I totally agree. However, knowing the particular advantages of one approach vs another takes time to learn and understand. All well beyond the experience level of a young engineer, On the other hand, I quess that was why the question was asked in the first place.
One other bit of advice comes to mind. Several years after building my FPGA base real time Kalman Image processor I came back to look at the problem with the GPU also added to the signal processing mix. Everything had flipped. Even though I was using high performance FPGAs, there was far more horse power in the GPU. My suggestion would be to use the FPGA to get your image over to a GPU say via an ethernet port, perform your algorithms there at the GPU, and then send back control vectors back to the robot.