UniACT-dataset

arXiv

From six public datasets, we clean, standardize, and balance samples to construct UniACT-dataset, a large-scale dataset with over 6 million trajectories and 9,500 hours of data,

Action Representation: Delta Actions in End-Effector Position with Rotation Vectors.

• Unified Single- and Dual-Arm Training via Zero Padding.
  • all single-arm data are uniformly treated as right-arm executions within a dual-arm configuration.
  • the action dimensions corresponding to the ununsed arm are filled with zeros.
• Data Distribution: Diverse Embodiments and Task Scenarios. more than six million trajectories spans more than twenty unique robot embodiments and 9500+ hours
  • dataset scale distribution
  • embodiment distribution