tests(update): updating tests

Per review, log a warning when compute_episode_stats skips a feature with a
zero-width shape, so users know stats were intentionally not computed for it.

docs/source/hardware_guide.mdx

@@ -82,18 +82,18 @@ VRAM is the first filter. Within a tier, pick by budget and availability — the
 
 ### Hugging Face Jobs
 
-[Hugging Face Jobs](https://huggingface.co/docs/hub/jobs) lets you run training on managed HF infrastructure, billed by the second. The repo publishes a ready-to-use image: **`huggingface/lerobot-gpu:latest`**, rebuilt **every night at 02:00 UTC from `main`** ([`docker_publish.yml`](https://github.com/huggingface/lerobot/blob/main/.github/workflows/docker_publish.yml)) — so it tracks the current state of the repo, not a tagged release.
+[Hugging Face Jobs](https://huggingface.co/docs/hub/jobs) lets you run training on managed HF infrastructure, billed by the second, without owning a GPU. `lerobot-train` submits and streams the job for you — just add `--job.target=<flavor>` to a normal training command:
 
 ```bash
-hf jobs run --flavor a10g-large huggingface/lerobot-gpu:latest \
-  bash -c "nvidia-smi && lerobot-train \
-    --policy.type=act --dataset.repo_id=<USER>/<DATASET> \
-    --policy.repo_id=<USER>/act_<task> --batch_size=8 --steps=50000"
+lerobot-train \
+  --policy.type=act --dataset.repo_id=<USER>/<DATASET> \
+  --policy.repo_id=<USER>/act_<task> \
+  --job.target=a10g-large
 ```
 
 Notes:
 
-- The leading `nvidia-smi` is a quick sanity check that CUDA is visible inside the container — useful to fail fast if the flavor or driver mismatched.
-- The default Job timeout is 30 minutes; pass `--timeout 4h` (or longer) for real training.
-- `--flavor` maps onto the table above: `t4-small`/`t4-medium` (T4, ACT only), `l4x1`/`l4x4` (L4 24 GB), `a10g-small/large/largex2/largex4` (A10G 24 GB scaled out), `a100-large` (A100). For the current full catalogue + pricing see [https://huggingface.co/docs/hub/jobs](https://huggingface.co/docs/hub/jobs).
-- Prefer not to write the `hf jobs run` wrapper yourself? `lerobot-train` can submit the job for you: just add `--job.target=<flavor>` to a normal training command and it handles dataset upload, log streaming, and the final model push. See the [imitation-learning training guide](./il_robots).
+- Run `hf auth login` once before submitting, the job runs under your token.
+- `--job.target` maps onto the table above: `t4-small`/`t4-medium` (T4, ACT only), `l4x1`/`l4x4` (L4 24 GB), `a10g-small/large/largex2/largex4` (A10G 24 GB scaled out), `a100-large` (A100). List the current catalogue with pricing via `hf jobs hardware`, or see [https://huggingface.co/docs/hub/jobs](https://huggingface.co/docs/hub/jobs).
+- The job defaults to a `2d` (48h) timeout. Override it with `--job.timeout=4h` (or any other valid duration string) to shorten or extend the timeout. The job automatically stops when the command completes.
+- For the full walkthrough — dataset upload, checkpoint streaming, resuming a run on a job — see the [imitation-learning training guide](./il_robots#train-using-hugging-face-jobs).

docs/source/il_robots.mdx

@@ -532,84 +532,7 @@ If your local computer doesn't have a powerful GPU you could utilize Google Cola
 
 Hugging Face jobs let's you easily select hardware and run the training in the cloud. So if you don't have a powerful GPU or you need more VRAM or just want to train a model much faster use HF Jobs! It's pay as you go and you simply pay for each second of use, you can see the pricing and additional information [here](https://huggingface.co/docs/hub/jobs).
 
-> **Tip:** if you just want to launch a standard training run, you can skip building the command below and use the integrated **Train on HF Jobs via `--job.target`** flow described further down — `lerobot-train` then submits the job, uploads a local-only dataset for you, and streams the logs.
-
-To run the training manually use this command:
-
-<hfoptions id="train_with_hf_jobs">
-<hfoption id="Command">
-```bash
-hf jobs run \
-  --flavor a10g-small \
-  --timeout 4h \
-  --secrets HF_TOKEN \
-  huggingface/lerobot-gpu:latest \
-  -- \
-  python -m lerobot.scripts.lerobot_train \
-    --dataset.repo_id=username/dataset \
-    --policy.type=act \
-    --steps=5000 \
-    --batch_size=16 \
-    --policy.device=cuda \
-    --policy.repo_id=username/your_policy \
-    --log_freq=100
-```
-</hfoption>
-<hfoption id="API example">
-
-<!-- prettier-ignore-start -->
-```python
-from huggingface_hub import run_job, get_token
-
-run_name = "act_so101_hf_jobs"
-dataset_id = "username/dataset"
-user_hub_id = "username"
-
-command_args = [
-    "python", "-m", "lerobot.scripts.lerobot_train",
-    "--dataset.repo_id", dataset_id,
-    "--policy.type", "act",
-    "--steps", "5000",
-    "--batch_size", "16",
-    "--num_workers", "4",
-    "--policy.device", "cuda",
-    "--log_freq", "100",
-    "--save_freq", "1000",
-    "--save_checkpoint", "true",
-    "--wandb.enable", "false",
-    "--policy.repo_id", f"{user_hub_id}/{run_name}"
-]
-
-print(f"Submitting job '{run_name}' to Hugging Face Infrastructure...")
-
-job_info = run_job(
-    image="huggingface/lerobot-gpu:latest",
-    command=command_args,
-    flavor="a10g-small",
-    timeout="4h",
-    secrets={"HF_TOKEN": get_token()}
-)
-
-print("\n🚀 Job successfully launched!")
-print(f"🔹 Job ID: {job_info.id}")
-print(f"🔗 Live UI Dashboard & Logs: {job_info.url}")
-```
-<!-- prettier-ignore-end -->
-
-</hfoption>
-</hfoptions>
-
-You can modify the `--flavor` to use different hardware, for example: `t4-small`, `a100-large`, `h200`. Use `hf jobs hardware` to see the full list with pricing.
-Depending on the model you want to train and the hardware you selected you can also modify the `--batch_size` and `--number_of_workers`.
-For longer training sessions increase the timeout.
-
-Once the training is started you can go to [Jobs](https://huggingface.co/settings/jobs) and see if your jobs is running as well as all the outputs. Sometimes it takes a few minutes to schedule your job so be patient.
-
-After training the model will be pushed to hub and you can use it as any other model with LeRobot.
-
-#### Train on HF Jobs via `--job.target` (integrated CLI)
-
-`lerobot-train` runs locally by default. To run on a HuggingFace GPU without constructing the Docker command yourself, pass `--job.target` with a hardware flavor name:
+`lerobot-train` runs locally by default. To run on a HuggingFace GPU, pass `--job.target` with a hardware flavor name:
 
 ```bash
 lerobot-train \

src/lerobot/configs/__init__.py

@@ -34,6 +34,8 @@ from .types import (
 )
 from .video import (
     DEFAULT_DEPTH_UNIT,
+    DEPTH_METER_UNIT,
+    DEPTH_MILLIMETER_UNIT,
     VALID_VIDEO_CODECS,
     VIDEO_ENCODER_INFO_KEYS,
     DepthEncoderConfig,
@@ -41,6 +43,7 @@ from .video import (
     VideoEncoderConfig,
     depth_encoder_defaults,
     encoder_config_from_video_info,
+    infer_depth_unit,
     rgb_encoder_defaults,
 )
 
@@ -70,8 +73,11 @@ __all__ = [
     "depth_encoder_defaults",
     # Factories
     "encoder_config_from_video_info",
+    "infer_depth_unit",
     # Constants
     "DEFAULT_DEPTH_UNIT",
+    "DEPTH_METER_UNIT",
+    "DEPTH_MILLIMETER_UNIT",
     "VALID_VIDEO_CODECS",
     "VIDEO_ENCODER_INFO_KEYS",
 ]

src/lerobot/configs/video.py

@@ -22,6 +22,8 @@ import logging
 from dataclasses import dataclass, field
 from typing import Any, ClassVar, Self
 
+import numpy as np
+
 from lerobot.utils.import_utils import require_package
 
 logger = logging.getLogger(__name__)
@@ -36,7 +38,9 @@ HW_VIDEO_CODECS = [
     "h264_vaapi",  # Linux Intel/AMD
     "h264_qsv",  # Intel Quick Sync
 ]
-VALID_VIDEO_CODECS: frozenset[str] = frozenset({"h264", "hevc", "libsvtav1", "auto", *HW_VIDEO_CODECS})
+VALID_VIDEO_CODECS: frozenset[str] = frozenset(
+    {"h264", "hevc", "libsvtav1", "libaom-av1", "auto", *HW_VIDEO_CODECS}
+)
 # Aliases for legacy video codec names.
 VIDEO_CODECS_ALIASES: dict[str, str] = {"av1": "libsvtav1"}
 
@@ -65,6 +69,15 @@ DEPTH_METER_UNIT: str = "m"
 DEPTH_MILLIMETER_UNIT: str = "mm"
 DEFAULT_DEPTH_UNIT: str = DEPTH_MILLIMETER_UNIT
 
+
+def infer_depth_unit(dtype: np.dtype | type) -> str:
+    """Infer the physical unit of raw depth frames from their dtype.
+
+    Floating-point frames are assumed to be in metres, integer frames in millimetres.
+    """
+    return DEPTH_METER_UNIT if np.issubdtype(np.dtype(dtype), np.floating) else DEPTH_MILLIMETER_UNIT
+
+
 # Depth-specific tuning fields persisted under ``features[*]["info"]`` as ``video.<name>``.
 DEPTH_ENCODER_INFO_FIELD_NAMES: frozenset[str] = frozenset({"depth_min", "depth_max", "shift", "use_log"})
 
@@ -213,18 +226,24 @@ class VideoEncoderConfig:
             if encoder_threads is not None:
                 svtav1_parts.append(f"lp={encoder_threads}")
             if svtav1_parts:
-                opts["svtav1-params"] = ":".join(svtav1_parts)
+                set_if("svtav1-params", ":".join(svtav1_parts))
         elif self.vcodec in ("h264", "hevc"):
             set_if("crf", self.crf)
             set_if("preset", self.preset)
             if self.fast_decode:
-                opts["tune"] = "fastdecode"
+                set_if("tune", "fastdecode")
             set_if("threads", encoder_threads)
+        elif self.vcodec == "libaom-av1":
+            set_if("crf", self.crf)
+            set_if("preset", self.preset)
+            if encoder_threads is not None:
+                set_if("threads", encoder_threads)
+                set_if("row-mt", 1)
         elif self.vcodec in ("h264_videotoolbox", "hevc_videotoolbox"):
             if self.crf is not None:
-                opts["q:v"] = max(1, min(100, 100 - self.crf * 2))
+                set_if("q:v", max(1, min(100, 100 - self.crf * 2)))
         elif self.vcodec in ("h264_nvenc", "hevc_nvenc"):
-            opts["rc"] = 0
+            set_if("rc", 0)
             set_if("qp", self.crf)
             set_if("preset", self.preset)
         elif self.vcodec == "h264_vaapi":

src/lerobot/datasets/compute_stats.py

@@ -509,7 +509,7 @@ def compute_episode_stats(
         For 'image'/'video' features, stats are computed per channel and kept with a
         leading channel axis (e.g. shape (3, 1, 1) for RGB). RGB stats are divided by
         255 to land in [0, 1]; depth maps (features flagged with ``is_depth_map``) skip
-        this rescaling and remain in their stored units.
+        this rescaling and remain in their stored units (stored in ``depth_unit``).
     """
     if quantile_list is None:
         quantile_list = DEFAULT_QUANTILES
@@ -519,6 +519,13 @@ def compute_episode_stats(
         if features[key]["dtype"] in {"string", "language"}:
             continue
 
+        # Features with zero-width shapes are skipped (no data to compute stats on)
+        if any(d == 0 for d in features[key].get("shape", ())):
+            logging.debug(
+                f"Skipping statistics computation for feature '{key}' with a zero-width shape {features[key]['shape']}."
+            )
+            continue
+
         if features[key]["dtype"] in ["image", "video"]:
             ep_ft_array = sample_images(data)
             axes_to_reduce = (0, 2, 3)

src/lerobot/datasets/dataset_metadata.py

@@ -26,12 +26,13 @@ import pyarrow as pa
 import pyarrow.parquet as pq
 from huggingface_hub import snapshot_download
 
-from lerobot.configs import VideoEncoderConfig
+from lerobot.configs import DEPTH_METER_UNIT, VideoEncoderConfig
 from lerobot.utils.constants import DEFAULT_FEATURES, HF_LEROBOT_HOME, HF_LEROBOT_HUB_CACHE
 from lerobot.utils.feature_utils import _validate_feature_names
 from lerobot.utils.utils import flatten_dict
 
 from .compute_stats import aggregate_stats
+from .depth_utils import MM_PER_METRE
 from .feature_utils import create_empty_dataset_info
 from .io_utils import (
     get_file_size_in_mb,
@@ -358,6 +359,35 @@ class LeRobotDatasetMetadata:
 
         return [key for key, ft in self.features.items() if _is_depth(ft)]
 
+    def rescale_depth_stats(self, output_unit: str) -> None:
+        """Rescale depth feature stats in place from their recorded unit to ``output_unit``.
+
+        Depth stats are stored in the unit the frames were recorded in
+        (``features[key]["info"]["depth_unit"]``), while frames are returned in
+        ``output_unit`` on read. This converts the unit-bearing stat entries so
+        stats match the frames consumers see.
+        """
+        missing_unit_keys = [
+            key for key in self.depth_keys if (self.features[key].get("info") or {}).get("depth_unit") is None
+        ]
+        if missing_unit_keys:
+            logging.warning(
+                f"Depth feature(s) {missing_unit_keys} have no recorded 'depth_unit' in their info. "
+                f"Depth maps and stats for these keys will be returned AS IS, with no unit conversion "
+                f"to the requested output unit {output_unit!r}. Re-record the dataset or set 'depth_unit' "
+                f"in the feature info (meta/info.json) to enable conversion."
+            )
+        if self.stats is None:
+            return
+        for key in self.depth_keys:
+            stored_unit = (self.features[key].get("info") or {}).get("depth_unit")
+            if stored_unit is None or stored_unit == output_unit or key not in self.stats:
+                continue
+            factor = MM_PER_METRE if stored_unit == DEPTH_METER_UNIT else 1.0 / MM_PER_METRE
+            self.stats[key] = {
+                stat: value if stat == "count" else value * factor for stat, value in self.stats[key].items()
+            }
+
     @property
     def camera_keys(self) -> list[str]:
         """Keys to access visual modalities (regardless of their storage method)."""

src/lerobot/datasets/dataset_reader.py

@@ -22,10 +22,14 @@ from pathlib import Path
 import datasets
 import torch
 
-from lerobot.configs import DEFAULT_DEPTH_UNIT, DepthEncoderConfig
+from lerobot.configs import (
+    DEFAULT_DEPTH_UNIT,
+    DEPTH_METER_UNIT,
+    DepthEncoderConfig,
+)
 
 from .dataset_metadata import LeRobotDatasetMetadata
-from .depth_utils import dequantize_depth
+from .depth_utils import MM_PER_METRE, dequantize_depth
 from .feature_utils import (
     check_delta_timestamps,
     get_delta_indices,
@@ -102,6 +106,13 @@ class DatasetReader:
             for vid_key in self._meta.depth_keys
         }
 
+        # Get the input unit of each depth feature stored as raw images.
+        self._image_depth_units: dict[str, str | None] = {
+            key: (self._meta.features[key].get("info") or {}).get("depth_unit")
+            for key in self._meta.depth_keys
+            if key in self._meta.image_keys
+        }
+
     def set_image_transforms(self, image_transforms: Callable | None) -> None:
         """Replace the transform applied to visual observations."""
         if image_transforms is not None and not callable(image_transforms):
@@ -329,6 +340,13 @@ class DatasetReader:
                     continue
                 item[cam] = self._image_transforms(item[cam])
 
+        # Convert depth features to the output unit.
+        for key, stored_unit in self._image_depth_units.items():
+            if key in item and stored_unit is not None and stored_unit != self._depth_output_unit:
+                item[key] = (
+                    item[key] * MM_PER_METRE if stored_unit == DEPTH_METER_UNIT else item[key] / MM_PER_METRE
+                )
+
         # Add task as a string
         task_idx = item["task_index"].item()
         item["task"] = self._meta.tasks.iloc[task_idx].name

src/lerobot/datasets/dataset_writer.py

@@ -36,6 +36,7 @@ from lerobot.configs import (
     RGBEncoderConfig,
     VideoEncoderConfig,
     depth_encoder_defaults,
+    infer_depth_unit,
     rgb_encoder_defaults,
 )
 
@@ -209,6 +210,15 @@ class DatasetWriter:
         self.episode_buffer["timestamp"].append(timestamp)
         self.episode_buffer["task"].append(frame.pop("task"))
 
+        # Record each depth feature's input unit once, inferred from the first frame's dtype.
+        if frame_index == 0:
+            for depth_key in self._meta.depth_keys:
+                if depth_key not in frame:
+                    continue
+                info = self._meta.features[depth_key].setdefault("info", {})
+                if info.get("depth_unit") is None:
+                    info["depth_unit"] = infer_depth_unit(np.asarray(frame[depth_key]).dtype)
+
         # Start streaming encoder on first frame of episode
         if frame_index == 0 and self._streaming_encoder is not None:
             self._streaming_encoder.start_episode(

src/lerobot/datasets/depth_utils.py

@@ -34,12 +34,13 @@ from lerobot.configs.video import (
     DEPTH_METER_UNIT,
     DEPTH_MILLIMETER_UNIT,
     DEPTH_QMAX,
+    infer_depth_unit,
 )
 
 from .image_writer import squeeze_single_channel
 from .pyav_utils import write_u16_plane
 
-_MM_PER_METRE = 1000.0
+MM_PER_METRE = 1000.0
 _UINT16_MAX = 65535
 
 
@@ -57,11 +58,7 @@ def _depth_input_to_float32_and_unit(
     input_unit: Literal["auto", DEPTH_METER_UNIT, DEPTH_MILLIMETER_UNIT],
 ) -> tuple[NDArray[np.float32], Literal[DEPTH_METER_UNIT, DEPTH_MILLIMETER_UNIT]]:
     """Convert depth to float32 in the chosen unit, and return the resolved unit."""
-    resolved_unit = (
-        (DEPTH_METER_UNIT if np.issubdtype(depth.dtype, np.floating) else DEPTH_MILLIMETER_UNIT)
-        if input_unit == "auto"
-        else input_unit
-    )
+    resolved_unit = infer_depth_unit(depth.dtype) if input_unit == "auto" else input_unit
     return depth.astype(np.float32, order="K"), resolved_unit
 
 
@@ -126,12 +123,12 @@ def quantize_depth(
 
     # Convert depth_min, depth_max, and shift to the resolved input unit.
     depth_min_u = (
-        np.float32(depth_min) if resolved_unit == DEPTH_METER_UNIT else np.float32(depth_min * _MM_PER_METRE)
+        np.float32(depth_min) if resolved_unit == DEPTH_METER_UNIT else np.float32(depth_min * MM_PER_METRE)
     )
     depth_max_u = (
-        np.float32(depth_max) if resolved_unit == DEPTH_METER_UNIT else np.float32(depth_max * _MM_PER_METRE)
+        np.float32(depth_max) if resolved_unit == DEPTH_METER_UNIT else np.float32(depth_max * MM_PER_METRE)
     )
-    shift_u = np.float32(shift) if resolved_unit == DEPTH_METER_UNIT else np.float32(shift * _MM_PER_METRE)
+    shift_u = np.float32(shift) if resolved_unit == DEPTH_METER_UNIT else np.float32(shift * MM_PER_METRE)
 
     # Normalization and quantization is performed in the resolved input unit.
     if use_log:
@@ -236,7 +233,7 @@ def dequantize_depth(
 
         # mm path: round + clamp in float32, skipping the uint16 round-trip
         # when returning a tensor (torch.uint16 is poorly supported).
-        buf.mul_(_MM_PER_METRE).round_().clamp_(0.0, _UINT16_MAX)
+        buf.mul_(MM_PER_METRE).round_().clamp_(0.0, _UINT16_MAX)
         if output_tensor:
             return buf
         return buf.cpu().numpy().astype(np.uint16, copy=False)
@@ -259,7 +256,7 @@ def dequantize_depth(
     if output_unit == DEPTH_METER_UNIT:
         return torch.from_numpy(buf) if output_tensor else buf
 
-    np.multiply(buf, _MM_PER_METRE, out=buf)
+    np.multiply(buf, MM_PER_METRE, out=buf)
     np.rint(buf, out=buf)
     np.clip(buf, 0.0, _UINT16_MAX, out=buf)
     if output_tensor:

src/lerobot/datasets/feature_utils.py

@@ -67,9 +67,9 @@ def get_hf_features_from_features(features: dict) -> datasets.Features:
         elif ft["shape"] == (1,):
             hf_features[key] = datasets.Value(dtype=ft["dtype"])
         elif len(ft["shape"]) == 1:
-            hf_features[key] = datasets.Sequence(
-                length=ft["shape"][0], feature=datasets.Value(dtype=ft["dtype"])
-            )
+            # pyarrow rejects fixed-size lists of length 0, so use a variable length list instead
+            length = ft["shape"][0] if ft["shape"][0] > 0 else -1
+            hf_features[key] = datasets.Sequence(length=length, feature=datasets.Value(dtype=ft["dtype"]))
         elif len(ft["shape"]) == 2:
             hf_features[key] = datasets.Array2D(shape=ft["shape"], dtype=ft["dtype"])
         elif len(ft["shape"]) == 3:

src/lerobot/datasets/lerobot_dataset.py

@@ -224,6 +224,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
         )
         self.root = self.meta.root
         self.revision = self.meta.revision
+        self.meta.rescale_depth_stats(self._depth_output_unit)
 
         if episodes is not None and any(
             episode >= self.meta.total_episodes or episode < 0 for episode in episodes
@@ -350,6 +351,11 @@ class LeRobotDataset(torch.utils.data.Dataset):
         """Frames per second used during data collection."""
         return self.meta.fps
 
+    @property
+    def depth_output_unit(self) -> str:
+        """Physical unit (``"m"`` or ``"mm"``) depth maps and statistics are returned in on read."""
+        return self._depth_output_unit
+
     @property
     def num_frames(self) -> int:
         """Number of frames in selected episodes."""

src/lerobot/datasets/streaming_dataset.py

@@ -22,11 +22,11 @@ import numpy as np
 import torch
 from datasets import load_dataset
 
-from lerobot.configs import DEFAULT_DEPTH_UNIT, DepthEncoderConfig
+from lerobot.configs import DEFAULT_DEPTH_UNIT, DEPTH_METER_UNIT, DepthEncoderConfig
 from lerobot.utils.constants import HF_LEROBOT_HOME, LOOKAHEAD_BACKTRACKTABLE, LOOKBACK_BACKTRACKTABLE
 
 from .dataset_metadata import CODEBASE_VERSION, LeRobotDatasetMetadata
-from .depth_utils import dequantize_depth
+from .depth_utils import MM_PER_METRE, dequantize_depth
 from .feature_utils import get_delta_indices
 from .io_utils import item_to_torch
 from .utils import (
@@ -310,6 +310,7 @@ class StreamingLeRobotDataset(torch.utils.data.IterableDataset):
         )
         self.root = self.meta.root
         self.revision = self.meta.revision
+        self.meta.rescale_depth_stats(self._depth_output_unit)
         # Check version
         check_version_compatibility(self.repo_id, self.meta._version, CODEBASE_VERSION)
 
@@ -318,6 +319,13 @@ class StreamingLeRobotDataset(torch.utils.data.IterableDataset):
             for vid_key in self.meta.depth_keys
         }
 
+        # Input unit of each depth feature stored as raw images (dequantized separately from videos).
+        self._image_depth_units: dict[str, str | None] = {
+            key: (self.meta.features[key].get("info") or {}).get("depth_unit")
+            for key in self.meta.depth_keys
+            if key in self.meta.image_keys
+        }
+
         self.delta_timestamps = None
         self.delta_indices = None
 
@@ -348,6 +356,11 @@ class StreamingLeRobotDataset(torch.utils.data.IterableDataset):
     def fps(self):
         return self.meta.fps
 
+    @property
+    def depth_output_unit(self) -> str:
+        """Physical unit (``"m"`` or ``"mm"``) depth maps are returned in on read."""
+        return self._depth_output_unit
+
     @staticmethod
     def _iter_random_indices(
         rng: np.random.Generator, buffer_size: int, random_batch_size=100
@@ -530,6 +543,15 @@ class StreamingLeRobotDataset(torch.utils.data.IterableDataset):
         for update in updates:
             result.update(update)
 
+        # Convert raw-image depth features to the output unit (video depth is already converted).
+        for key, stored_unit in self._image_depth_units.items():
+            if key in result and stored_unit is not None and stored_unit != self._depth_output_unit:
+                result[key] = (
+                    result[key] * MM_PER_METRE
+                    if stored_unit == DEPTH_METER_UNIT
+                    else result[key] / MM_PER_METRE
+                )
+
         result["task"] = self.meta.tasks.iloc[item["task_index"]].name
 
         yield result

src/lerobot/scripts/lerobot_dataset_viz.py

@@ -84,6 +84,7 @@ import torch
 import torch.utils.data
 import tqdm
 
+from lerobot.configs import DEPTH_MILLIMETER_UNIT
 from lerobot.datasets import LeRobotDataset
 from lerobot.utils.constants import ACTION, DONE, OBS_STATE, REWARD, SUCCESS
 from lerobot.utils.utils import init_logging
@@ -228,6 +229,9 @@ def visualize_dataset(
 
     logging.info("Logging to Rerun")
 
+    # Depth frames and stats are dequantized to the dataset's depth_output_unit on load.
+    depth_meter = 1000.0 if dataset.depth_output_unit == DEPTH_MILLIMETER_UNIT else 1.0
+
     # Use the dataset's q01/q99 depth statistics for robust depth range bounds
     depth_ranges = {}
     for key in dataset.meta.depth_keys:
@@ -254,6 +258,7 @@ def visualize_dataset(
                     depth = to_hwc_float32_numpy(batch[key][i])
                     depth_entity = rr.DepthImage(
                         depth,
+                        meter=depth_meter,
                         colormap=rr.components.Colormap.Viridis,
                         depth_range=depth_ranges.get(key),
                     )

src/lerobot/scripts/lerobot_train.py

@@ -211,8 +211,12 @@ def train(cfg: TrainPipelineConfig, accelerator: "Accelerator | None" = None):
         # Accelerate auto-detects the device based on the available hardware and ignores the policy.device setting.
         # Force the device to be CPU when the active config's device is set to CPU (works for both policy and reward model training).
         force_cpu = cfg.trainable_config.device == "cpu"
+        # Drive Accelerate's autocast from policy.dtype (bf16/fp16 activate it; float32/absent -> launcher default).
+        policy_dtype = getattr(cfg.trainable_config, "dtype", None)
+        mixed_precision = {"bfloat16": "bf16", "float16": "fp16", "float32": "no"}.get(policy_dtype)
         accelerator = Accelerator(
             step_scheduler_with_optimizer=False,
+            mixed_precision=mixed_precision,
             kwargs_handlers=[ddp_kwargs],
             cpu=force_cpu,
         )

src/lerobot/utils/rerun_visualization.py

@@ -24,6 +24,7 @@ import os
 
 import numpy as np
 
+from lerobot.configs import DEPTH_MILLIMETER_UNIT, infer_depth_unit
 from lerobot.types import RobotAction, RobotObservation
 
 from .constants import ACTION, ACTION_PREFIX, OBS_PREFIX, OBS_STR
@@ -161,7 +162,13 @@ def log_rerun_data(
                     observation_paths.add(key)
                 else:
                     if arr.shape[-1] == 1:
-                        img_entity = rr.DepthImage(arr, colormap=rr.components.Colormap.Viridis)
+                        # At record time, the depth unit is inferred from the frame type.
+                        depth_unit = infer_depth_unit(arr.dtype)
+                        img_entity = rr.DepthImage(
+                            arr,
+                            meter=1000.0 if depth_unit == DEPTH_MILLIMETER_UNIT else 1.0,
+                            colormap=rr.components.Colormap.Viridis,
+                        )
                     else:
                         img_entity = rr.Image(arr).compress() if compress_images else rr.Image(arr)
                     rr.log(key, entity=img_entity, static=True)

tests/datasets/test_compute_stats.py

@@ -13,6 +13,7 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+import logging
 from unittest.mock import patch
 
 import numpy as np
@@ -687,6 +688,28 @@ def test_compute_episode_stats_string_features_skipped():
     assert "q01" in stats["action"]
 
 
+def test_compute_episode_stats_zero_width_features_skipped(caplog):
+    """Test that features with a zero-width dim (e.g. shape=(0,)) are skipped with a debug log."""
+    episode_data = {
+        "empty": np.zeros((100, 0), dtype=np.float32),  # Zero-width feature
+        "action": np.random.normal(0, 1, (100, 5)),
+    }
+    features = {
+        "empty": {"dtype": "float32", "shape": (0,)},
+        "action": {"dtype": "float32", "shape": (5,)},
+    }
+
+    with caplog.at_level(logging.DEBUG):
+        stats = compute_episode_stats(episode_data, features)
+
+    # Zero-width features should be skipped with a debug log, others computed as usual
+    assert "empty" not in stats
+    assert "empty" in caplog.text
+    assert "action" in stats
+    assert "q01" in stats["action"]
+    assert stats["action"]["mean"].shape == (5,)
+
+
 def test_aggregate_feature_stats_with_quantiles():
     """Test aggregating feature stats that include quantiles."""
     stats_ft_list = [

tests/datasets/test_datasets.py

@@ -1531,6 +1531,7 @@ def test_valid_video_codecs_constant():
     assert "h264" in VALID_VIDEO_CODECS
     assert "hevc" in VALID_VIDEO_CODECS
     assert "libsvtav1" in VALID_VIDEO_CODECS
+    assert "libaom-av1" in VALID_VIDEO_CODECS
     assert "auto" in VALID_VIDEO_CODECS
     assert "h264_videotoolbox" in VALID_VIDEO_CODECS
     assert "h264_nvenc" in VALID_VIDEO_CODECS
@@ -1538,7 +1539,7 @@ def test_valid_video_codecs_constant():
     assert "h264_qsv" in VALID_VIDEO_CODECS
     assert "hevc_videotoolbox" in VALID_VIDEO_CODECS
     assert "hevc_nvenc" in VALID_VIDEO_CODECS
-    assert len(VALID_VIDEO_CODECS) == 10
+    assert len(VALID_VIDEO_CODECS) == 11
 
 
 def test_delta_timestamps_with_episodes_filter(tmp_path, empty_lerobot_dataset_factory):
@@ -1803,3 +1804,11 @@ def test_episode_filter_unknown_key_raises(tmp_path, lerobot_dataset_factory):
             root=dataset.root,
             episode_filter=lambda ep: ep["not_a_real_field"] > 0,
         )
+
+
+def test_get_hf_features_zero_width_feature_does_not_raise_on_from_dict():
+    import datasets
+
+    features = {"empty": {"dtype": "float32", "shape": (0,), "names": ["empty"]}}
+    hf_features = get_hf_features_from_features(features)
+    datasets.Dataset.from_dict({"empty": [[], []]}, features=hf_features)

tests/datasets/test_depth.py

@@ -32,6 +32,7 @@ from lerobot.configs.video import (
 )
 from lerobot.datasets.depth_utils import dequantize_depth, quantize_depth
 from lerobot.datasets.image_writer import image_array_to_pil_image, write_image
+from lerobot.utils.constants import DEFAULT_FEATURES
 from tests.fixtures.constants import (
     DEFAULT_FPS,
     DUMMY_CAMERA_FEATURES,
@@ -245,3 +246,91 @@ class TestFeatureFileRouting:
 
         dataset.save_episode()
         dataset.finalize()
+
+
+class TestDepthUnitMetadata:
+    """The depth unit is inferred once from dtype, stored in ``info``, and drives stats + reads."""
+
+    NUM_FRAMES = 4
+
+    def _record(self, root, features_factory, depth_dtype, value, use_videos):
+        from lerobot.datasets.lerobot_dataset import LeRobotDataset
+
+        features = features_factory(camera_features=DUMMY_CAMERA_FEATURES_WITH_DEPTH, use_videos=use_videos)
+        dataset = LeRobotDataset.create(
+            repo_id=DUMMY_REPO_ID,
+            fps=DEFAULT_FPS,
+            features=features,
+            root=root,
+            use_videos=use_videos,
+            streaming_encoding=use_videos,
+        )
+        for _ in range(self.NUM_FRAMES):
+            frame: dict = {"task": "test"}
+            for key, ft in dataset.meta.features.items():
+                if key in DEFAULT_FEATURES:
+                    continue
+                if key in dataset.meta.depth_keys:
+                    frame[key] = np.full(ft["shape"], value, dtype=depth_dtype)
+                elif key in dataset.meta.camera_keys:
+                    frame[key] = np.random.randint(0, 256, ft["shape"], dtype=np.uint8)
+                else:
+                    frame[key] = np.zeros(ft["shape"], dtype=np.float32)
+            dataset.add_frame(frame)
+        return dataset
+
+    @pytest.mark.parametrize("use_videos", [False, True])
+    @pytest.mark.parametrize(
+        ("depth_dtype", "value", "expected_unit"),
+        [(np.float32, 2.0, DEPTH_METER_UNIT), (np.uint16, 2000, DEPTH_MILLIMETER_UNIT)],
+    )
+    def test_recorded_unit_inferred_persisted_and_kept_in_stats(
+        self, tmp_path, features_factory, use_videos, depth_dtype, value, expected_unit
+    ):
+        """Unit is inferred from the first frame's dtype, drives stats (raw, never canonicalized), and survives a reload."""
+        from lerobot.datasets.lerobot_dataset import LeRobotDataset
+
+        dataset = self._record(tmp_path / "ds", features_factory, depth_dtype, value, use_videos)
+        assert dataset.meta.features[DEPTH_KEY]["info"]["depth_unit"] == expected_unit
+        dataset.save_episode()
+        mean = float(np.asarray(dataset.meta.stats[DEPTH_KEY]["mean"]).reshape(-1)[0])
+        np.testing.assert_allclose(mean, value, rtol=0.05)
+        dataset.finalize()
+
+        reloaded = LeRobotDataset(repo_id=DUMMY_REPO_ID, root=tmp_path / "ds")
+        assert reloaded.meta.features[DEPTH_KEY]["info"]["depth_unit"] == expected_unit
+
+    @pytest.mark.parametrize("use_videos", [False, True])
+    @pytest.mark.parametrize(
+        ("output_unit", "expected"),
+        [(DEPTH_MILLIMETER_UNIT, 2000.0), (DEPTH_METER_UNIT, 2.0)],
+    )
+    def test_read_honors_output_unit_for_frames_and_stats(
+        self, tmp_path, features_factory, use_videos, output_unit, expected
+    ):
+        """Reloading with a ``depth_output_unit`` converts metre frames (image mode) and rescales stats while preserving count."""
+        from lerobot.datasets.lerobot_dataset import LeRobotDataset
+
+        dataset = self._record(tmp_path / "ds", features_factory, np.float32, 2.0, use_videos=use_videos)
+        dataset.save_episode()
+        count = float(np.asarray(dataset.meta.stats[DEPTH_KEY]["count"]).reshape(-1)[0])
+        dataset.finalize()
+
+        read_dataset = LeRobotDataset(
+            repo_id=DUMMY_REPO_ID, root=tmp_path / "ds", depth_output_unit=output_unit
+        )
+        stats = read_dataset.meta.stats[DEPTH_KEY]
+        np.testing.assert_allclose(float(np.asarray(stats["mean"]).reshape(-1)[0]), expected, rtol=0.05)
+        np.testing.assert_allclose(float(np.asarray(stats["count"]).reshape(-1)[0]), count)
+
+        if not use_videos:
+            depth = read_dataset[0][DEPTH_KEY]
+            assert torch.allclose(depth, torch.full_like(depth, expected))
+
+            from lerobot.datasets.streaming_dataset import StreamingLeRobotDataset
+
+            stream_dataset = StreamingLeRobotDataset(
+                repo_id=DUMMY_REPO_ID, root=tmp_path / "ds", depth_output_unit=output_unit
+            )
+            stream_depth = next(iter(stream_dataset))[DEPTH_KEY]
+            assert torch.allclose(stream_depth, torch.full_like(stream_depth, expected))

tests/datasets/test_video_encoding.py

@@ -345,7 +345,9 @@ class TestExtraOptions:
         opts = cfg.get_codec_options()
         assert opts["qp"] == 20
         assert isinstance(opts["qp"], int)
-        assert cfg.get_codec_options(as_strings=True)["qp"] == "20"
+        str_opts = cfg.get_codec_options(as_strings=True)
+        assert str_opts["qp"] == "20"
+        assert all(isinstance(v, str) for v in str_opts.values())
 
     @require_libsvtav1
     def test_structured_fields_win_on_collision(self):

tests/fixtures/dataset_factories.py

@@ -26,6 +26,7 @@ import pytest
 import torch
 from datasets import Dataset
 
+from lerobot.configs.video import infer_depth_unit
 from lerobot.datasets.dataset_metadata import CODEBASE_VERSION, LeRobotDatasetMetadata
 from lerobot.datasets.feature_utils import get_hf_features_from_features
 from lerobot.datasets.io_utils import flatten_dict, hf_transform_to_torch
@@ -535,6 +536,13 @@ def lerobot_dataset_factory(
                 chunks_size=chunks_size,
                 **info_kwargs,
             )
+            # This synthetic path skips add_frame, so record the depth unit the writer would
+            # have stored (dummy depth is uint16) to keep ``depth_unit`` present in info.json.
+            # Reassign a fresh info dict to avoid mutating the shared feature constants.
+            for ft in info.features.values():
+                ft_info = ft.get("info")
+                if ft_info is not None and ft_info.get("is_depth_map") and "depth_unit" not in ft_info:
+                    ft["info"] = {**ft_info, "depth_unit": infer_depth_unit(np.uint16)}
         if stats is None:
             stats = stats_factory(features=info.features)
         if tasks is None:

tests/utils/test_rerun_visualization.py

@@ -50,8 +50,9 @@ def mock_rerun(monkeypatch):
             return self
 
     class DummyDepthImage:
-        def __init__(self, arr, colormap=None):
+        def __init__(self, arr, meter=None, colormap=None):
             self.arr = arr
+            self.meter = meter
             self.colormap = colormap
 
     def dummy_log(key, obj=None, **kwargs):