RoboTurk Real Robot Dataset
Provider：Stanford University (Fei-Fei Li, Silvio Savarese et al., Vision and Learning Lab)
RoboTurk is a crowdsourced real teleoperation dataset: 54 remote users drove a Sawyer on three long-horizon tasks, 2,144 demos / 111 hours, with MIT code licensing.
At a glance
- Readiness
- L1 · Rigid-body ready
- Rationale
- Manipulation L1: RoboTurk let 54 remote operators control a real Sawyer arm via phone or browser through a crowdsourcing platform to perform three long-horizon tasks, collecting 2,144 real teleoperation demonstrations (111 hours); it is real data rather than simulation-loadable 3D models, hence L1 as a real behavior data layer, consistent with DROID and BridgeData V2.
- Scale
- 2,144 real teleoperation demonstrations / 111 hours from 54 remote non-expert users on a Sawyer single arm across three long-horizon tasks (laundry layout, tower creation, object search), each frame with Kinect RGB-D, a USB camera stream, joint data and control signals in HDF5 with aligned MP4; about 54.8GB total (Demo and hour counts come from the official RoboTurk real dataset page (roboturk.stanford.edu/dataset_real.html): 2,144 demos / 111 hours / 54 users / 3 tasks. The Hugging Face LeRobot conversion (roboturk_lerobot) holds only 1,796 episodes / 168,423 frames (no camera video) as a format subset and does not replace the official full-count basis. Checked 2026-10-28.)
- License
- MIT — The official code repository is MIT; the real dataset release page does not spell out a separate data license (catalogs such as kineticblocks and humanoidsdata flag that terms must be confirmed with the publisher), so the data is described per the MIT code repository. Many community LeRobot conversions list apache-2.0, but that is a conversion-script default and must not be read as the upstream data term.
- Access
- Open
- Formats
- other
- Runtimes
- other
- Physics properties
- Joint structureSensor attributes
- Domain
- Other
- Published
- 2026-10-28
- Updated
- 2026-10-28
Why it is listed
RoboTurk is a pioneering crowdsourced dataset for robotic imitation learning, built at Stanford's Vision and Learning Lab under Fei-Fei Li and Silvio Savarese. Over one week, 54 remote operators controlled physical Sawyer arms through phones or browsers to perform three long-horizon tabletop tasks, producing one of the largest human-teleoperation corpora of its time and demonstrating that robot data can be collected at scale without lab presence.
Readiness assessment
Each of the 2,144 demonstrations stores Kinect RGB and depth video, a USB webcam stream, robot joint data and the operator's control signals with timestamps in HDF5 alongside aligned MP4 files, covering laundry layout, tower creation and object search. The companion robomimic framework ships the dataset plus learning algorithms, and a Hugging Face LeRobot conversion (roboturk_lerobot) reformats 1,796 episodes into the Open X / LeRobot layout, so the data reaches both classic HDF5 and modern RLDS pipelines.
Practical notes
The official code repository is MIT, but the real dataset release page does not state a separate data license, so confirm terms with the publisher before redistribution; community LeRobot conversions list apache-2.0 as a conversion default rather than an upstream term. Full per-task sets are 17 to 19GB each with 10-demo mini sets available, so download sizing matters on shared clusters. Because the corpus predates modern VLA models, it is most useful as a long-horizon teleoperation benchmark and a historical reference rather than a primary pre-training source.
Disclaimer: license and usage terms follow the official source — verify them yourself before use. This site indexes and links only; it does not host or mirror any asset file.
License and usage terms are those of the official source — verify before use. No asset file is hosted or mirrored on this site.