DexScale

Dexterous manipulation is essential for building general-purpose embodied agents, yet learning such skills remains challenging due to high-dimensional hand control, complex contact dynamics, and the lack of large-scale, high-quality datasets. Existing robot manipulation datasets are often limited to simple grippers, visual observations, single embodiments, or short-horizon tasks, making them insufficient for studying general dexterous manipulation.

In this work, we introduce DexScale, a diverse and large-scale vision-tactile dataset designed for general dexterous robotic manipulation. DexScale supports dexterous hands with different degrees of freedom and provides synchronized multimodal observations, including robot actions, RGB-D visual perception, and in-hand tactile sensing. Built upon a high-fidelity isomorphic data collection system, DexScale captures fine-grained, smooth, and contact-rich manipulation trajectories that are difficult to obtain with conventional teleoperation pipelines.

The dataset covers a broad range of tasks and scenarios, including tool use, dexterous object manipulation, diverse grasping, bimanual coordination, long-horizon execution, and interactions in varied environments. By scaling dexterous manipulation data across embodiments, modalities, tasks, and scenes, DexScale provides a foundation for training and evaluating policies that integrate spatial perception, tactile feedback, and precise hand control. Together, DexScale aims to bridge the gap between large-scale robot learning and fine-grained dexterous control, offering a data foundation for developing general-purpose vision-tactile manipulation policies.