Merge remote-tracking branch 'origin/main' into feat/smolvla-on-steerable

# Conflicts:
#	docs/source/annotation_pipeline.mdx
#	examples/annotations/run_hf_job.py
#	pyproject.toml
#	src/lerobot/annotations/steerable_pipeline/config.py
#	src/lerobot/annotations/steerable_pipeline/frames.py
#	src/lerobot/annotations/steerable_pipeline/modules/plan_subtasks_memory.py
#	src/lerobot/annotations/steerable_pipeline/vlm_client.py
#	src/lerobot/annotations/steerable_pipeline/writer.py
#	src/lerobot/datasets/__init__.py
#	src/lerobot/datasets/sampler.py
#	src/lerobot/scripts/lerobot_annotate.py
#	src/lerobot/scripts/lerobot_train.py
#	tests/annotations/test_frames.py
#	tests/annotations/test_modules.py
#	tests/annotations/test_writer.py
#	tests/datasets/test_sampler.py
#	tests/scripts/test_lerobot_annotate.py
#	uv.lock

tests/datasets/test_aggregate.py

@@ -32,6 +32,26 @@ from lerobot.datasets.lerobot_dataset import LeRobotDataset
 from tests.fixtures.constants import DUMMY_REPO_ID
 
 
+def assert_data_shards_one_row_group_per_episode(root):
+    """Every aggregated DATA shard must have exactly one parquet row group per episode."""
+    import pyarrow.parquet as pq
+
+    shards = sorted((root / "data").rglob("*.parquet"))
+    assert shards, f"no data shards found under {root}/data"
+    n_episodes = 0
+    for shard in shards:
+        pf = pq.ParquetFile(shard)
+        episodes = pf.read(columns=["episode_index"]).column("episode_index").to_pylist()
+        assert pf.metadata.num_row_groups == len(set(episodes)), shard
+        for i in range(pf.metadata.num_row_groups):
+            rg_episodes = set(
+                pf.read_row_group(i, columns=["episode_index"]).column("episode_index").to_pylist()
+            )
+            assert len(rg_episodes) == 1, f"{shard} row group {i} spans episodes {rg_episodes}"
+        n_episodes += len(set(episodes))
+    return n_episodes
+
+
 def assert_episode_and_frame_counts(aggr_ds, expected_episodes, expected_frames):
     """Test that total number of episodes and frames are correctly aggregated."""
     assert aggr_ds.num_episodes == expected_episodes, (
@@ -289,6 +309,52 @@ def test_aggregate_datasets(tmp_path, lerobot_dataset_factory):
     assert_dataset_iteration_works(aggr_ds)
 
 
+def test_aggregate_datasets_without_concatenation(tmp_path, lerobot_dataset_factory):
+    """With concatenation disabled, each source file is kept as its own destination file."""
+    ds_0 = lerobot_dataset_factory(
+        root=tmp_path / "no_stitch_0",
+        repo_id=f"{DUMMY_REPO_ID}_no_stitch_0",
+        total_episodes=3,
+        total_frames=60,
+    )
+    ds_1 = lerobot_dataset_factory(
+        root=tmp_path / "no_stitch_1",
+        repo_id=f"{DUMMY_REPO_ID}_no_stitch_1",
+        total_episodes=4,
+        total_frames=80,
+    )
+
+    aggr_root = tmp_path / "no_stitch_aggr"
+    aggregate_datasets(
+        repo_ids=[ds_0.repo_id, ds_1.repo_id],
+        roots=[ds_0.root, ds_1.root],
+        aggr_repo_id=f"{DUMMY_REPO_ID}_no_stitch_aggr",
+        aggr_root=aggr_root,
+        concatenate_videos=False,
+        concatenate_data=False,
+    )
+
+    with (
+        patch("lerobot.datasets.dataset_metadata.get_safe_version") as mock_get_safe_version,
+        patch("lerobot.datasets.dataset_metadata.snapshot_download") as mock_snapshot_download,
+    ):
+        mock_get_safe_version.return_value = "v3.0"
+        mock_snapshot_download.return_value = str(aggr_root)
+        aggr_ds = LeRobotDataset(f"{DUMMY_REPO_ID}_no_stitch_aggr", root=aggr_root)
+
+    assert_episode_and_frame_counts(
+        aggr_ds, ds_0.num_episodes + ds_1.num_episodes, ds_0.num_frames + ds_1.num_frames
+    )
+    assert_dataset_iteration_works(aggr_ds)
+    assert_video_timestamps_within_bounds(aggr_ds)
+
+    # Two single-file sources stay as two files each, instead of being packed together.
+    assert len(list((aggr_root / "data").rglob("*.parquet"))) == 2
+    assert aggr_ds.meta.video_keys, "Test fixture should produce at least one video feature"
+    for key in aggr_ds.meta.video_keys:
+        assert len(list((aggr_root / "videos" / key).rglob("*.mp4"))) == 2
+
+
 @pytest.mark.parametrize("mutation", ["mismatched_value", "missing_key"])
 def test_aggregate_incomplete_video_encoder_info_warns_and_nuls_encoders(
     tmp_path, lerobot_dataset_factory, caplog, mutation
@@ -520,6 +586,41 @@ def assert_image_frames_integrity(aggr_ds, ds_0, ds_1):
             )
 
 
+@pytest.mark.parametrize("use_videos", [True, False], ids=["video", "image"])
+def test_aggregate_one_row_group_per_episode(tmp_path, lerobot_dataset_factory, use_videos):
+    """Aggregated DATA shards keep one row group per episode (not one collapsed group).
+
+    Covers both the non-image (``df.to_parquet``) and image
+    (``to_parquet_with_hf_images``) write branches, including the merge-into-
+    existing-file branch via a low file-size threshold that forces packing.
+    """
+    ds_0 = lerobot_dataset_factory(
+        root=tmp_path / "rg_0",
+        repo_id=f"{DUMMY_REPO_ID}_rg_0",
+        total_episodes=3,
+        total_frames=60,
+        use_videos=use_videos,
+    )
+    ds_1 = lerobot_dataset_factory(
+        root=tmp_path / "rg_1",
+        repo_id=f"{DUMMY_REPO_ID}_rg_1",
+        total_episodes=4,
+        total_frames=80,
+        use_videos=use_videos,
+    )
+
+    aggr_root = tmp_path / "rg_aggr"
+    aggregate_datasets(
+        repo_ids=[ds_0.repo_id, ds_1.repo_id],
+        roots=[ds_0.root, ds_1.root],
+        aggr_repo_id=f"{DUMMY_REPO_ID}_rg_aggr",
+        aggr_root=aggr_root,
+    )
+
+    n_episodes = assert_data_shards_one_row_group_per_episode(aggr_root)
+    assert n_episodes == ds_0.num_episodes + ds_1.num_episodes
+
+
 def test_aggregate_image_datasets(tmp_path, lerobot_dataset_factory):
     """Test aggregation of image-based datasets preserves HuggingFace Image schema.
 

tests/datasets/test_compute_stats.py

@@ -83,6 +83,29 @@ def test_get_feature_stats_images():
     assert stats["min"].shape == stats["max"].shape == stats["mean"].shape == stats["std"].shape
 
 
+def test_get_feature_stats_uint8_images_preserves_std():
+    data = np.array(
+        [
+            [
+                [[0, 64], [128, 255]],
+                [[255, 128], [64, 0]],
+                [[32, 96], [160, 224]],
+            ],
+            [
+                [[16, 80], [144, 240]],
+                [[240, 144], [80, 16]],
+                [[48, 112], [176, 208]],
+            ],
+        ],
+        dtype=np.uint8,
+    )
+
+    stats = get_feature_stats(data, axis=(0, 2, 3), keepdims=True)
+
+    expected_std = data.transpose(0, 2, 3, 1).reshape(-1, 3).std(axis=0).reshape(1, 3, 1, 1)
+    np.testing.assert_allclose(stats["std"], expected_std)
+
+
 def test_get_feature_stats_axis_0_keepdims(sample_array):
     expected = {
         "min": np.array([[1, 2, 3]]),

tests/datasets/test_datasets.py

@@ -51,7 +51,7 @@ from lerobot.robots import make_robot_from_config
 from lerobot.transforms import ImageTransforms, ImageTransformsConfig
 from lerobot.utils.constants import ACTION, DONE, OBS_IMAGES, OBS_STATE, OBS_STR, REWARD
 from lerobot.utils.feature_utils import hw_to_dataset_features
-from tests.fixtures.constants import DUMMY_CHW, DUMMY_HWC, DUMMY_REPO_ID
+from tests.fixtures.constants import DUMMY_CHW, DUMMY_HWC, DUMMY_MOTOR_FEATURES, DUMMY_REPO_ID
 from tests.mocks.mock_robot import MockRobotConfig
 from tests.utils import require_x86_64_kernel
 
@@ -133,6 +133,21 @@ def test_dataset_feature_with_forward_slash_raises_error():
         )
 
 
+def test_create_does_not_mutate_input_features(tmp_path, empty_lerobot_dataset_factory):
+    # ``create`` must deep-copy features so a dataset built from another's features stays independent.
+    dataset = empty_lerobot_dataset_factory(
+        root=tmp_path / "ds1", features=DUMMY_MOTOR_FEATURES, use_videos=False
+    )
+    dataset_copy = empty_lerobot_dataset_factory(
+        root=tmp_path / "ds2", features=dataset.meta.features, use_videos=False
+    )
+
+    original_shape = dataset.meta.info.features["state"]["shape"]
+    dataset_copy.meta.info.features["state"]["shape"] = (999,)
+
+    assert dataset.meta.info.features["state"]["shape"] == original_shape
+
+
 def test_add_frame_missing_task(tmp_path, empty_lerobot_dataset_factory):
     features = {"state": {"dtype": "float32", "shape": (1,), "names": None}}
     dataset = empty_lerobot_dataset_factory(root=tmp_path / "test", features=features)

tests/datasets/test_sampler.py

@@ -25,7 +25,7 @@ from datasets import Dataset  # noqa: E402
 from lerobot.datasets.io_utils import (
     hf_transform_to_torch,
 )
-from lerobot.datasets.sampler import EpisodeAwareSampler, WeightedEpisodeAwareSampler
+from lerobot.datasets.sampler import EpisodeAwareSampler, WeightedEpisodeAwareSampler, compute_sampler_state
 
 
 def calculate_episode_data_index(hf_dataset: Dataset) -> dict[str, torch.Tensor]:
@@ -114,6 +114,19 @@ def test_shuffle():
     assert set(sampler) == {0, 1, 2, 3, 4, 5}
 
 
+def test_shuffle_is_reproducible_across_instances():
+    # The order is a pure function of (seed, epoch), so two fresh samplers (e.g. two ranks)
+    # produce the same permutation without any generator synchronization.
+    sampler_a = EpisodeAwareSampler([0], [6], shuffle=True, seed=42)
+    sampler_b = EpisodeAwareSampler([0], [6], shuffle=True, seed=42)
+    epoch_0 = list(sampler_a)
+    assert list(sampler_b) == epoch_0
+    # Desyncing the global RNG must not affect the permutation.
+    sampler_c = EpisodeAwareSampler([0], [6], shuffle=True, seed=42)
+    torch.randperm(1000)  # consume global RNG, as rank-asymmetric code (e.g. eval) would
+    assert list(sampler_c) == epoch_0
+
+
 def test_negative_drop_first_frames_raises():
     with pytest.raises(ValueError, match="drop_n_first_frames must be >= 0"):
         EpisodeAwareSampler([0], [10], drop_n_first_frames=-1)
@@ -183,3 +196,85 @@ def test_weighted_sampler_zero_weights_fall_back_to_uniform():
 def test_weighted_sampler_rejects_short_weight_vector():
     with pytest.raises(ValueError, match="frame_weights"):
         WeightedEpisodeAwareSampler([0], [10], frame_weights=torch.ones(5))
+
+
+# --- seeded (seed, epoch) shuffling, resume, and state ---
+
+EPISODE_BOUNDS = ([0, 2, 3], [2, 3, 6])  # episodes of 2, 1 and 3 frames
+
+
+@pytest.mark.parametrize("num_frames", [1, 2, 3, 37, 64, 100])
+def test_deterministic_sampler_shuffle_is_permutation(num_frames):
+    for seed in (0, 1, 1234):
+        sampler = EpisodeAwareSampler([0], [num_frames], shuffle=True, seed=seed)
+        assert sorted(sampler) == list(range(num_frames))
+
+
+def test_deterministic_sampler_epochs_reproduce_and_differ():
+    sampler_a = EpisodeAwareSampler([0], [100], shuffle=True, seed=42)
+    sampler_b = EpisodeAwareSampler([0], [100], shuffle=True, seed=42)
+    epoch_0 = list(sampler_a)
+    assert list(sampler_b) == epoch_0  # same (seed, epoch) -> same order on any process
+    epoch_1 = list(sampler_a)  # __iter__ auto-advances the epoch
+    assert epoch_1 != epoch_0
+    assert sorted(epoch_1) == sorted(epoch_0)
+    sampler_a.set_epoch(0)
+    assert list(sampler_a) == epoch_0
+    assert list(EpisodeAwareSampler([0], [100], shuffle=True, seed=7)) != epoch_0
+
+
+def test_deterministic_sampler_resume_mid_epoch():
+    reference = EpisodeAwareSampler(*EPISODE_BOUNDS, shuffle=True, seed=42)
+    epoch_0 = list(reference)
+    epoch_1 = list(reference)
+    for start in (0, 1, 4, len(epoch_0)):
+        resumed = EpisodeAwareSampler(*EPISODE_BOUNDS, shuffle=True, seed=42)
+        resumed.load_state_dict({"epoch": 0, "start_index": start})
+        assert list(resumed) == epoch_0[start:]
+        # the resumed sampler continues into the same epoch 1 as the uninterrupted one
+        assert list(resumed) == epoch_1
+
+
+def test_deterministic_sampler_construction_stores_only_boundaries():
+    # Construction is O(num_episodes), not O(num_frames): a million-frame single episode
+    # instantiates from just its boundaries without materializing a per-frame index list.
+    num_frames = 1_000_000
+    sampler = EpisodeAwareSampler([0], [num_frames], shuffle=True, seed=0)
+    assert len(sampler) == num_frames
+    assert sampler._starts.shape == (1,) and sampler._cum_lengths.shape == (1,)
+
+
+def test_deterministic_sampler_resume_is_exact_at_scale():
+    # Seeded randperm makes resume sample-exact at non-trivial sizes: regenerating the epoch's
+    # permutation and slicing from the saved offset reproduces the remaining order exactly.
+    num_frames = 100_000
+    reference = EpisodeAwareSampler([0], [num_frames], shuffle=True, seed=0)
+    epoch_0 = list(reference)
+    assert sorted(epoch_0) == list(range(num_frames))
+    start = num_frames - 5
+    resumed = EpisodeAwareSampler([0], [num_frames], shuffle=True, seed=0)
+    resumed.load_state_dict({"epoch": 0, "start_index": start})
+    assert list(resumed) == epoch_0[start:]
+
+
+def test_compute_sampler_state():
+    # 100 frames, batch 10, 2 ranks -> 10 underlying batches, 5 per rank per epoch.
+    assert compute_sampler_state(step=0, num_frames=100, batch_size=10, num_processes=2) == {
+        "epoch": 0,
+        "start_index": 0,
+    }
+    # step 7 -> epoch 1, 2 per-rank batches in = 2 * 10 * 2 = 40 samples in
+    assert compute_sampler_state(step=7, num_frames=100, batch_size=10, num_processes=2) == {
+        "epoch": 1,
+        "start_index": 40,
+    }
+    # uneven epoch: 95 frames -> 10 underlying batches (last short), still 5 per rank
+    assert compute_sampler_state(step=12, num_frames=95, batch_size=10, num_processes=2) == {
+        "epoch": 2,
+        "start_index": 40,
+    }
+    # uneven sharding: 105 frames -> 11 underlying batches, 6 per rank (even_batches pads)
+    assert compute_sampler_state(step=11, num_frames=105, batch_size=10, num_processes=2) == {
+        "epoch": 1,
+        "start_index": 100,
+    }

tests/datasets/test_video_encoding.py

@@ -504,6 +504,19 @@ class TestReencodeVideo:
         assert info["video.g"] == 6
         assert info["video.crf"] == 23
 
+    @require_h264
+    def test_reencode_video_trim_window(self, tmp_path):
+        src = TEST_ARTIFACTS_DIR / "clip_6frames.mp4"
+        out = tmp_path / "trim_window.mp4"
+        cfg = VideoEncoderConfig(vcodec="h264")
+        reencode_video(src, out, camera_encoder=cfg, start_time_s=0.05, end_time_s=0.12, overwrite=True)
+
+        with av.open(str(out)) as container:
+            frames = list(container.decode(video=0))
+        # Only the frames at 0.067 and 0.1 s fall inside [0.05, 0.12).
+        assert len(frames) == 2
+        assert frames[0].time == pytest.approx(0.0, abs=1e-3)
+
 
 class TestConcatenateVideoFiles:
     def test_two_clips_frame_count(self, tmp_path):