Data Conversion¶

Convert collected .vla trajectories to LeRobot / DROID format for use with the openpi training pipeline.

DROID Schema¶

The exporter targets the DROID LeRobot schema:

Three named cameras: wrist_image_left, exterior_image_1_left, exterior_image_2_left
joint_position — first num_joints dims of observation/proprio_joints
actions — joint velocities (num_joints) concatenated with a zeroed gripper channel

Configuration¶

All fields can be overridden at the command line with tyro.

DatasetCfg fields:

Field	Default	Description
`image_height`	`180`	Output image height in pixels
`image_width`	`320`	Output image width in pixels
`fps`	`50`	Dataset frames per second
`robot_type`	`"panda"`	Robot type label in LeRobot metadata
`repo_id`	`"rio"`	HuggingFace repo ID for the dataset
`num_joints`	`None`	Number of joints; inferred from data if `None`
`action_dim`	`None`	Action dimensionality; inferred if `None`
`camera_mapping`	`None`	Dict mapping camera keys → DROID names; inferred if `None`

Script-specific Args fields:

Field	Default	Description
`input`	`"/tmp/dummy_data/"`	Path to a `.vla` file or directory
`output`	`None`	Output directory (defaults to `~/.cache/huggingface/lerobot/{repo_id}`)
`robot_type`	`"xarm"`	Overrides `DatasetCfg` default for xarm datasets
`verbose`	`False`	Enable verbose logging
`clean`	`False`	Delete output directory before converting

Auto-Inference¶

When num_joints, action_dim, or camera_mapping are left as None, the script infers them from the first trajectory:

num_joints — last dimension of observation/proprio_joints
action_dim — num_joints + gripper dimensions
camera_mapping — assigns DROID names by ascending resolution (smallest → wrist_image_left)

Convert to DROID¶

Minimal — auto-infer everything:

uv run examples/data/convert_to_lerobot_droid.py --input /data/rollouts/my_task/

Output lands in ~/.cache/huggingface/lerobot/rio by default.

Specify output path and repo ID:

uv run examples/data/convert_to_lerobot_droid.py \
    --input /data/rollouts/my_task/ \
    --output /data/lerobot/my_task/ \
    --repo-id myuser/my_task

Re-run from scratch (wipe existing output):

uv run examples/data/convert_to_lerobot_droid.py \
    --input /data/rollouts/my_task/ \
    --output /data/lerobot/my_task/ \
    --clean

Override inferred fields:

uv run examples/data/convert_to_lerobot_droid.py \
    --input /data/my_task \
    --num-joints 7 \
    --action-dim 8

Compute Normalization Statistics¶

After conversion, compute normalization stats required by the openpi training pipeline:

cd third_party/openpi
uv run scripts/compute_norm_stats.py --config-name pi05_droid_finetune_test