huggingface · nicolas-rabault · Sep 14, 2026 · Jul 15, 2026 · Jul 20, 2026 · Jul 20, 2026
diff --git a/.github/workflows/quality.yml b/.github/workflows/quality.yml
@@ -50,10 +50,11 @@ jobs:
           cache: npm
           cache-dependency-path: frontend/package-lock.json
 
-      - name: Verify committed frontend bundle
+      - name: Verify frontend tests and committed bundle
         working-directory: frontend
         run: |
           npm ci
+          npm test
           npm run build
           if [[ -n "$(git status --porcelain -- dist)" ]]; then
             git diff --stat -- dist

diff --git a/CLAUDE.md b/CLAUDE.md
@@ -39,6 +39,8 @@ Run the Python tests with `pytest` (config in [pyproject.toml](pyproject.toml);
 - [teleoperate.py](lelab/teleoperate.py) — leader→follower teleoperation (wraps `lerobot.teleoperate`).
 - [calibrate.py](lelab/calibrate.py) — step-by-step web calibration with a `CalibrationManager` singleton and `_step_complete` threading.Event.
 - [train.py](lelab/train.py) — wraps the LeRobot training CLI as a subprocess (psutil for lifecycle, queue for log streaming).
+- [datasets.py](lelab/datasets.py) — dataset listing (local cache + Hub, merged under a `source` field) plus the episode-browsing handlers behind `/dataset-episodes`, `/dataset-episode`, `/dataset-frame`, `/dataset-thumbnails`, `/dataset-motion` and `/dataset-video`.
+- [episode_media.py](lelab/episode_media.py) — reads the LeRobot v3.0 on-disk layout (`meta/episodes/*.parquet` + `videos/`) and decodes frames with PyAV. Deliberately does **not** go through `LeRobotDataset`, so browsing a dataset never builds a robot config or imports torch. Several episodes share one mp4, so locating an episode resolves both the file and the `from_/to_timestamp` window inside it — which is why `/dataset-video` is addressed by chunk/file rather than by episode.
 - [dataset_repair.py](lelab/dataset_repair.py) — rebuilds `meta/episodes/` for a recording interrupted before `LeRobotDataset.finalize()` ran. Without an episode index LeRobot reads a local dataset as "not downloaded yet" and fetches it from the Hub, which 404s for a dataset that was never pushed. Call `repair_local_dataset(repo_id)` before opening a recorded dataset with `LeRobotDataset`.
 - [utils/config.py](lelab/utils/config.py) — shared paths and persistence: calibration JSON, saved ports, saved config selections. **Import shared constants from here, do not hardcode paths in feature modules.**
 

diff --git a/README.md b/README.md
@@ -40,6 +40,10 @@ A page will automatically open in your browser and you are ready to go.
       <td>📹 <b>Record</b></td>
       <td>Capture episodes into a LeRobotDataset, with cameras.</td>
     </tr>
+    <tr>
+      <td>🔎 <b>Browse</b></td>
+      <td>Watch the episodes of a dataset on disk — every camera, frame by frame, with a joint-motion trace. No upload required.</td>
+    </tr>
     <tr>
       <td>🧠 <b>Train</b></td>
       <td>Kick off a LeRobot training job, watch logs live.</td>

diff --git a/frontend/dist/assets/index-1vZ-UTJX.css b/frontend/dist/assets/index-1vZ-UTJX.css
diff --git a/frontend/dist/assets/index-BX9JKcmw.js b/frontend/dist/assets/index-BX9JKcmw.js
diff --git a/frontend/dist/assets/index-Cl3yeJ80.js b/frontend/dist/assets/index-Cl3yeJ80.js
diff --git a/frontend/dist/assets/index-DJuBJMLM.css b/frontend/dist/assets/index-DJuBJMLM.css
diff --git a/frontend/dist/index.html b/frontend/dist/index.html
@@ -10,8 +10,8 @@
     <meta property="og:title" content="LeLab" />
     <meta property="og:description" content="LeRobot but on your browser." />
     <meta property="og:type" content="website" />
-    <script type="module" crossorigin src="/assets/index-Cl3yeJ80.js"></script>
-    <link rel="stylesheet" crossorigin href="/assets/index-1vZ-UTJX.css">
+    <script type="module" crossorigin src="/assets/index-BX9JKcmw.js"></script>
+    <link rel="stylesheet" crossorigin href="/assets/index-DJuBJMLM.css">
   </head>
 
   <body>