> Is it training initialization or is it a form of transfer learning?

I doubt it's randomly initialized. The brain has centers for motion which might be honed in at birth for these animals.

Let's not ignore genetics, which pre-populate weights and subnetworks in the form of motion-control centers in the brain. Once you consider the gene-brain interaction over the course of evolution as part of your training, the huge data required for modern ML starts to make more sense.