feat(dataset): add streaming video encoding + HW encoder support (#2974)

* feat(dataset): init stream encoding

* feat(dataset): use threads to fix frame pickle latency

* refactor(dataset): remove HW encoded related changes

* add lp (#2977)

* feat(dataset): add Hw encoding + log drop frames (#2978)

* chore(docs): add streaming video encoding guide

* fix(dataset): style docs + testing

* chore(docs): simplify sttreaming video encoding guide

* chore(dataset): add commands + streaming encoding default false + print note if false + queue default is now 30

* chore(docs): add verification note advice

* chore(dataset): adjusting defaults & docs for streaming encoding

* docs(scripts): improve docstrings

* test(dataset): polish streaming encoding tests

* chore(dataset): move FYI log related to streaming

* chore(dataset): add arg vcodec to suggestions

* refactor(dataset): better handling for auto and available vcodec

* chore(dataset): change log level

* docs(dataset): add note related to training performance vcodec

* docs(dataset): add more notes to streaming encoding

---------

Co-authored-by: Caroline Pascal <caroline8.pascal@gmail.com>
Co-authored-by: Pepijn <pepijn@huggingface.co>
This commit is contained in:
Steven Palma
2026-02-23 13:57:43 +01:00
committed by GitHub
parent 5865170d36
commit e96339a3b4
16 changed files with 1532 additions and 55 deletions
+30 -4
View File
@@ -26,8 +26,10 @@ lerobot-record \
--dataset.repo_id=<my_username>/<my_dataset_name> \
--dataset.num_episodes=2 \
--dataset.single_task="Grab the cube" \
--dataset.streaming_encoding=true \
--dataset.encoder_threads=2 \
--display_data=true
# <- Optional: specify video codec (h264, hevc, libsvtav1). Default is libsvtav1. \
# <- Optional: specify video codec (auto, h264, hevc, libsvtav1). Default is libsvtav1. \
# --dataset.vcodec=h264 \
# <- Teleop optional if you want to teleoperate to record or in between episodes with a policy \
# --teleop.type=so100_leader \
@@ -58,7 +60,10 @@ lerobot-record \
--display_data=true \
--dataset.repo_id=${HF_USER}/bimanual-so-handover-cube \
--dataset.num_episodes=25 \
--dataset.single_task="Grab and handover the red cube to the other arm"
--dataset.single_task="Grab and handover the red cube to the other arm" \
--dataset.streaming_encoding=true \
# --dataset.vcodec=auto \
--dataset.encoder_threads=2
```
"""
@@ -179,9 +184,19 @@ class DatasetRecordConfig:
# Number of episodes to record before batch encoding videos
# Set to 1 for immediate encoding (default behavior), or higher for batched encoding
video_encoding_batch_size: int = 1
# Video codec for encoding videos. Options: 'h264', 'hevc', 'libsvtav1'.
# Use 'h264' for faster encoding on systems where AV1 encoding is CPU-heavy.
# Video codec for encoding videos. Options: 'h264', 'hevc', 'libsvtav1', 'auto',
# or hardware-specific: 'h264_videotoolbox', 'h264_nvenc', 'h264_vaapi', 'h264_qsv'.
# Use 'auto' to auto-detect the best available hardware encoder.
vcodec: str = "libsvtav1"
# Enable streaming video encoding: encode frames in real-time during capture instead
# of writing PNG images first. Makes save_episode() near-instant. More info in the documentation: https://huggingface.co/docs/lerobot/streaming_video_encoding
streaming_encoding: bool = False
# Maximum number of frames to buffer per camera when using streaming encoding.
# ~1s buffer at 30fps. Provides backpressure if the encoder can't keep up.
encoder_queue_maxsize: int = 30
# Number of threads per encoder instance. None = auto (codec default).
# Lower values reduce CPU usage, maps to 'lp' (via svtav1-params) for libsvtav1 and 'threads' for h264/hevc..
encoder_threads: int | None = None
# Rename map for the observation to override the image and state keys
rename_map: dict[str, str] = field(default_factory=dict)
@@ -452,6 +467,9 @@ def record(cfg: RecordConfig) -> LeRobotDataset:
root=cfg.dataset.root,
batch_encoding_size=cfg.dataset.video_encoding_batch_size,
vcodec=cfg.dataset.vcodec,
streaming_encoding=cfg.dataset.streaming_encoding,
encoder_queue_maxsize=cfg.dataset.encoder_queue_maxsize,
encoder_threads=cfg.dataset.encoder_threads,
)
if hasattr(robot, "cameras") and len(robot.cameras) > 0:
@@ -474,6 +492,9 @@ def record(cfg: RecordConfig) -> LeRobotDataset:
image_writer_threads=cfg.dataset.num_image_writer_threads_per_camera * len(robot.cameras),
batch_encoding_size=cfg.dataset.video_encoding_batch_size,
vcodec=cfg.dataset.vcodec,
streaming_encoding=cfg.dataset.streaming_encoding,
encoder_queue_maxsize=cfg.dataset.encoder_queue_maxsize,
encoder_threads=cfg.dataset.encoder_threads,
)
# Load pretrained policy
@@ -497,6 +518,11 @@ def record(cfg: RecordConfig) -> LeRobotDataset:
listener, events = init_keyboard_listener()
if not cfg.dataset.streaming_encoding:
logging.info(
"Streaming encoding is disabled. If you have capable hardware, consider enabling it for way faster episode saving. --dataset.streaming_encoding=true --dataset.encoder_threads=2 # --dataset.vcodec=auto. More info in the documentation: https://huggingface.co/docs/lerobot/streaming_video_encoding"
)
with VideoEncodingManager(dataset):
recorded_episodes = 0
while recorded_episodes < cfg.dataset.num_episodes and not events["stop_recording"]: