DROID — A Large-Scale In-The-Wild Robot Manipulation Dataset
DROID is a large-scale in-the-wild robot manipulation dataset collected by Stanford University, UC Berkeley, Carnegie Mellon University, and eight additional institutions using the Franka Panda robotic arm. Released in 2024 under the Apache 2.0 license for unrestricted commercial use, it contains 76,000 demonstration episodes totalling approximately 350 hours of real-world manipulation data collected across diverse environments at multiple research sites. Tasks span object rearrangement, pick-and-place, tool use, and environmental interaction across kitchens, offices, labs, and unstructured spaces. Data is distributed in Zarr and HDF5 formats via Hugging Face and is included in the Open X-Embodiment collection. DROID is distinguished by its scale of institutional collaboration and in-the-wild collection methodology — demonstrations were collected by multiple operators across geographically distributed sites rather than in controlled single-lab settings, making it one of the most diverse manipulation datasets available.
| Year | 2024 |
|---|---|
| Episodes | 76,000 |
| Total hours | 350 |
| Embodiments | Franka Panda |
| Modalities | rgb, proprioception |
| View type | third-person, wrist-cam |
| Task categories | manipulation, pick-and-place, long-horizon |
| Data format | zarr, lerobot, hdf5 |
| License | Apache 2.0 |
| Access | open — commercial use permitted |
| Maintainer | Stanford University, UC Berkeley, CMU, 8 additional institutions |
| Origin country | US |
What is it?
DROID is a large-scale in-the-wild robot manipulation dataset collected by Stanford University, UC Berkeley, Carnegie Mellon University, and eight additional institutions using the Franka Panda robotic arm. Released in 2024 under Apache 2.0, it contains 76,000 demonstration episodes totalling approximately 350 hours of real-world manipulation data collected across geographically distributed research sites. Tasks span object rearrangement, pick-and-place, tool use, and environmental interaction across kitchens, offices, labs, and unstructured spaces.
Who is it for?
DROID targets researchers developing generalised manipulation policies that must handle the diversity of real-world deployment environments. It is particularly valuable for teams studying in-the-wild generalisation — the challenge of maintaining policy performance when the environment, objects, and lighting conditions are unfamiliar. The Franka Panda is the dominant research arm globally, making DROID relevant to the largest segment of academic manipulation researchers.
Key specifications
- Episodes: 76,000 demonstrations
- Duration: ~350 hours
- Embodiment: Franka Panda (7-DoF arm)
- Tasks: Pick-and-place, object rearrangement, tool use, environmental interaction
- Format: Zarr, LeRobot, HDF5
- License: Apache 2.0 — commercial use permitted
- Access: Open — Hugging Face
How it compares
DROID is the largest single-embodiment Franka Panda dataset. BridgeData V2 (60,096 episodes) covers more environment diversity on a lower-cost arm. Open X-Embodiment includes DROID as a subset but mixes it with 21 other embodiments. DROID's distinguishing feature is coordinated multi-site collection — data comes from 10 different institutions across diverse geographic locations, providing environmental diversity without the quality heterogeneity of Open X-Embodiment.
Limitations and access notes
DROID covers single-arm manipulation only — no bimanual, legged, or humanoid data. The Franka Panda is an industrial-grade arm costing approximately $15,000-20,000, limiting accessibility for smaller labs. Apache 2.0 permits unrestricted commercial use.
Linked professions
Frequently asked questions
What robot is DROID collected on?
DROID uses the Franka Panda, a 7-DoF research arm manufactured by Franka Robotics (Germany). The Franka Panda is the most widely used robot arm in academic manipulation research globally, appearing in hundreds of published papers.
What makes DROID different from other manipulation datasets?
DROID was collected across 10 different institutions in diverse geographic locations — offices, kitchens, labs, and unstructured spaces at Stanford, UC Berkeley, CMU, and seven other sites. This multi-site in-the-wild collection methodology provides environmental diversity that single-lab datasets cannot match.
Can DROID be used commercially?
Yes. DROID is licensed under Apache 2.0, permitting unrestricted commercial use, modification, and redistribution.
How do I access DROID?
DROID is available on Hugging Face at huggingface.co/datasets/fracapuano/droid and via the project GitHub at github.com/droid-dataset/droid. No registration required.
Is DROID included in Open X-Embodiment?
Yes. DROID is one of the constituent datasets in the Open X-Embodiment collection. However, it is also available as a standalone dataset with full metadata and documentation separately from the Open X-Embodiment unified format.