User Guide · Data Pipelines · RLDS export

RLDS export

Some VLA stacks — most notably OpenVLA / OpenVLA-OFT — consume the RLDS / TFDS format rather than LeRobot. utils/convert_to_rlds.py scans recorded trajectory dirs and emits a TFDS-style RLDS dataset.

Convert recorded trajectories

-i/--input-dirs takes one or more directories to scan (by -p/--pattern); -o/--output-dir is where the RLDS dataset is written.

$ python utils/convert_to_rlds.py \
    -i demos/grasp_part/100221 \
    -o demos/datasets/rlds \
    --dataset-name grasp_part_rlds \
    --image-size 256

Flags

Flag	Default	Notes
`-i`, `--input-dirs`	— (required)	One or more dirs to scan for trajectories
`-o`, `--output-dir`	`/nat/demos/datasets/rlds`	RLDS dataset output dir
`--dataset-name`	`fgmanip_rlds`	TFDS dataset name
`--image-size`	`256`	Square image resize
`--val-ratio`	`0.0`	Validation split fraction
`--all-episodes`	off	Include non-successful episodes too
`--max-episodes`	`0`	Cap episodes (0 = no cap)
`-p`, `--pattern`	replay pattern	Filename pattern to collect
`--seed`	`42`	Split shuffle seed

Use with OpenVLA / OpenVLA-OFT

Point the OpenVLA training config at --output-dir; both OpenVLA and OpenVLA-OFT consume RLDS shards directly. Mixing MetaFine RLDS with OXE / OpenX-style datasets is the usual path for joint pretrain + fine-tune runs.

Canonical source: python utils/convert_to_rlds.py --help.