Data Flywheel#

Use --record-trajectory with the api planner to save model requests, responses, and tool calls. Supported environments also record executed actions and their observations for later export to LeRobot training data.

Installation#

The example below uses LIBERO-PRO. Follow Quick Start for environment setup and simulator assets, and Planner Configuration for API credentials. From the RPent repository root, install LIBERO-PRO with the optional export dependencies:

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

Record a Trajectory#

Recording is off by default and requires --planner api. Action recording supports LIBERO, RoboCasa365, RoboTwin, and RoboDojo. Enable recording and choose a run directory:

rpent --robot libero \
  --suite libero_goal --task 0 --seed 0 \
  --planner api --model anthropic:claude-opus-4-8 \
  --record-trajectory \
  --output-dir /path/to/run

--record-trajectory: Enable trajectory recording for model requests, responses, and tool calls. Supported environments also record executed actions and their observations. Recording is off by default and requires --planner api.
--output-dir /path/to/run: Set the output root directory for this run. Replace /path/to/run with your desired save path. Trajectory records are saved in its trajectories/<task_id>/ subdirectory.

After the run, records are saved under /path/to/run/trajectories/<task_id>/: task.json, planners.jsonl, actions.jsonl, predictions.jsonl, and an artifacts/ directory for attachments. VLA records link each instruction to the actions actually executed; predicted actions that were not executed are not saved.

Export Training Data#

Export recorded trajectories from one robot to a new output directory:

rpent-flywheel \
  --input-root /path/to/run \
  --output-dir /path/to/lerobot/goal-task-00

--input-root accepts a trajectory directory containing task.json or a parent directory containing multiple runs. Export each robot separately; --output-dir must not already exist.

Add --image-writer-threads 4 to compress and save images with four threads. The default, 0, writes images synchronously.

Only tasks confirmed successful by the environment are eligible; planner completion alone is insufficient. The exporter selects and organizes training data using the following rules:

Success criterion: environment_success in the task result must explicitly be true. Failed tasks, missing results, and tasks whose success cannot be confirmed are excluded.
Action source: Only executed actions uniquely linked to a valid VLA prediction record are retained. The prediction must contain a non-empty instruction and belong to the same tool call as the action. Scripted actions, other non-VLA actions, and unexecuted predicted actions are excluded.
Data organization: Each continuous segment is saved as a separate LeRobot episode. Each executed action corresponds to one frame containing the observation before execution, the action, and the corresponding VLA instruction.

The dataset is written to /path/to/lerobot/goal-task-00/train/ without an automatic validation split. report.json in the output directory lists export counts and exclusions. If no actions qualify, only the report is written. Original recordings are left unchanged.

Train with RLinf#

For the LIBERO example, use the exported LeRobot dataset for 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 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/train \
  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.

SFT Experiment Results#

On 2026-10-23, successful trajectories collected with RPent were used to fine-tune RLinf-Pi05-LIBERO-130-fullshot-SFT. The resulting VLA model was evaluated against the original model within RPent.

Training used 199 usable successful trajectories from LIBERO-PRO’s three goal suites. Language-branch LoRA and full fine-tuning of the action branch ran for 500 steps. The saved checkpoint at step 300 was selected by validation loss for evaluation.

Evaluation covered tasks 0–9 and seeds 25–34 in libero_goal_swap, for 100 runs per model. Both groups used the same DeepSeek-V4.1-Flash-1017 planner (reasoning effort: none) and official memory, with only the VLA model replaced.

Task success rates for the full RPent system were:

Model	Successes / runs	Success rate
Original model	70/100	70%
SFT model (step 300)	73/100	73%

In this evaluation, the SFT model succeeded in three more runs, increasing the success rate by three percentage points.