RoboMIND Multi-Embodiment Robot Manipulation Dataset and Benchmark
Provider：Beijing Innovation Center of Humanoid Robotics, Peking University and Beijing Academy of Artificial Intelligence
RoboMIND provides 107k multi-embodiment teleoperated trajectories over 479 tasks and 96 object classes, with 5k failure demonstrations, an Isaac Sim digital twin and Apache-2.0 licensing.
At a glance
- Readiness
- L1 · Rigid-body ready
- Rationale
- Manipulation L1: 107k real teleoperated demonstration trajectories span 4 embodiments, 479 tasks and 96 object classes, with multi-view observations, proprioceptive state and language descriptions, plus an Isaac Sim digital twin; but the main assets are real-world video and action logs rather than simulation-loadable 3D models, hence L1 as a real behavior data layer, consistent with how BridgeData V2 and RLBench are graded.
- Scale
- 107k real demonstration trajectories over 4 embodiments (Franka 52,926, Tien Kung 19,152, AgileX 10,629, UR-5e 25,170), 479 tasks and 96 object classes, plus 5k failure demonstrations (Per-embodiment counts follow the official project page and the RSS 2025 paper (DOI 10.15607/RSS.2025.XXI.152). The site also announces the complete RoboMIND V2.0 collection hosted on ModelScope, but publishes no single unified trajectory total for V2.0, so only the paper's 107k figure is recorded here. Checked 2026-10-26.)
- License
- Apache-2.0 — The official Hugging Face dataset card x-humanoid-robomind/RoboMIND declares Apache-2.0 (written apache-2 in this site's controlled vocabulary), and the ModelScope-hosted copy carries the same license. Access requires accepting the terms and passing gated auto approval on Hugging Face, so this is an open license behind a registration step and access is recorded as gated rather than open.
- Access
- Sign-up / registration
- Formats
- other
- Runtimes
- other
- Physics properties
- Collision geometryJoint structureSensor attributes
- Domain
- Other
- Published
- 2026-10-26
- Updated
- 2026-10-26
Why it is listed
RoboMIND targets a specific gap in the open manipulation data ecosystem: individual labs publish their own recordings, but there is no unified collection standard, so quality and coverage vary. The project built one platform and one protocol, then collected everything through it. The result is described as the largest multi-embodiment teleoperation dataset gathered on a single unified platform, spanning a Franka Emika Panda, a UR5e, an AgileX dual-arm robot and the Tien Kung humanoid with dual dexterous hands.
Readiness assessment
Every trajectory is human-teleoperated and stored on a unified intelligence platform, carrying multi-view observations, proprioceptive state and a language description. Data is delivered as HDF5 files organized per embodiment and camera configuration, with a documented layout, and about 10,000 successful trajectories carry refined segment-level language annotations that were machine-drafted and then manually revised. A digital twin in Isaac Sim replicates the real tasks and assets, which supports low-cost extra data collection and evaluation.
Practical notes
Two features distinguish this release. First, roughly 5,000 real failure demonstrations are included, each annotated with its cause, so failure cases can be studied rather than discarded during quality inspection. Second, trajectory length differs sharply by embodiment: Franka and UR episodes are usually under 200 steps and suit fundamental skills, while Tien Kung and AgileX episodes commonly exceed 500 steps and fit long-horizon, multi-skill tasks. The dataset card notes that 675 trajectories in one folder contain only left and right camera images because of equipment maintenance, so per-folder completeness varies.
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.