The DARRL dataset: Demonstrations for Action Recognition and Robot Learning

The DARRL dataset (Demonstrations for Action Recognition and Robot Learning) is a collection of 760 RGB-D videos of humans performing various manipulation tasks. It is provided with object and action annotations (in the COCO format) for 30 of those videos; segmentation masks are also provided. Additionnal gaze data collected from outside observers is shared for research on human attention and observation strategy analysis. It can also be used as a basis for learning from demonstrations for a robotic arm, for instance. This work is supported by Région Pays de la Loire.