DROID: Distributed Robot Interaction Dataset

76,000 real-world teleop trajectories across 564 scenes — the largest single-arm manipulation dataset and the canonical pretraining corpus for OpenVLA, π0, and Octo.

TL;DR

Metric	Value
Task count	86 distinct task descriptions (open vocabulary language)
Robots	Franka Emika Panda with parallel-jaw gripper
Modalities	2x Zed 2 stereo RGB-D + 1x Zed Mini wrist + language + 7-DoF joints
License	CC-BY 4.0
Size	76K trajectories, 350 hours, ~1.7 TB RLDS
Scenes	564 real-world environments across 18 institutions

What is DROID?

DROID (Distributed Robot Interaction Dataset) is the output of a Stanford- and UC Berkeley-led consortium of 18 institutions that together collected the largest publicly available in-the-wild robot manipulation dataset. The data release was announced in early 2024 and captures 350 hours of teleoperated Franka Emika Panda interaction spanning 76,000 episodes across 564 distinct real-world scenes — kitchens, offices, labs, hotel rooms, industrial workspaces, even a convenience store.

The key design choice that differentiates DROID from prior real-world datasets is its commitment to scene diversity over task depth. Rather than recording thousands of trials of a small set of manipulation primitives, DROID collected a long tail of 86 distinct task descriptions (open vocabulary language like "pour the cereal into the bowl" or "wipe the spilled coffee") across hundreds of unique backgrounds. The resulting distribution is much harder for policies to memorize and much more reflective of real deployment conditions.

Every DROID episode ships with three stereo RGB-D cameras — two externally mounted Zed 2 units plus one Zed Mini mounted on the wrist — at a consistent 15 Hz rate, alongside the Franka's 7-DoF joint state, end-effector pose, gripper width, and a natural language instruction. The dataset is distributed in the RLDS / TFDS format used by Open X-Embodiment and is mirrored on Hugging Face as LeRobot-format Parquet.

How to download & load

# Install TFDS + the DROID spec
pip install tensorflow_datasets rlds tensorflow
python -c "
import tensorflow_datasets as tfds
ds = tfds.load('droid', data_dir='gs://gresearch/robotics/droid/1.0.0')['train']
for ep in ds.take(1):
    print(ep.keys())"

# Or stream via LeRobot
pip install lerobot
python -c "
from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
d = LeRobotDataset('KarlP/droid', streaming=True)
print(next(iter(d)))"

The full mirror is hosted on Google Cloud Storage under gs://gresearch/robotics/droid/ and is free to egress. A 100-episode "DROID-100" sample is packaged on Hugging Face for quick experimentation and fits on a laptop.

Common use cases & pairings

• VLA pretraining. OpenVLA, π0, Octo, and the RT-2-X family all include DROID in their pretraining mixture — it provides the long-tail scene diversity that cross-embodiment data cannot supply alone.
• Fine-tuning on a single scene. Teams frequently pretrain on DROID and fine-tune on ~50 demos of their target task collected in their own lab.
• Language grounding. Because tasks are described in open vocabulary English, DROID is the best public dataset for evaluating CLIP / VLM grounding on real robot data.
• Scaling law studies. DROID's sheer size lets researchers characterize how success rate improves with dataset fraction — a signal not available in smaller datasets.

Benchmarks & leaderboards

DROID is primarily used as pretraining data rather than a fixed benchmark, but the policy learning repository publishes reference Diffusion Policy and ACT numbers on the DROID-100 split. See the Papers with Code DROID page and the project site for current results.