LIBERO Data Flywheel

The optional Flywheel records executed LIBERO evaluation trajectories without changing the planner or action primitives. It stores immutable raw episodes; conversion to a training format is a separate step.

Installation

Follow Installation for environment setup and simulator assets. From the RPent repository root, install LIBERO-PRO with Flywheel export support:

pip install -e ".[libero-pro,flywheel]"

Collect an episode

Enable collection on a normal LIBERO evaluation run and choose a data root:

rpent --robot libero \
  --suite libero_goal --task 0 --seed 0 \
  --planner codex \
  --collect-flywheel-data \
  --flywheel-root /path/to/datacollection

The run writes one episode below /path/to/datacollection/raw/libero/<suite>/task_<id>/seed_<seed>/. Each episode contains the policy observations, executed actions, rewards, terminal flags, primitive IDs, and VLA proposals. Collection is opt-in and is supported only for evaluation mode.

Validate and export successful episodes

Validate one raw episode before using it:

rpent-flywheel validate /path/to/raw/episode

Export every finalized successful episode for one suite and task to a LeRobot dataset:

rpent-flywheel export-lerobot \
  --data-root /path/to/datacollection \
  --suite libero_goal \
  --task 0 \
  --dataset-id goal-task-00 \
  --output-root /path/to/lerobot

--data-root is the raw data root, while --output-root is the parent directory of the exported dataset. This example writes the dataset to /path/to/lerobot/goal-task-00.

For each successful episode, the exporter includes all executed actions from the start through the first action that achieves task success, including scripted and VLA actions. Actions after the first success are excluded.

The exporter never rewrites the raw episodes. Failed episodes remain available for auditing but are not included in this supervised-training export.

Train with RLinf

After exporting successful trajectories to LeRobot, you can perform Pi0.5 supervised fine-tuning in a separate RLinf environment. RPent handles collection and export; RLinf handles model training, logging, and checkpoint saving.

For environment setup and training configuration, see the RLinf OpenPI_RLinf SFT guide. Point the training data path to the exported LeRobot dataset directory containing meta/info.json, not the raw episode directory.

Use an RLinf environment with LIBERO LeRobot data-loading support. Prepare a Pi0.5 LIBERO SFT configuration at /path/to/sft-config/libero_pi05_sft.yaml, then run:

cd /path/to/RLinf
unset PYTHONPATH
export PYTHONPATH="$PWD"
export EMBODIED_PATH="$PWD/examples/sft"

.venv/bin/python examples/sft/train_vla_sft.py \
  --config-path /path/to/sft-config \
  --config-name libero_pi05_sft \
  data.train_data_paths=/path/to/lerobot/goal-task-00 \
  actor.model.model_path=/path/to/pi05-checkpoint \
  runner.logger.log_path=/path/to/new-training-output

Match the configuration’s resource placement and batch size to your available GPUs, and use the checkpoint and normalization settings for the selected model and dataset.