Part 4/8:
Historically, a considerable workforce was employed to manually tag data needed for training Tesla's neural networks; however, this process was both slow and susceptible to human error. To address these challenges, Tesla has initiated the automation of data labeling through the Cortex system.
According to a Tesla patent application, the workflow for data labeling consists of three main steps:
Capture of image and navigational data from Tesla vehicles and Optimus robots.
Collection of data from multiple trips to establish an environment model, possibly utilizing a 3D representation.
Use of the generated model by the Cortex supercomputer to autolabel new data, thereby streamlining the training process significantly.