The BC-Z dataset is a large-scale robot learning dataset jointly developed by Google, Everyday Robots, UC Berkeley, and Stanford University in 2022 to promote the development of robot imitation learning.BC-Z: Zero-Shot Task Generalization with Robotic Imitation LearningThe core contribution of this dataset is that it supports zero-shot task generalization, which enables robots to perform new manipulation tasks through imitation learning without prior experience.

The BC-Z dataset contains more than 25,877 different manipulation task scenarios, covering 100 diverse manipulation tasks. These tasks were collected through expert teleoperation and shared autonomy processes, involving 12 robots and 7 different operators, with a total of 125 hours of robot operation time. The dataset supports training a 7-DOF multi-task policy that can be adjusted based on the language description of the task or human operation video to perform specific manipulation tasks.