RoboMIND: Benchmark on Multi-embodiment Intelligence Normative Data for Robot Manipulation
RoboMIND is a teleoperated, multi-embodiment manipulation dataset covering four robot types: a Franka Emika Panda, a UR5e, an AgileX dual-arm robot and a humanoid with dual dexterous hands. The project page reports about 107k real-world demonstration trajectories across 479 tasks and 96 object classes, including 5k real-world failure demonstrations, plus an Isaac Sim digital twin.
Multi-embodiment imitation learning and vision-language-action training on real manipulation tasks, including learning from failure demonstrations. The paper evaluates single-task imitation learning and multi-task VLA models trained on the data.
Human teleoperation records multi-view observations, robot joint states and a language description for each episode, organized by embodiment and task. An Isaac Sim digital twin replicates the real tasks and assets to support extra data collection and evaluation. The data are hosted on Hugging Face and ModelScope, according to the project page.
It adds real humanoid manipulation data with dual dexterous hands alongside arm platforms, which supports cross-embodiment comparisons. The exact humanoid model is not confirmed in the sources, so it is not linked to a roster robot.
Scoring
This record defines no leaderboard metrics.
Hardware
No in-dataset robot embodiments are linked.
Four embodiments: Franka Emika Panda, UR5e, AgileX dual-arm, and a humanoid with dual dexterous hands. The project page's hardware notes mention 'Tien Kung', which may name this humanoid, but the wording is ambiguous. The roster robot tiangong-ultra may or may not be this model, so it is not linked.
Recorded outcomes
No result files yet
Evidence
Related evaluations