Take your experiments further.

1. Export from the lab

Choose Export ↓ beside an episode, or Export dataset ↓ for all episodes. The JSON file includes states, commands, object motion, task labels, and any recorded camera images.

For camera datasets, select Include overhead camera before recording. Camera and state-only episodes must be converted separately.

2. Convert on your computer

This optional step needs Python 3.12 and LeRobot 0.4.4. It runs outside the browser and does not connect to a robot or upload your recordings.

Download the LeRobot converter ↓

Create a separate Python environment and install the pinned library:

python3.12 -m venv .venv
source .venv/bin/activate
python -m pip install "lerobot==0.4.4"

Then run the downloaded converter, replacing the input filename with your export. Choose a new output folder:

python export_lerobot.py verapulse-dataset.json ./my-dataset

The converter checks the recording, writes the dataset, and loads it again with LeRobot to verify it. Keep the original JSON as the complete source record.

3. Match your policy’s inputs

The dataset contains joint positions in radians, joint velocities, task descriptions, terminal outcome labels, and optional 320 × 240 overhead images. Actions are joint velocities in radians per second. An existing SO-101 policy that expects position commands needs its own action adapter before these recordings can be used with it.

Samples use a regular 20 Hz simulation-time grid. The most recent complete observation is retained; the original wall and simulation clocks are included. Images are embedded in the dataset rather than encoded as video.

The meta/verapulse.json file retains scene configuration, camera details, task evaluation, and the original command log. Contacts and full object trajectories remain in the original JSON. Success labels describe this reference simulation, not validated physical performance.

Current support: simulation episodes with the same robot revision, joint ordering, and camera setup. Live robot recording and custom policy execution are still planned.