SKU DEM-6003 · Sold by External
DROID
Product specifications
| SKU | DEM-6003 |
|---|---|
| Data type | Demonstrations |
| Volume | 76,000 teleoperated demonstration trajectories (episodes) |
| Size on disk | ~2 TB (RLDS/TFDS format); ~400 GB (LeRobot video-compressed conversion) |
| Format | RLDS / TensorFlow Datasets (tfds.load("droid")); also raw HDF5 and LeRobot parquet+mp4 conversions |
| Access model | PUBLIC LICENSE |
| Pricing | Free · open dataset (CC-BY 4.0) |
| Quality score | — |
| License | Creative Commons Attribution 4.0 International (CC-BY 4.0) |
DROID (Distributed Robot Interaction Dataset) contains 76,000 teleoperated demonstration trajectories (about 350 hours of interaction) collected on a single shared, open-source hardware platform: a Franka Panda 7-DoF arm with two adjustable Zed 2 stereo cameras, a wrist-mounted Zed Mini, and an Oculus Quest 2 for teleoperation. The data spans 564 scenes, 86 tasks, and 52 buildings, collected by 50 data collectors across 18 robotics labs (13 institutions) in North America, Asia, and Europe over 12 months. Each episode provides three synchronized stereo RGB camera streams, camera calibration, robot proprioceptive state and control commands, and up to three independent natural-language instruction annotations per successful episode. The dataset is distributed in RLDS/TFDS format (loadable from gs://gresearch/robotics) and has community conversions to HDF5 (raw) and LeRobot (parquet + mp4). These are pure human demonstrations with no reward labels; the released set is filtered to successful episodes, and policies trained on DROID are evaluated by human-judged task completion. The authors open-source the full dataset under the CC-BY 4.0 license.