lerobot

mirror of https://github.com/huggingface/lerobot.git synced 2026-08-06 00:19:47 +00:00

Files

T

pepijn 674c990a39 feat(streaming): default episode pool 1024 and wire streaming into lerobot-train

Raise the default episode_pool_size to 1024 (DatasetConfig + StreamingLeRobotDataset)
for better default shuffle quality at scale.

Streaming is now a first-class option of the main train script: when cfg.dataset.streaming
is set, the dataloader is not handed to accelerate (the dataset is already rank-disjoint via
split_dataset_by_node, so IterableDatasetShard would drop (N-1)/N of each rank's stream),
batches are moved to device manually, and the episode-aware sampler is skipped. Remove the
standalone examples/scaling/train_streaming_multinode.py example in favor of this wiring.

Co-authored-by: Cursor <cursoragent@cursor.com>

2026-06-12 09:24:32 +00:00

backward_compatibility

feat(dependencies): minimal default tag install (#3362 )

2026-04-12 20:03:04 +02:00

dataset

refactor: support custom progress parquet overlays (#3640 )

2026-05-21 14:32:10 +02:00

lekiwi

feat(rollout): decouple policy deployment from data recording with new lerobot-rollout CLI (#3413 )

2026-04-28 00:57:35 +02:00

notebooks

docs(quickstart): adding rollout (#3598 )