RoboSet Real Multi-Task Robot Manipulation Dataset
Provider：Carnegie Mellon University and Meta AI (RoboAgent / MT-ACT)
RoboSet is a real multi-task manipulation dataset: Franka plus Robotiq in kitchen scenes, 28,500 trajectories across 38 tasks and 12 skills, under MIT.
At a glance
- Readiness
- L1 · Rigid-body ready
- Rationale
- Manipulation L1: RoboSet collected 28,500 trajectories (9,500 teleoperation plus 19,000 kinesthetic playback) on Franka Emika Panda arms with Robotiq grippers in real kitchen scenes across 38 tasks and 12 skills, with four cameras plus joint / end-effector / gripper data; it is real data rather than simulation-loadable 3D models, hence L1 as a real behavior data layer, consistent with DROID and RH20T.
- Scale
- 28,500 real trajectories (9,500 teleoperation plus 19,000 kinesthetic playback) across 38 tasks and 12 manipulation skills on Franka Emika Panda arms with Robotiq 2F-85 grippers, four cameras (top/left/right plus wrist) with RGB and depth in HDF5; the teleoperation subset Hub mirror is about 425GB (The 28,500 total (9,500 teleoperation plus 19,000 kinesthetic playback), 38 tasks and 12 skills come from the official RoboHive wiki and OpenBot entry; some secondary catalogs misreport '100k+', 'IIT Delhi' and 'CC BY 4.0', which conflict with the official CMU/Meta, MIT attribution and are not adopted. Checked 2026-10-28.)
- License
- MIT — The official RoboHive dataset documentation states the data is 'made available using a MIT license'; secondary catalogs (e.g. geppettorobotdirectory) mislabel it CC BY 4.0, which conflicts with the official MIT attribution, so defer to the official source.
- 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
RoboSet is the real-robot dataset behind RoboAgent and the MT-ACT system, created by researchers at CMU, UT Austin, the University of Washington and Meta AI. It was built to fill the gap of no large-scale robotics dataset in a uniform format suited to robot learning: every instance is a robot trajectory capturing observations, actions, rewards, RGB-D visuals from multiple camera views and other environmental information at each timestep.
Readiness assessment
Collected on Franka Emika Panda arms with Robotiq 2F-85 grippers across everyday kitchen scenes, RoboSet combines 9,500 human teleoperation trajectories with 19,000 expert (kinesthetic-playback) trajectories over 38 tasks expressed as 12 manipulation skills, recording four camera views plus joint positions/velocities, end-effector and gripper pose/velocities and applied controls in HDF5. It reaches both classic HDF5 learners and a Hugging Face Hub mirror of the teleoperation subset, so it serves as a structured multi-task benchmark that complements in-the-wild corpora like DROID.
Practical notes
The official RoboHive documentation states the data is made available under an MIT license; secondary catalogs that label it CC BY 4.0 are incorrect, so defer to the official source. The teleoperation subset Hub mirror is about 425GB, and the wrist camera is one of the four views, so storage and camera-mode handling matter when building loaders. Because the corpus is single-arm Franka in standardized kitchen setups, it offers controlled task-transfer studies rather than environmental diversity.
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.