
Quick start
- Install the extension from the VS Code marketplace.
- Open a folder containing a LeRobot dataset (
meta/info.json is the marker).
- Click the LeRobot icon in the activity bar — datasets are auto-detected.
- Click an episode to preview, or click Metadata to open the full schema viewer.
Remote (SSH) datasets
The toolbar's 🌐 button opens a wizard:
- Pick a host from
~/.ssh/config, or enter user@host[:port] manually.
- Browse the remote filesystem with QuickPick — folders that look like LeRobot datasets are tagged.
- Pick the folder. The extension mirrors
meta/ locally and lazy-downloads parquet/video on first preview.
Authentication priority: ssh-agent → identity file → password / passphrase prompt. If you hit the ssh2 OpenSSH integrity-check bug, run ssh-add <key> once and retry.
Hugging Face datasets
Click the cloud icon and enter a namespace/name repo id. Metadata is fetched into a local cache; data files load on demand from the HF CDN.
Settings
| Setting |
Default |
What it does |
lerobotViewer.autoScanWorkspace |
true |
Auto-detect datasets in workspace folders on activation |
lerobotViewer.workspaceScanDepth |
3 |
How deep to recurse when scanning |
lerobotViewer.huggingfaceCacheDir |
"" |
Override ~/.cache/lerobot-viewer |
lerobotViewer.rerunExecutable |
"rerun" |
Path to the optional Rerun viewer binary |
lerobotViewer.enableRerun |
false |
Surface the "Open in Rerun" button (requires Rerun installed) |
Architecture
src/
├── dataset/
│ ├── adapters/ ← v2.x and v3.0 dataset format adapters
│ ├── ssh/ ← config / connection / browser / fetch
│ ├── DatasetVersionDetector.ts
│ ├── datasetLoader.ts ← top-level facade
│ ├── datasetService.ts
│ ├── huggingface.ts
│ └── parquetReader.ts ← hyparquet decoding
├── webview/
│ ├── baseWebviewPanel.ts ← shared panel scaffolding
│ ├── episodePreviewPanel.ts
│ ├── metadataViewerPanel.ts
│ └── protocol.ts
└── extension.ts
webview/
├── App.tsx ← episode preview React tree
├── MetadataView.tsx ← metadata viewer React tree
├── components/ ← Header, TransportBar, SignalGraph, ...
└── hooks/ ← usePlayback, usePlaybackShortcuts
Adapters are pure-Node (no vscode import) so they're unit-testable. The DatasetVersionDetector reads codebase_version from info.json and falls back to layout heuristics for ambiguous datasets.
Development
npm install
npm run watch # esbuild + tailwind in watch mode
# F5 in VS Code → "Run Extension"
npm test # node:test against adapters / config parser
npm run typecheck
npm run build # production bundle
License
Apache-2.0
| |
