feat(vla): add SmolVLA conditioning and experiment artifacts

tests/test_simple_robot_dataset_image_loading.py

@@ -12,10 +12,13 @@ from roboimi.vla.data.simpe_robot_dataset import SimpleRobotDataset
 
 
 class SimpleRobotDatasetImageLoadingTest(unittest.TestCase):
-    def _write_episode(self, dataset_dir: Path) -> None:
-        episode_path = dataset_dir / "episode_0.hdf5"
+    def _write_episode(self, dataset_dir: Path, episode_idx: int = 0, action_offset: float = 0.0) -> None:
+        episode_path = dataset_dir / f"episode_{episode_idx}.hdf5"
         with h5py.File(episode_path, "w") as root:
-            root.create_dataset("action", data=np.arange(8, dtype=np.float32).reshape(4, 2))
+            root.create_dataset(
+                "action",
+                data=(np.arange(8, dtype=np.float32).reshape(4, 2) + action_offset),
+            )
             root.create_dataset(
                 "observations/qpos",
                 data=np.arange(16, dtype=np.float32).reshape(4, 4),
@@ -79,3 +82,22 @@ class SimpleRobotDatasetImageLoadingTest(unittest.TestCase):
 
         fake_cv2.resize.assert_not_called()
         self.assertEqual(tuple(sample["observation.front"].shape), (2, 3, 8, 8))
+
+    def test_dataset_can_filter_by_episode_indices_and_expose_available_episode_indices(self):
+        with tempfile.TemporaryDirectory() as tmpdir:
+            dataset_dir = Path(tmpdir)
+            self._write_episode(dataset_dir, episode_idx=3, action_offset=0.0)
+            self._write_episode(dataset_dir, episode_idx=7, action_offset=100.0)
+
+            dataset = SimpleRobotDataset(
+                dataset_dir,
+                obs_horizon=2,
+                pred_horizon=3,
+                camera_names=["front"],
+                episode_indices=[7],
+            )
+            sample = dataset[0]
+
+        self.assertEqual(dataset.available_episode_indices, [7])
+        self.assertEqual(len(dataset), 4)
+        self.assertEqual(sample["action"][0, 0].item(), 100.0)