Abstract: Techniques for orchestrating activities at an earthmoving site are described. In an example, a stream of images are captured by an image capturing device removably fixed to a mobile earthmoving machine with a field of view overlapping with a working range of an attachment coupled to the mobile earthmoving machine. Using a selected subset of images from the stream captured during a time period, a first trained algorithm generates a probability that the subset of images depicts an action performable by the machine. Based on the probability, it is determined that the machine was performing the action during the time period, and in response, an image from the subset of images is selected. The image and an indication that the machine was performing the action during the time period are transmitted to a remote server system through a network.