This dataset was created by the Psi-Robot team in 2025. The research background is based on the need for high success rate of dexterous grasping in cluttered scenes, especially achieving a success rate of more than 90% under unseen objects, lighting and background combinations. The related paper results are "DexGraspVLA: A Vision-Language-Action Framework Towards General Dexterous GraspingThis framework uses a pre-trained vision-language model as a high-level task planner and learns a diffusion-based strategy as a low-level action controller. Its innovation lies in using the basic model to achieve strong generalization capabilities and using diffusion-based imitation learning to obtain dexterous actions.

This is a small dataset with 51 samples of human demonstration data, which is useful for understanding the data and format, and running the code to experience the training process.